yacy_search_server/source/de/anomic/plasma/plasmaCrawlStacker.java

890 lines
39 KiB
Java
Raw Normal View History

// plasmaCrawlStacker.java
// -----------------------
// part of YaCy
// (C) by Michael Peter Christen; mc@anomic.de
// first published on http://www.anomic.de
// Frankfurt, Germany, 2005
//
// This file was contributed by Martin Thelian
//
// $LastChangedDate$
// $LastChangedRevision$
// $LastChangedBy$
//
// This program is free software; you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation; either version 2 of the License, or
// (at your option) any later version.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program; if not, write to the Free Software
// Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
//
// Using this software in any meaning (reading, learning, copying, compiling,
// running) means that you agree that the Author(s) is (are) not responsible
// for cost, loss of data or any harm that may be caused directly or indirectly
// by usage of this softare or this documentation. The usage of this software
// is on your own risk. The installation and usage (starting/running) of this
// software may allow other people or application to access your computer and
// any attached devices and is highly dependent on the configuration of the
// software which must be done by the user of the software; the author(s) is
// (are) also not responsible for proper configuration and usage of the
// software, even if provoked by documentation provided together with
// the software.
//
// Any changes to this file according to the GPL as documented in the file
// gpl.txt aside this file in the shipment you received can be done to the
// lines that follows this copyright notice here, but changes must not be
// done inside the copyright notive above. A re-distribution must contain
// the intact and unchanged copyright notice.
// Contributions and changes to the program code must be marked as such.
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
package de.anomic.plasma;
import java.io.File;
import java.io.IOException;
import java.net.InetAddress;
import java.net.MalformedURLException;
import java.util.Date;
import java.util.Iterator;
import java.util.LinkedList;
import org.apache.commons.pool.impl.GenericObjectPool;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
import de.anomic.data.robotsParser;
import de.anomic.http.httpc;
import de.anomic.plasma.plasmaURL;
import de.anomic.index.indexURLEntry;
import de.anomic.kelondro.kelondroCache;
import de.anomic.kelondro.kelondroException;
import de.anomic.kelondro.kelondroFlexTable;
import de.anomic.kelondro.kelondroFlexWidthArray;
import de.anomic.kelondro.kelondroIndex;
import de.anomic.kelondro.kelondroRow;
import de.anomic.kelondro.kelondroRowSet;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
import de.anomic.kelondro.kelondroTree;
import de.anomic.net.URL;
import de.anomic.plasma.urlPattern.plasmaURLPattern;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
import de.anomic.server.serverSemaphore;
import de.anomic.server.logging.serverLog;
import de.anomic.yacy.yacyCore;
public final class plasmaCrawlStacker {
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
// keys for different database types
public static final int QUEUE_DB_TYPE_RAM = 0;
public static final int QUEUE_DB_TYPE_TREE = 1;
public static final int QUEUE_DB_TYPE_FLEX = 2;
final WorkerPool theWorkerPool;
private GenericObjectPool.Config theWorkerPoolConfig = null;
final ThreadGroup theWorkerThreadGroup = new ThreadGroup("stackCrawlThreadGroup");
final serverLog log = new serverLog("STACKCRAWL");
final plasmaSwitchboard sb;
//private boolean stopped = false;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
private stackCrawlQueue queue;
public plasmaCrawlStacker(plasmaSwitchboard sb, File dbPath, long preloadTime, int dbtype) {
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
this.sb = sb;
this.queue = new stackCrawlQueue(dbPath, preloadTime, dbtype);
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
this.log.logInfo(this.queue.size() + " entries in the stackCrawl queue.");
this.log.logInfo("STACKCRAWL thread initialized.");
// configuring the thread pool
// implementation of session thread pool
this.theWorkerPoolConfig = new GenericObjectPool.Config();
// The maximum number of active connections that can be allocated from pool at the same time,
// 0 for no limit
this.theWorkerPoolConfig.maxActive = Integer.parseInt(sb.getConfig("stacker.MaxActiveThreads","50"));
// The maximum number of idle connections connections in the pool
// 0 = no limit.
