深入剖析SolrCloud（四）
在上一篇中介绍了连接Zookeeper集群的方法，这一篇将围绕一个有趣的话题---来展开，这就是Replication（索引复制），关于Solr Replication的详细介绍，可以参考http://wiki.apache.org/solr/SolrReplication。
在开始这个话题之前，先从我最近在应用中引入solr的master/slave架构时，遇到的一个让我困扰的实际问题。
应用场景简单描述如下：
1）首先master节点下载索引分片，然后创建配置文件，加入master节点的replication配置片段，再对索引分片进行合并(关于mergeIndex，可以参考http://wiki.apache.org/solr/MergingSolrIndexes)，然后利用上述配置文件和索引数据去创建一个solr核。
2）slave节点创建配置文件，加入slave节点的replication配置片段，创建一个空的solr核，等待从master节点进行索引数据同步
出现的问题：slave节点没有从master节点同步到数据。
问题分析：
1）首先检查master节点，获取最新的可复制索引的版本号，
http://master_host:port/solr/replication?command=indexversion
发现返回的索引版本号是0，这说明mater节点根本没有触发replication动作，
2）为了确认上述判断，在slave节点上进一步查看replication的详细信息
http://slave_host:port/solr/replication?command=details
发现确实如此，尽管master节点的索引版本号和slave节点的索引版本号不一致，但索引却没有同步过来，再分别查看master节点和slave节点的日志，发现索引复制动作确实没有开始。
综上所述，确实是master节点没有触发索引复制动作，那究竟是为何呢？先将原因摆出来，后面会通过源码的分析来加以说明。
原因：solr合并索引时，不管你是通过mergeindexes的http命令，还是调用底层lucene的IndexWriter，记得最后一定要提交一个commit，否则，不仅索引不仅不会对查询可见，更是对于master/slave架构的solr集群来说，master节点的replication动作不会触发，因为indexversion没有感知到变化。
好了，下面开始对Solr的Replication的分析。
Solr容器在加载solr核的时候，会对已经注册的各个实现SolrCoreAware接口的Handler进行回调，调用其inform方法。
对于ReplicationHandler来说，就是在这里对自己是属于master节点还是slave节点进行判断，若是slave节点，则创建一个SnapPuller对象，定时负责从master节点主动拉索引数据下来；若是master节点，则只设置相应的参数。
this.core = core;
registerFileStreamResponseWriter();
registerCloseHook();
NamedList slave = (NamedList) initArgs.get("slave");
boolean enableSlave = isEnabled( slave );
if (enableSlave) {
tempSnapPuller = snapPuller = new SnapPuller(slave, this, core);
isSlave = true;
}
NamedList master = (NamedList) initArgs.get("master");
boolean enableMaster = isEnabled( master );
if (!enableSlave && !enableMaster) {
enableMaster = true;
master = new NamedList<Object>();
}
if (enableMaster) {
includeConfFiles = (String) master.get(CONF_FILES);
if (includeConfFiles != null && includeConfFiles.trim().length() > 0) {
List<String> files = Arrays.asList(includeConfFiles.split(","));
for (String file : files) {
if (file.trim().length() == 0) continue;
String[] strs = file.split(":");
// if there is an alias add it or it is null
confFileNameAlias.add(strs[0], strs.length > 1 ? strs[1] : null);
}
LOG.info("Replication enabled for following config files: " + includeConfFiles);
}
List backup = master.getAll("backupAfter");
boolean backupOnCommit = backup.contains("commit");
boolean backupOnOptimize = !backupOnCommit && backup.contains("optimize");
List replicateAfter = master.getAll(REPLICATE_AFTER);
replicateOnCommit = replicateAfter.contains("commit");
replicateOnOptimize = !replicateOnCommit && replicateAfter.contains("optimize");
if (!replicateOnCommit && ! replicateOnOptimize) {
replicateOnCommit = true;
}
// if we only want to replicate on optimize, we need the deletion policy to
// save the last optimized commit point.
