Solr4.8.0源码分析(22)之SolrCloud的Recovery策略(三)

Solr4.8.0源码分析(22)之SolrCloud的Recovery策略(三)

    本文是SolrCloud的Recovery策略系列的第三篇文章，前面两篇主要介绍了Recovery的总体流程，以及PeerSync策略。本文以及后续的文章将重点介绍Replication策略。Replication策略不但可以在SolrCloud中起到leader到replica的数据同步，也可以在用多个单独的Solr来实现主从同步。本文先介绍在SolrCloud的leader到replica的数据同步，下一篇文章将介绍通过配置SolrConfig.xml来实现多个Solr节点间的主从同步。

一. Replication策略介绍

        Replication的作用在前文已经介绍过了，当需要同步的数据较多时候，Solr会放弃按document为单位的同步模式(PeerSync)而采用以文件为最小单位的同步模式。在Replication的过程中，重点使用了SnapPuller类，它封装了对leader数据快照以及通过快照来实现同步的代码。Replication流程原理如下图所示。接下来根据源码来介绍每一步骤。

• 开始Replication的时候，首先进行一次commitOnLeader操作，即发送commit命令到leader。它的作用就是将leader的update中的数据刷入到索引文件中，使得快照snap完整。

  private void commitOnLeader(String leaderUrl) throws SolrServerException,
      IOException {
    HttpSolrServer server = new HttpSolrServer(leaderUrl);
    try {
      server.setConnectionTimeout(30000);
      UpdateRequest ureq = new UpdateRequest();
      ureq.setParams(new ModifiableSolrParams());
      ureq.getParams().set(DistributedUpdateProcessor.COMMIT_END_POINT, true);
      ureq.getParams().set(UpdateParams.OPEN_SEARCHER, false);
      ureq.setAction(AbstractUpdateRequest.ACTION.COMMIT, false, true).process(
          server);
    } finally {
      server.shutdown();
    }
  }

• 获取leader的lastVersion与lastGeneration，同本分片的进行比较来确定是否需要进行同步。

      //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);

• 检查本分片是否打开IndexWriter，如果没有则Recovery失败。这是因为没有打开indexWriter就无法获取索引的generation以及version信息，replication无法进行下去。

      // TODO: make sure that getLatestCommit only returns commit points for the main index (i.e. no side-car indexes)
      IndexCommit commit = core.getDeletionPolicy().getLatestCommit();
      if (commit == null) {
        // Presumably the IndexWriter hasn't been opened yet, and hence the deletion policy hasn't been updated with commit points
        RefCounted<SolrIndexSearcher> searcherRefCounted = null;
        try {
          searcherRefCounted = core.getNewestSearcher(false);
          if (searcherRefCounted == null) {
            LOG.warn("No open searcher found - fetch aborted");
            return false;
          }
          commit = searcherRefCounted.get().getIndexReader().getIndexCommit();
        } finally {
          if (searcherRefCounted != null)
            searcherRefCounted.decref();
        }
      }

• 如果获取的leader的lastestVersion为0，则表示leader没有索引数据，那么根本就不需要进行replication。所以返回true结果。

      if (latestVersion == 0L) {
        if (forceReplication && commit.getGeneration() != 0) {
          // since we won't get the files for an empty index,
          // we just clear ours and commit
          RefCounted<IndexWriter> iw = core.getUpdateHandler().getSolrCoreState().getIndexWriter(core);
          try {
            iw.get().deleteAll();
          } finally {
            iw.decref();
          }
          SolrQueryRequest req = new LocalSolrQueryRequest(core,
              new ModifiableSolrParams());
          core.getUpdateHandler().commit(new CommitUpdateCommand(req, false));
        }
        
        //there is nothing to be replicated
        successfulInstall = true;
        return true;
      }

• 我们还需要通过比较分片的lastestVersion和leader的lastestVersion来确定是否需要继续进行replication，因为两者相等同样没必要进行replication，除非进行的时forceReplication

      if (!forceReplication && IndexDeletionPolicyWrapper.getCommitTimestamp(commit) == latestVersion) {
        //master and slave are already in sync just return
        LOG.info("Slave in sync with master.");
        successfulInstall = true;
        return true;
      }