this.theWorkerPoolConfig.maxIdle = Integer.parseInt(sb.getConfig("stacker.MaxIdleThreads","10"));
this.theWorkerPoolConfig.minIdle = Integer.parseInt(sb.getConfig("stacker.MinIdleThreads","5"));
// block undefinitely
this.theWorkerPoolConfig.maxWait = -1;
// Action to take in case of an exhausted DBCP statement pool
// 0 = fail, 1 = block, 2= grow
this.theWorkerPoolConfig.whenExhaustedAction = GenericObjectPool.WHEN_EXHAUSTED_BLOCK;
this.theWorkerPoolConfig.minEvictableIdleTimeMillis = 30000;
//this.theWorkerPoolConfig.timeBetweenEvictionRunsMillis = 30000;
// creating worker pool
this.theWorkerPool = new WorkerPool(new WorkterFactory(this.theWorkerThreadGroup),this.theWorkerPoolConfig);
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
public GenericObjectPool.Config getPoolConfig() {
return this.theWorkerPoolConfig;
}
public int getDBType() {
return this.queue.getDBType();
}
public void setPoolConfig(GenericObjectPool.Config newConfig) {
this.theWorkerPool.setConfig(newConfig);
}
public void close() {
try {
this.log.logFine("Shutdown. Terminating worker threads.");
if (this.theWorkerPool != null) this.theWorkerPool.close();
} catch (Exception e1) {
this.log.logSevere("Unable to shutdown all remaining stackCrawl threads", e1);
}
try {
this.log.logFine("Shutdown. Closing stackCrawl queue.");
if (this.queue != null) this.queue.close();
this.queue = null;
} catch (IOException e) {
this.log.logSevere("DB could not be closed properly.", e);
}
}
public int getNumActiveWorker() {
return this.theWorkerPool.getNumActive();
}
public int getNumIdleWorker() {
return this.theWorkerPool.getNumIdle();
}
public int size() {
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return this.queue.size();
}
public void job() {
try {
// getting a new message from the crawler queue
checkInterruption();
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
plasmaCrawlEntry theMsg = this.queue.waitForMessage();
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
if (theMsg != null) {
// getting a free session thread from the pool
checkInterruption();
Worker worker = (Worker) this.theWorkerPool.borrowObject();
// processing the new request
worker.execute(theMsg);
}
} catch (Exception e) {
if (e instanceof InterruptedException) {
this.log.logFine("Interruption detected.");
} else if ((e instanceof IllegalStateException) &&
(e.getMessage() != null) &&
(e.getMessage().indexOf("Pool not open") >= -1)) {
this.log.logFine("Pool was closed.");
} else {
this.log.logSevere("plasmaStackCrawlThread.run/loop", e);
}
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
}
public void enqueue(
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
URL nexturl,
String referrerhash,
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
String initiatorHash,
String name,
Date loadDate,
int currentdepth,
plasmaCrawlProfile.entry profile) {
if (profile != null) try {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
this.queue.addMessage(new plasmaCrawlEntry(
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
initiatorHash,
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
nexturl,
referrerhash,
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
name,
loadDate,
profile.handle(),
currentdepth,
0,
0
));
} catch (Exception e) {
e.printStackTrace();
}
}
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
public String dequeue(plasmaCrawlEntry theMsg) throws InterruptedException {
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
plasmaCrawlProfile.entry profile = this.sb.profiles.getEntry(theMsg.profileHandle());
if (profile == null) {
String errorMsg = "LOST PROFILE HANDLE '" + theMsg.profileHandle() + "' for URL " + theMsg.url();
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
this.log.logSevere(errorMsg);
throw new IllegalStateException(errorMsg);
}
return stackCrawl(
theMsg.url().toString(),
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
theMsg.referrerhash(),
theMsg.initiator(),
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
theMsg.name(),
theMsg.loaddate(),
theMsg.depth(),
profile);
}
public void checkInterruption() throws InterruptedException {
Thread curThread = Thread.currentThread();
if (curThread.isInterrupted()) throw new InterruptedException("Shutdown in progress ...");
}
public String stackCrawl(String nexturlString, String referrerString, String initiatorHash, String name, Date loadDate, int currentdepth, plasmaCrawlProfile.entry profile) throws InterruptedException {
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
// stacks a crawl item. The position can also be remote
// returns null if successful, a reason string if not successful
//this.log.logFinest("stackCrawl: nexturlString='" + nexturlString + "'");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
long startTime = System.currentTimeMillis();
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
String reason = null; // failure reason
// getting the initiator peer hash
if ((initiatorHash == null) || (initiatorHash.length() == 0)) initiatorHash = plasmaURL.dummyHash;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