if (replicateOnOptimize) {
IndexDeletionPolicyWrapper wrapper = core.getDeletionPolicy();
IndexDeletionPolicy policy = wrapper == null ? null : wrapper.getWrappedDeletionPolicy();
if (policy instanceof SolrDeletionPolicy) {
SolrDeletionPolicy solrPolicy = (SolrDeletionPolicy)policy;
if (solrPolicy.getMaxOptimizedCommitsToKeep() < 1) {
solrPolicy.setMaxOptimizedCommitsToKeep(1);
}
} else {
LOG.warn("Replication can't call setMaxOptimizedCommitsToKeep on " + policy);
}
}
if (replicateOnOptimize || backupOnOptimize) {
core.getUpdateHandler().registerOptimizeCallback(getEventListener(backupOnOptimize, replicateOnOptimize));
}
if (replicateOnCommit || backupOnCommit) {
replicateOnCommit = true;
core.getUpdateHandler().registerCommitCallback(getEventListener(backupOnCommit, replicateOnCommit));
}
if (replicateAfter.contains("startup")) {
replicateOnStart = true;
RefCounted<SolrIndexSearcher> s = core.getNewestSearcher(false);
try {
DirectoryReader reader = s==null ? null : s.get().getIndexReader();
if (reader!=null && reader.getIndexCommit() != null && reader.getIndexCommit().getGeneration() != 1L) {
try {
if(replicateOnOptimize){
Collection<IndexCommit> commits = DirectoryReader.listCommits(reader.directory());
for (IndexCommit ic : commits) {
if(ic.getSegmentCount() == 1){
if(indexCommitPoint == null || indexCommitPoint.getGeneration() < ic.getGeneration()) indexCommitPoint = ic;
}
}
} else{
indexCommitPoint = reader.getIndexCommit();
}
} finally {
// We don't need to save commit points for replication, the SolrDeletionPolicy
// always saves the last commit point (and the last optimized commit point, if needed)
/***
if(indexCommitPoint != null){
core.getDeletionPolicy().saveCommitPoint(indexCommitPoint.getGeneration());
}
***/
}
}
// reboot the writer on the new index
core.getUpdateHandler().newIndexWriter();
} catch (IOException e) {
LOG.warn("Unable to get IndexCommit on startup", e);
} finally {
if (s!=null) s.decref();
}
}
String reserve = (String) master.get(RESERVE);
if (reserve != null && !reserve.trim().equals("")) {
reserveCommitDuration = SnapPuller.readInterval(reserve);
}
LOG.info("Commits will be reserved for " + reserveCommitDuration);
isMaster = true;
}
}
ReplicationHandler可以响应多种命令：
1) indexversion。
这里需要了解的第一个概念是索引提交点(IndexCommit)，这是底层lucene的东西，可以自行查阅资料。首先获取最新的索引提交点，然后从其中获取索引版本号和索引所属代。
if (commitPoint != null && replicationEnabled.get()) {
core.getDeletionPolicy().setReserveDuration(commitPoint.getVersion(), reserveCommitDuration);
rsp.add(CMD_INDEX_VERSION, commitPoint.getVersion());
rsp.add(GENERATION, commitPoint.getGeneration());
2）backup。这个命令用来对索引做快照。首先获取最新的索引提交点，然后创建做一个SnapShooter，具体的快照动作由这个对象完成，
try {
int numberToKeep = params.getInt(NUMBER_BACKUPS_TO_KEEP, Integer.MAX_VALUE);
IndexDeletionPolicyWrapper delPolicy = core.getDeletionPolicy();
IndexCommit indexCommit = delPolicy.getLatestCommit();
if(indexCommit == null) {
indexCommit = req.getSearcher().getReader().getIndexCommit();
}
// small race here before the commit point is saved
new SnapShooter(core, params.get("location")).createSnapAsync(indexCommit, numberToKeep, this);
} catch (Exception e) {
LOG.warn("Exception during creating a snapshot", e);
rsp.add("exception", e);
}
}
快照对象会启动一个线程去异步地做一个索引备份。
void createSnapAsync(final IndexCommit indexCommit, final int numberToKeep, final ReplicationHandler replicationHandler) {
replicationHandler.core.getDeletionPolicy().saveCommitPoint(indexCommit.getVersion());
new Thread() {
@Override
public void run() {
createSnapshot(indexCommit, numberToKeep, replicationHandler);
}
}.start();
}
void createSnapshot(final IndexCommit indexCommit, int numberToKeep, ReplicationHandler replicationHandler) {
NamedList details = new NamedList();
details.add("startTime", new Date().toString());
File snapShotDir = null;
String directoryName = null;
Lock lock = null;
try {
if(numberToKeep<Integer.MAX_VALUE) {
deleteOldBackups(numberToKeep);
}
SimpleDateFormat fmt = new SimpleDateFormat(DATE_FMT, Locale.US);
directoryName = "snapshot." + fmt.format(new Date());