• 获取leader节点的lastestGeneration的索引文件列表以及相关文件信息，以及配置文件列表以及信息。如果文件列表为空，则退出replication。

      // get the list of files first
      fetchFileList(latestGeneration);
      // this can happen if the commit point is deleted before we fetch the file list.
      if(filesToDownload.isEmpty()) return false;

 private void fetchFileList(long gen) throws IOException {
    ModifiableSolrParams params = new ModifiableSolrParams();
    params.set(COMMAND,  CMD_GET_FILE_LIST);
    params.set(GENERATION, String.valueOf(gen));
    params.set(CommonParams.WT, "javabin");
    params.set(CommonParams.QT, "/replication");
    QueryRequest req = new QueryRequest(params);
    HttpSolrServer server = new HttpSolrServer(masterUrl, myHttpClient);  //XXX modify to use shardhandler
    try {
      server.setSoTimeout(60000);
      server.setConnectionTimeout(15000);
      NamedList response = server.request(req);

      List<Map<String, Object>> files = (List<Map<String,Object>>) response.get(CMD_GET_FILE_LIST);
      if (files != null)
        filesToDownload = Collections.synchronizedList(files);
      else {
        filesToDownload = Collections.emptyList();
        LOG.error("No files to download for index generation: "+ gen);
      }

      files = (List<Map<String,Object>>) response.get(CONF_FILES);
      if (files != null)
        confFilesToDownload = Collections.synchronizedList(files);

    } catch (SolrServerException e) {
      throw new IOException(e);
    } finally {
      server.shutdown();
    }
  }

• 建立临时的index目录来存放同步过来的数据，临时index目录的格式是index.timestamp。它存放在data目录下。

String tmpIdxDirName = "index." + new SimpleDateFormat(SnapShooter.DATE_FMT, Locale.ROOT).format(new Date());
      tmpIndex = createTempindexDir(core, tmpIdxDirName);

      tmpIndexDir = core.getDirectoryFactory().get(tmpIndex, DirContext.DEFAULT, core.getSolrConfig().indexConfig.lockType);
      
      // cindex dir...
      indexDirPath = core.getIndexDir();
      indexDir = core.getDirectoryFactory().get(indexDirPath, DirContext.DEFAULT, core.getSolrConfig().indexConfig.lockType);

• 判断isFullCopyNeeded是否为true来决定是否需要关闭IndexWriter。如果本分片(slave)的数据的version或者generation新于master(leader)或者是forceReplication，那么必须进行所有数据的完整同步。

// if the generation of master is older than that of the slave , it means they are not compatible to be copied
      // then a new index directory to be created and all the files need to be copied
      boolean isFullCopyNeeded = IndexDeletionPolicyWrapper
          .getCommitTimestamp(commit) >= latestVersion
          || commit.getGeneration() >= latestGeneration || forceReplication;

        if (isIndexStale(indexDir)) {
          isFullCopyNeeded = true;
        }
        
        if (!isFullCopyNeeded) {
          // rollback - and do it before we download any files
          // so we don't remove files we thought we didn't need
          // to download later
          solrCore.getUpdateHandler().getSolrCoreState()
          .closeIndexWriter(core, true);
        }

• 现在才开始真正的下载不同的索引文件，Replication是根据索引文件的大小来判断是否发生过变化.下载文件时候，Replication是以packet的大小为单位进行下载的，这可以在SolrConfig.xml中设置，下一篇文章将具体介绍这个。

  private void downloadIndexFiles(boolean downloadCompleteIndex,
      Directory indexDir, Directory tmpIndexDir, long latestGeneration)
      throws Exception {
    if (LOG.isDebugEnabled()) {
      LOG.debug("Download files to dir: " + Arrays.asList(indexDir.listAll()));
    }
    for (Map<String,Object> file : filesToDownload) {
      if (!slowFileExists(indexDir, (String) file.get(NAME))
          || downloadCompleteIndex) {
        dirFileFetcher = new DirectoryFileFetcher(tmpIndexDir, file,
            (String) file.get(NAME), false, latestGeneration);
        currentFile = file;
        dirFileFetcher.fetchFile();
        filesDownloaded.add(new HashMap<>(file));
      } else {
        LOG.info("Skipping download for " + file.get(NAME)
            + " because it already exists");
      }
    }
  }