// strange errors
if (nexturlString == null) {
reason = plasmaCrawlEURL.DENIED_URL_NULL;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
this.log.logSevere("Wrong URL in stackCrawl: url=null");
return reason;
}
// getting the referer url and url hash
URL referrerURL = null;
if (referrerString != null) {
try {
referrerURL = new URL(referrerString);
} catch (MalformedURLException e) {
referrerURL = null;
referrerString = null;
}
}
String referrerHash = (referrerString==null)?null:plasmaURL.urlHash(referrerString);
// check for malformed urls
URL nexturl = null;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
try {
nexturl = new URL(nexturlString);
} catch (MalformedURLException e) {
reason = plasmaCrawlEURL.DENIED_MALFORMED_URL;
this.log.logSevere("Wrong URL in stackCrawl: " + nexturlString +
". Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return reason;
}
// check if the protocol is supported
String urlProtocol = nexturl.getProtocol();
if (!this.sb.cacheLoader.isSupportedProtocol(urlProtocol)) {
reason = plasmaCrawlEURL.DENIED_UNSUPPORTED_PROTOCOL;
this.log.logSevere("Unsupported protocol in URL '" + nexturlString + "'. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
return reason;
}
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
// check if ip is local ip address
checkInterruption(); // TODO: this is protocol specific
InetAddress hostAddress = httpc.dnsResolve(nexturl.getHost());
if(this.sb.getConfig("yacyDebugMode", "true").equals("true")){
//just ignore the check in debugmode (useful for tor(.eff.org)
}else if (hostAddress == null) {
// if a http proxy is configured name resolution may not work
if (this.sb.remoteProxyConfig == null || !this.sb.remoteProxyConfig.useProxy()) {
reason = plasmaCrawlEURL.DENIED_UNKNOWN_HOST;
this.log.logFine("Unknown host in URL '" + nexturlString + "'. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
return reason;
}
} else if (hostAddress.isSiteLocalAddress()) {
reason = plasmaCrawlEURL.DENIED_PRIVATE_IP_ADDRESS;
this.log.logFine("Host in URL '" + nexturlString + "' has private IP address. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
return reason;
} else if (hostAddress.isLoopbackAddress()) {
reason = plasmaCrawlEURL.DENIED_LOOPBACK_IP_ADDRESS;
this.log.logFine("Host in URL '" + nexturlString + "' has loopback IP address. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
return reason;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
// check blacklist
checkInterruption();
if (plasmaSwitchboard.urlBlacklist.isListed(plasmaURLPattern.BLACKLIST_CRAWLER,nexturl)) {
reason = plasmaCrawlEURL.DENIED_URL_IN_BLACKLIST;
this.log.logFine("URL '" + nexturlString + "' is in blacklist. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return reason;
}
// filter deny
if ((currentdepth > 0) && (profile != null) && (!(nexturlString.matches(profile.generalFilter())))) {
reason = plasmaCrawlEURL.DENIED_URL_DOES_NOT_MATCH_FILTER;
this.log.logFine("URL '" + nexturlString + "' does not match crawling filter '" + profile.generalFilter() + "'. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return reason;
}
// deny cgi
if (plasmaHTCache.isCGI(nexturlString)) {
reason = plasmaCrawlEURL.DENIED_CGI_URL;
this.log.logFine("URL '" + nexturlString + "' is CGI URL. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return reason;
}
// deny post properties
if ((plasmaHTCache.isPOST(nexturlString)) && (profile != null) && (!(profile.crawlingQ()))) {
reason = plasmaCrawlEURL.DENIED_POST_URL;
this.log.logFine("URL '" + nexturlString + "' is post URL. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return reason;
}
// add domain to profile domain list
if ((profile.domFilterDepth() != Integer.MAX_VALUE) || (profile.domMaxPages() != Integer.MAX_VALUE)) {
profile.domInc(nexturl.getHost(), (referrerURL == null) ? null : referrerURL.getHost().toLowerCase(), currentdepth);
}
// deny urls that do not match with the profile domain list
if (!(profile.grantedDomAppearance(nexturl.getHost()))) {
reason = plasmaCrawlEURL.DENIED_NO_MATCH_WITH_DOMAIN_FILTER;
this.log.logFine("URL '" + nexturlString + "' is not listed in granted domains. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
return reason;
}
// deny urls that exceed allowed number of occurrences
if (!(profile.grantedDomCount(nexturl.getHost()))) {
reason = plasmaCrawlEURL.DENIED_DOMAIN_COUNT_EXCEEDED;
this.log.logFine("URL '" + nexturlString + "' appeared too often, a maximum of " + profile.domMaxPages() + " is allowed. "+
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
return reason;
}
// check if the url is double registered
checkInterruption();
String nexturlhash = plasmaURL.urlHash(nexturl);
String dbocc = this.sb.urlExists(nexturlhash);
indexURLEntry oldEntry = null;
oldEntry = this.sb.wordIndex.loadedURL.load(nexturlhash, null);
boolean recrawl = (oldEntry != null) && (((System.currentTimeMillis() - oldEntry.loaddate().getTime()) / 60000) > profile.recrawlIfOlder());
// FIXME: this does not work correctly?