lock = lockFactory.makeLock(directoryName + ".lock");
if (lock.isLocked()) return;
snapShotDir = new File(snapDir, directoryName);
if (!snapShotDir.mkdir()) {
LOG.warn("Unable to create snapshot directory: " + snapShotDir.getAbsolutePath());
return;
}
Collection<String> files = indexCommit.getFileNames();
FileCopier fileCopier = new FileCopier(solrCore.getDeletionPolicy(), indexCommit);
fileCopier.copyFiles(files, snapShotDir);
details.add("fileCount", files.size());
details.add("status", "success");
details.add("snapshotCompletedAt", new Date().toString());
} catch (Exception e) {
SnapPuller.delTree(snapShotDir);
LOG.error("Exception while creating snapshot", e);
details.add("snapShootException", e.getMessage());
} finally {
replicationHandler.core.getDeletionPolicy().releaseCommitPoint(indexCommit.getVersion());
replicationHandler.snapShootDetails = details;
if (lock != null) {
try {
lock.release();
} catch (IOException e) {
LOG.error("Unable to release snapshoot lock: " + directoryName + ".lock");
}
}
}
}
3）fetchindex。响应来自slave节点的取索引文件的请求，会启动一个线程来实现索引文件的获取。
String masterUrl = solrParams.get(MASTER_URL);
if (!isSlave && masterUrl == null) {
rsp.add(STATUS,ERR_STATUS);
rsp.add("message","No slave configured or no 'masterUrl' Specified");
return;
}
final SolrParams paramsCopy = new ModifiableSolrParams(solrParams);
new Thread() {
@Override
public void run() {
doFetch(paramsCopy);
}
}.start();
rsp.add(STATUS, OK_STATUS);
具体的获取动作是通过SnapPuller对象来实现的，首先尝试获取pull对象锁，如果请求锁失败，则说明还有取索引数据动作未结束，如果请求锁成功，就调用SnapPuller对象的fetchLatestIndex方法来取最新的索引数据。
void doFetch(SolrParams solrParams) {
String masterUrl = solrParams == null ? null : solrParams.get(MASTER_URL);
if (!snapPullLock.tryLock())
return;
try {
tempSnapPuller = snapPuller;
if (masterUrl != null) {
NamedList<Object> nl = solrParams.toNamedList();
nl.remove(SnapPuller.POLL_INTERVAL);
tempSnapPuller = new SnapPuller(nl, this, core);
}
tempSnapPuller.fetchLatestIndex(core);
} catch (Exception e) {
LOG.error("SnapPull failed ", e);
} finally {
tempSnapPuller = snapPuller;
snapPullLock.unlock();
}
}
最后真正的取索引数据过程，首先，若mastet节点的indexversion为0，则说明master节点根本没有提供可供复制的索引数据，若master节点和slave节点的indexversion相同，则说明slave节点目前与master节点索引数据状态保持一致，无需同步。若两者的indexversion不同，则开始索引复制过程，首先从master节点上下载指定索引版本号的索引文件列表，然后创建一个索引文件同步服务线程来完成同并工作。
这里需要区分的是，如果master节点的年代比slave节点要老，那就说明两者已经不相容，此时slave节点需要新建一个索引目录，再从master节点做一次全量索引复制。还需要注意的一点是，索引同步也是可以同步配置文件的，若配置文件发生变化，则需要对solr核进行一次reload操作。最对了，还有，和文章开头一样， slave节点同步完数据后，别忘了做一次commit操作，以便刷新自己的索引提交点到最新的状态。最后，关闭并等待同步服务线程结束。此外，具体的取索引文件是通过FileFetcher对象来完成。
boolean fetchLatestIndex(SolrCore core) throws IOException {
replicationStartTime = System.currentTimeMillis();
try {
//get the current 'replicateable' index version in the master
NamedList response = null;
try {
response = getLatestVersion();
} catch (Exception e) {
LOG.error("Master at: " + masterUrl + " is not available. Index fetch failed. Exception: " + e.getMessage());
return false;
}
long latestVersion = (Long) response.get(CMD_INDEX_VERSION);
long latestGeneration = (Long) response.get(GENERATION);
if (latestVersion == 0L) {
//there is nothing to be replicated
return false;
}
IndexCommit commit;
RefCounted<SolrIndexSearcher> searcherRefCounted = null;
try {
searcherRefCounted = core.getNewestSearcher(false);
commit = searcherRefCounted.get().getReader().getIndexCommit();
} finally {
if (searcherRefCounted != null)
searcherRefCounted.decref();
}
if (commit.getVersion() == latestVersion && commit.getGeneration() == latestGeneration) {
//master and slave are alsready in sync just return
LOG.info("Slave in sync with master.");
return false;
}
LOG.info("Master's version: " + latestVersion + ", generation: " + latestGeneration);
LOG.info("Slave's version: " + commit.getVersion() + ", generation: " + commit.getGeneration());
LOG.info("Starting replication process");
// get the list of files first
fetchFileList(latestVersion);
// this can happen if the commit point is deleted before we fetch the file list.