if ((dbocc != null) && (!(recrawl))) {
reason = plasmaCrawlEURL.DOUBLE_REGISTERED + dbocc + ")";
//this.log.logFine("URL '" + nexturlString + "' is double registered in '" + dbocc + "'. " + "Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return reason;
}
// checking robots.txt for http(s) resources
checkInterruption();
if ((urlProtocol.equals("http") || urlProtocol.equals("https")) && robotsParser.isDisallowed(nexturl)) {
reason = plasmaCrawlEURL.DENIED_ROBOTS_TXT;
this.log.logFine("Crawling of URL '" + nexturlString + "' disallowed by robots.txt. " +
"Stack processing time: " + (System.currentTimeMillis()-startTime) + "ms");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return reason;
}
// show potential re-crawl
if (recrawl) {
this.log.logFine("RE-CRAWL of URL '" + nexturlString + "': this url was crawled " +
((System.currentTimeMillis() - oldEntry.loaddate().getTime()) / 60000 / 60 / 24) + " days ago.");
}
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
// store information
boolean local = ((initiatorHash.equals(plasmaURL.dummyHash)) || (initiatorHash.equals(yacyCore.seedDB.mySeed.hash)));
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
boolean global =
(profile != null) &&
(profile.remoteIndexing()) /* granted */ &&
(currentdepth == profile.generalDepth()) /* leaf node */ &&
//(initiatorHash.equals(yacyCore.seedDB.mySeed.hash)) /* not proxy */ &&
(
(yacyCore.seedDB.mySeed.isSenior()) ||
(yacyCore.seedDB.mySeed.isPrincipal())
) /* qualified */;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
if ((!local)&&(!global)&&(!profile.handle().equals(this.sb.defaultRemoteProfile.handle()))) {
this.log.logSevere("URL '" + nexturlString + "' can neither be crawled local nor global.");
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
// add the url into the crawling queue
checkInterruption();
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
plasmaCrawlEntry ne = new plasmaCrawlEntry(initiatorHash, /* initiator, needed for p2p-feedback */
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
nexturl, /* url clear text string */
referrerHash, /* last url in crawling queue */
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
name, /* load date */
loadDate, /* the anchor name */
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
(profile == null) ? null : profile.handle(), // profile must not be null!
currentdepth, /*depth so far*/
0, /*anchors, default value */
0 /*forkfactor, default value */
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
);
this.sb.noticeURL.push(
((global) ? plasmaCrawlNURL.STACK_TYPE_LIMIT :
((local) ? plasmaCrawlNURL.STACK_TYPE_CORE : plasmaCrawlNURL.STACK_TYPE_REMOTE)) /*local/remote stack*/,
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
ne);
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
return null;
}
final class stackCrawlQueue {
private final serverSemaphore readSync;
private final serverSemaphore writeSync;
private final LinkedList urlEntryHashCache;
private kelondroIndex urlEntryCache;
private File cacheStacksPath;
private long preloadTime;
private int dbtype;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
public stackCrawlQueue(File cacheStacksPath, long preloadTime, int dbtype) {
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
// init the read semaphore
this.readSync = new serverSemaphore (0);
// init the write semaphore
this.writeSync = new serverSemaphore (1);
// init the message list
this.urlEntryHashCache = new LinkedList();
// create a stack for newly entered entries
this.cacheStacksPath = cacheStacksPath;
this.preloadTime = preloadTime;
this.dbtype = dbtype;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
openDB();
try {
// loop through the list and fill the messageList with url hashs
Iterator rows = this.urlEntryCache.rows(true, null);
kelondroRow.Entry entry;
while (rows.hasNext()) {
entry = (kelondroRow.Entry) rows.next();
if (entry == null) {
System.out.println("ERROR! null element found");
continue;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
this.urlEntryHashCache.add(entry.getColString(0, null));
this.readSync.V();
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
} catch (kelondroException e) {
/* if we have an error, we start with a fresh database */
plasmaCrawlStacker.this.log.logSevere("Unable to initialize crawl stacker queue, kelondroException:" + e.getMessage() + ". Reseting DB.\n", e);
// deleting old db and creating a new db
try {this.urlEntryCache.close();} catch (Exception ex) {}
deleteDB();
openDB();
} catch (IOException e) {
/* if we have an error, we start with a fresh database */
plasmaCrawlStacker.this.log.logSevere("Unable to initialize crawl stacker queue, IOException:" + e.getMessage() + ". Reseting DB.\n", e);
// deleting old db and creating a new db
try {this.urlEntryCache.close();} catch (Exception ex) {}
deleteDB();
openDB();
}
}
private void deleteDB() {
if (this.dbtype == QUEUE_DB_TYPE_RAM) {
// do nothing..
}
if (this.dbtype == QUEUE_DB_TYPE_FLEX) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
kelondroFlexWidthArray.delete(cacheStacksPath, "urlNoticeStacker7.db");
}
if (this.dbtype == QUEUE_DB_TYPE_TREE) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
File cacheFile = new File(cacheStacksPath, "urlNoticeStacker.db");
cacheFile.delete();
}
}
private void openDB() {
if (!(cacheStacksPath.exists())) cacheStacksPath.mkdir(); // make the path
if (this.dbtype == QUEUE_DB_TYPE_RAM) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
this.urlEntryCache = new kelondroRowSet(plasmaCrawlEntry.rowdef, 0);
}
if (this.dbtype == QUEUE_DB_TYPE_FLEX) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
String newCacheName = "urlNoticeStacker7.db";
cacheStacksPath.mkdirs();
try {
this.urlEntryCache = new kelondroCache(new kelondroFlexTable(cacheStacksPath, newCacheName, preloadTime, plasmaCrawlEntry.rowdef, true), true, false);
} catch (Exception e) {
e.printStackTrace();
// kill DB and try again
kelondroFlexTable.delete(cacheStacksPath, newCacheName);
try {
this.urlEntryCache = new kelondroCache(new kelondroFlexTable(cacheStacksPath, newCacheName, preloadTime, plasmaCrawlEntry.rowdef, true), true, false);
} catch (Exception ee) {
ee.printStackTrace();
System.exit(-1);
}
}
}
if (this.dbtype == QUEUE_DB_TYPE_TREE) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
File cacheFile = new File(cacheStacksPath, "urlNoticeStacker.db");
cacheFile.getParentFile().mkdirs();
this.urlEntryCache = new kelondroCache(kelondroTree.open(cacheFile, true, preloadTime, plasmaCrawlEntry.rowdef), true, true);
}
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
public void close() throws IOException {
// closing the db
this.urlEntryCache.close();
// clearing the hash list
this.urlEntryHashCache.clear();
}
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
public void addMessage(plasmaCrawlEntry newMessage)
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
throws InterruptedException, IOException {
if (newMessage == null) throw new NullPointerException();
this.writeSync.P();
try {
boolean insertionDoneSuccessfully = false;
synchronized(this.urlEntryHashCache) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
kelondroRow.Entry oldValue = this.urlEntryCache.put(newMessage.toRow());
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
if (oldValue == null) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
insertionDoneSuccessfully = this.urlEntryHashCache.add(newMessage.urlhash());
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
}
if (insertionDoneSuccessfully) {
this.readSync.V();
}
} finally {
this.writeSync.V();
}
}
public int size() {