if(filesToDownload.isEmpty()) return false;
LOG.info("Number of files in latest index in master: " + filesToDownload.size());
// Create the sync service
fsyncService = Executors.newSingleThreadExecutor();
// use a synchronized list because the list is read by other threads (to show details)
filesDownloaded = Collections.synchronizedList(new ArrayList<Map<String, Object>>());
// if the generateion of master is older than that of the slave , it means they are not compatible to be copied
// then a new index direcory to be created and all the files need to be copied
boolean isFullCopyNeeded = commit.getGeneration() >= latestGeneration;
File tmpIndexDir = createTempindexDir(core);
if (isIndexStale())
isFullCopyNeeded = true;
successfulInstall = false;
boolean deleteTmpIdxDir = true;
File indexDir = null ;
try {
indexDir = new File(core.getIndexDir());
downloadIndexFiles(isFullCopyNeeded, tmpIndexDir, latestVersion);
LOG.info("Total time taken for download : " + ((System.currentTimeMillis() - replicationStartTime) / 1000) + " secs");
Collection<Map<String, Object>> modifiedConfFiles = getModifiedConfFiles(confFilesToDownload);
if (!modifiedConfFiles.isEmpty()) {
downloadConfFiles(confFilesToDownload, latestVersion);
if (isFullCopyNeeded) {
successfulInstall = modifyIndexProps(tmpIndexDir.getName());
deleteTmpIdxDir = false;
} else {
successfulInstall = copyIndexFiles(tmpIndexDir, indexDir);
}
if (successfulInstall) {
LOG.info("Configuration files are modified, core will be reloaded");
logReplicationTimeAndConfFiles(modifiedConfFiles, successfulInstall);//write to a file time of replication and conf files.
reloadCore();
}
} else {
terminateAndWaitFsyncService();
if (isFullCopyNeeded) {
successfulInstall = modifyIndexProps(tmpIndexDir.getName());
deleteTmpIdxDir = false;
} else {
successfulInstall = copyIndexFiles(tmpIndexDir, indexDir);
}
if (successfulInstall) {
logReplicationTimeAndConfFiles(modifiedConfFiles, successfulInstall);
doCommit();
}
}
replicationStartTime = 0;
return successfulInstall;
} catch (ReplicationHandlerException e) {
LOG.error("User aborted Replication");
} catch (SolrException e) {
throw e;
} catch (Exception e) {
throw new SolrException(SolrException.ErrorCode.SERVER_ERROR, "Index fetch failed : ", e);
} finally {
if (deleteTmpIdxDir) delTree(tmpIndexDir);
else delTree(indexDir);
}
return successfulInstall;
} finally {
if (!successfulInstall) {
logReplicationTimeAndConfFiles(null, successfulInstall);
}
filesToDownload = filesDownloaded = confFilesDownloaded = confFilesToDownload = null;
replicationStartTime = 0;
fileFetcher = null;
if (fsyncService != null && !fsyncService.isShutdown()) fsyncService.shutdownNow();
fsyncService = null;
stop = false;
fsyncException = null;
}
}
作者：洞庭散人
出处：http://phinecos.cnblogs.com/
posted on 2012-02-29 00:07 Phinecos(洞庭散人) 阅读(9523) 评论(1) 编辑 收藏 举报