synchronized(this.urlEntryHashCache) {
return this.urlEntryHashCache.size();
}
}
public int getDBType() {
return this.dbtype;
}
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
public plasmaCrawlEntry waitForMessage() throws InterruptedException, IOException {
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
this.readSync.P();
this.writeSync.P();
if (this.urlEntryHashCache.size() == 0) return null;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
String urlHash = null;
kelondroRow.Entry entry = null;
try {
synchronized(this.urlEntryHashCache) {
urlHash = (String) this.urlEntryHashCache.removeFirst();
if (urlHash == null) throw new IOException("urlHash is null");
entry = this.urlEntryCache.remove(urlHash.getBytes());
}
} finally {
this.writeSync.V();
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
if ((urlHash == null) || (entry == null)) return null;
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
return new plasmaCrawlEntry(entry);
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
}
public final class WorkterFactory implements org.apache.commons.pool.PoolableObjectFactory {
final ThreadGroup workerThreadGroup;
public WorkterFactory(ThreadGroup theWorkerThreadGroup) {
super();
if (theWorkerThreadGroup == null)
throw new IllegalArgumentException("The threadgroup object must not be null.");
this.workerThreadGroup = theWorkerThreadGroup;
*) Asynchronous queuing of crawl job URLs (stackCrawl) various checks like the blacklist check or the robots.txt disallow check are now done by a separate thread to unburden the indexer thread(s) TODO: maybe we have to introduce a threadpool here if it turn out that this single thread is a bottleneck because of the time consuming robots.txt downloads *) improved index transfer The index selection and transmission is done in parallel now to improve index transfer performance. TODO: maybe we could speed up performance by unsing multiple transmission threads in parallel instead of only a single one. *) gzip encoded post requests it is now configureable if a gzip encoded post request should be send on intex transfer/distribution *) storage Peer (very experimentell and not optimized yet) Now it's possible to send the result of the yacy indexer thread to a remote peer istead of storing the indexed words locally. This could be done by setting the property "storagePeerHash" in the yacy config file - Please note that if the index transfer fails, the index ist stored locally. - TODO: currently this index transfer is done by the indexer thread. To seedup the indexer a) this transmission should be done in parallel and b) multiple chunks should be bundled and transfered together *) general performance improvements - better memory cleanup after http request processing has finished - replacing some string concatenations with stringBuffers - replacing BufferedInputStreams with serverByteBuffer - replacing vectors with arraylists wherever possible - replacing hashtables with hashmaps wherever possible This was done because function calls to verctor or hashtable functions take 3 time longer than calls to functions of arraylists or hashmaps. TODO: we should take a look on the class serverObject which is inherited from hashmap Do we realy need a synchronization for this class? TODO: replace arraylists with linkedLists if random access to the list elements is not needed *) Robots Parser supports if-modified-since downloads now If the downloaded robots.txt file is older than 7 days the robots parser tries to download the robots.txt with the if-modified-since header to avoid unnecessary downloads if the file was not changed. Additionally the ETag header is used to detect changes. *) Crawler: better handling of unsupported mimeTypes + FileExtension *) Bugfix: plasmaWordIndexEntity was not closed correctly in - query.java - plasmaswitchboard.java *) function minimizeUrlDB added to yacy.java this function tests the current urlHashDB for unused urls ATTENTION: please don't use this function at the moment because it causes the wordIndexDB to flush all words into the word directory! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@853 6c8d7289-2bf4-0310-a012-ef5d649a1542
2005-10-05 12:45:33 +02:00
}
public Object makeObject() {
Worker newWorker = new Worker(this.workerThreadGroup);
newWorker.setPriority(Thread.MAX_PRIORITY);
return newWorker;
}
/**
* @see org.apache.commons.pool.PoolableObjectFactory#destroyObject(java.lang.Object)
*/
public void destroyObject(Object obj) {
if (obj instanceof Worker) {
Worker theWorker = (Worker) obj;
synchronized(theWorker) {
theWorker.setName("stackCrawlThread_destroyed");
theWorker.destroyed = true;
theWorker.setStopped(true);
theWorker.interrupt();
}
}
}
/**
* @see org.apache.commons.pool.PoolableObjectFactory#validateObject(java.lang.Object)
*/
public boolean validateObject(Object obj) {
return true;
}
/**
* @param obj
*
*/
public void activateObject(Object obj) {
//log.debug(" activateObject...");
}
/**
* @param obj
*
*/
public void passivateObject(Object obj) {
//log.debug(" passivateObject..." + obj);
// if (obj instanceof Session) {
// Session theSession = (Session) obj;
// }
}
}
public final class WorkerPool extends GenericObjectPool {
public boolean isClosed = false;
/**
* First constructor.
* @param objFactory
*/
public WorkerPool(WorkterFactory objFactory) {
super(objFactory);
this.setMaxIdle(10); // Maximum idle threads.
this.setMaxActive(50); // Maximum active threads.
this.setMinEvictableIdleTimeMillis(30000); //Evictor runs every 30 secs.
//this.setMaxWait(1000); // Wait 1 second till a thread is available
}
public WorkerPool(plasmaCrawlStacker.WorkterFactory objFactory,
GenericObjectPool.Config config) {
super(objFactory, config);
}
public Object borrowObject() throws Exception {
return super.borrowObject();
}
public void returnObject(Object obj) {
if (obj == null) return;
if (obj instanceof Worker) {
try {
((Worker)obj).setName("stackCrawlThread_inPool");
super.returnObject(obj);
} catch (Exception e) {
((Worker)obj).setStopped(true);
serverLog.logSevere("STACKCRAWL-POOL","Unable to return stackcrawl thread to pool.",e);
}
} else {
serverLog.logSevere("STACKCRAWL-POOL","Object of wront type '" + obj.getClass().getName() +
"' returned to pool.");
}
}
public void invalidateObject(Object obj) {
if (obj == null) return;
if (this.isClosed) return;
if (obj instanceof Worker) {
try {
((Worker)obj).setName("stackCrawlThread_invalidated");
((Worker)obj).setStopped(true);
super.invalidateObject(obj);
} catch (Exception e) {
serverLog.logSevere("STACKCRAWL-POOL","Unable to invalidate stackcrawl thread.",e);
}
}
}
public synchronized void close() throws Exception {
/*
* shutdown all still running session threads ...
*/
this.isClosed = true;
/* waiting for all threads to finish */
int threadCount = theWorkerThreadGroup.activeCount();
Thread[] threadList = new Thread[threadCount];
threadCount = theWorkerThreadGroup.enumerate(threadList);
try {
// trying to gracefull stop all still running sessions ...
log.logInfo("Signaling shutdown to " + threadCount + " remaining stackCrawl threads ...");
for ( int currentThreadIdx = 0; currentThreadIdx < threadCount; currentThreadIdx++ ) {
Thread currentThread = threadList[currentThreadIdx];
if (currentThread.isAlive()) {
((Worker)currentThread).setStopped(true);
}
}
// waiting a frew ms for the session objects to continue processing
try { Thread.sleep(500); } catch (InterruptedException ex) {}
// interrupting all still running or pooled threads ...
log.logInfo("Sending interruption signal to " + theWorkerThreadGroup.activeCount() + " remaining stackCrawl threads ...");
theWorkerThreadGroup.interrupt();
// if there are some sessions that are blocking in IO, we simply close the socket
log.logFine("Trying to abort " + theWorkerThreadGroup.activeCount() + " remaining stackCrawl threads ...");
for ( int currentThreadIdx = 0; currentThreadIdx < threadCount; currentThreadIdx++ ) {
Thread currentThread = threadList[currentThreadIdx];
if (currentThread.isAlive()) {
log.logInfo("Trying to shutdown stackCrawl thread '" + currentThread.getName() + "' [" + currentThreadIdx + "].");
((Worker)currentThread).close();
}
}
// we need to use a timeout here because of missing interruptable session threads ...
log.logFine("Waiting for " + theWorkerThreadGroup.activeCount() + " remaining stackCrawl threads to finish shutdown ...");
for ( int currentThreadIdx = 0; currentThreadIdx < threadCount; currentThreadIdx++ ) {
Thread currentThread = threadList[currentThreadIdx];
if (currentThread.isAlive()) {
log.logFine("Waiting for stackCrawl thread '" + currentThread.getName() + "' [" + currentThreadIdx + "] to finish shutdown.");
try { currentThread.join(500); } catch (InterruptedException ex) {}
}
}
log.logInfo("Shutdown of remaining stackCrawl threads finished.");
} catch (Exception e) {
log.logSevere("Unexpected error while trying to shutdown all remaining stackCrawl threads.",e);
}
super.close();
}
}
public final class Worker extends Thread {
boolean destroyed = false;
private boolean running = false;
private boolean stopped = false;
private boolean done = false;
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
private plasmaCrawlEntry theMsg;
public Worker(ThreadGroup theThreadGroup) {
super(theThreadGroup,"stackCrawlThread_created");
}
public void setStopped(boolean stopped) {
this.stopped = stopped;
}
public void close() {
if (this.isAlive()) {
try {
// trying to close all still open httpc-Sockets first
int closedSockets = httpc.closeOpenSockets(this);
if (closedSockets > 0) {
log.logInfo(closedSockets + " HTTP-client sockets of thread '" + this.getName() + "' closed.");
}
} catch (Exception e) {}
}
}
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
public synchronized void execute(plasmaCrawlEntry newMsg) {
this.theMsg = newMsg;
this.done = false;
if (!this.running) {
// this.setDaemon(true);
this.start();
} else {
this.notifyAll();
}
}
public void reset() {
this.done = true;
this.theMsg = null;
}
public boolean isRunning() {
return this.running;
}
public void run() {
this.running = true;
try {
// The thread keeps running.
while (!this.stopped && !this.isInterrupted() && !plasmaCrawlStacker.this.theWorkerPool.isClosed) {
if (this.done) {
synchronized (this) {
// return thread back into pool
plasmaCrawlStacker.this.theWorkerPool.returnObject(this);
// We are waiting for a new task now.
if (!this.stopped && !this.destroyed && !this.isInterrupted()) {
this.wait();
}
}
} else {
try {
// executing the new task
execute();
} finally {
// reset thread
reset();
}
}
}
} catch (InterruptedException ex) {
serverLog.logFiner("STACKCRAWL-POOL","Interruption of thread '" + this.getName() + "' detected.");
} finally {
if (plasmaCrawlStacker.this.theWorkerPool != null && !this.destroyed)
plasmaCrawlStacker.this.theWorkerPool.invalidateObject(this);
}
}
private void execute() throws InterruptedException {
try {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
this.setName("stackCrawlThread_" + this.theMsg.url());
String rejectReason = dequeue(this.theMsg);
// check for interruption
checkInterruption();
// if the url was rejected we store it into the error URL db
if (rejectReason != null) {
redesigned NURL-handling: - the general NURL-index for all crawl stack types was splitted into separate indexes for these stacks - the new NURL-index is managed by the crawl balancer - the crawl balancer does not need an internal index any more, it is replaced by the NURL-index - the NURL.Entry was generalized and is now a new class plasmaCrawlEntry - the new class plasmaCrawlEntry replaces also the preNURL.Entry class, and will also replace the switchboardEntry class in the future - the new class plasmaCrawlEntry is more accurate for date entries (holds milliseconds) and can contain larger 'name' entries (anchor tag names) - the EURL object was replaced by a new ZURL object, which is a container for the plasmaCrawlEntry and some tracking information - the EURL index is now filled with ZURL objects - a new index delegatedURL holds ZURL objects about plasmaCrawlEntry obects to track which url is handed over to other peers - redesigned handling of plasmaCrawlEntry - handover, because there is no need any more to convert one entry object into another - found and fixed numerous bugs in the context of crawl state handling - fixed a serious bug in kelondroCache which caused that entries could not be removed - fixed some bugs in online interface and adopted monitor output to new entry objects - adopted yacy protocol to handle new delegatedURL entries all old crawl queues will disappear after this update! git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3483 6c8d7289-2bf4-0310-a012-ef5d649a1542
2007-03-16 14:25:56 +01:00
plasmaCrawlZURL.Entry ee = sb.errorURL.newEntry(
this.theMsg, yacyCore.seedDB.mySeed.hash, null,
0, rejectReason);
ee.store();
sb.errorURL.stackPushEntry(ee);
}
} catch (Exception e) {
if (e instanceof InterruptedException) throw (InterruptedException) e;
plasmaCrawlStacker.this.log.logWarning("Error while processing stackCrawl entry.\n" +
"Entry: " + this.theMsg.toString() +
"Error: " + e.toString(),e);
} finally {
this.done = true;
}
}
}
}