 /**
     * The main method which downloads file
     */
    void fetchFile() throws Exception {
      try {
        while (true) {
          final FastInputStream is = getStream();
          int result;
          try {
            //fetch packets one by one in a single request
            result = fetchPackets(is);
            if (result == 0 || result == NO_CONTENT) {

              return;
            }
            //if there is an error continue. But continue from the point where it got broken
          } finally {
            IOUtils.closeQuietly(is);
          }
        }
      } finally {
        cleanup();
        //if cleanup suceeds . The file is downloaded fully. do an fsync
        fsyncService.submit(new Runnable(){
          @Override
          public void run() {
            try {
              copy2Dir.sync(Collections.singleton(saveAs));
            } catch (IOException e) {
              fsyncException = e;
            }
          }
        });
      }
    }

• 到这里已经完成了索引文件的同步，但是整一个同步过程才进行了一半。接下来要获取已经发生过修改的配置文件，如果没有修改过的配置文件则不需要下载配置文件。而比较配置文件是否发生变化则是比较文件的checksum信息。下载配置文件的过程downloadConfFiles()与下载索引文件的过程类似，就不具体介绍了。

    //get the details of the local conf files with the same alias/name
    List<Map<String, Object>> localFilesInfo = replicationHandler.getConfFileInfoFromCache(names, confFileInfoCache);
    //compare their size/checksum to see if
    for (Map<String, Object> fileInfo : localFilesInfo) {
      String name = (String) fileInfo.get(NAME);
      Map<String, Object> m = nameVsFile.get(name);
      if (m == null) continue; // the file is not even present locally (so must be downloaded)
      if (m.get(CHECKSUM).equals(fileInfo.get(CHECKSUM))) {
        nameVsFile.remove(name); //checksums are same so the file need not be downloaded
      }
    }

  private void downloadConfFiles(List<Map<String, Object>> confFilesToDownload, long latestGeneration) throws Exception {
    LOG.info("Starting download of configuration files from master: " + confFilesToDownload);
    confFilesDownloaded = Collections.synchronizedList(new ArrayList<Map<String, Object>>());
    File tmpconfDir = new File(solrCore.getResourceLoader().getConfigDir(), "conf." + getDateAsStr(new Date()));
    try {
      boolean status = tmpconfDir.mkdirs();
      if (!status) {
        throw new SolrException(SolrException.ErrorCode.SERVER_ERROR,
                "Failed to create temporary config folder: " + tmpconfDir.getName());
      }
      for (Map<String, Object> file : confFilesToDownload) {
        String saveAs = (String) (file.get(ALIAS) == null ? file.get(NAME) : file.get(ALIAS));
        localFileFetcher = new LocalFsFileFetcher(tmpconfDir, file, saveAs, true, latestGeneration);
        currentFile = file;
        localFileFetcher.fetchFile();
        confFilesDownloaded.add(new HashMap<>(file));
      }
      // this is called before copying the files to the original conf dir
      // so that if there is an exception avoid corrupting the original files.
      terminateAndWaitFsyncService();
      copyTmpConfFiles2Conf(tmpconfDir);
    } finally {
      delTree(tmpconfDir);
    }
  }

• 下载完索引数据以及配置文件后，现在需要处理临时的索引数据了。不同于索引文件的下载，配置文件在下载的过程中就已经替换了原先的配置文件了，这是在copyTmpConfFiles2Conf过程中。而索引数据的替换则需要根据isFullCopyNeeded这个参数，如果该值为true，则临时的索引文件将全部替换旧的索引文件，否则只是部分的文件的替换，他们的实现分别为：modifyIndexProps和moveIndexFiles。

            if (isFullCopyNeeded) {
              successfulInstall = modifyIndexProps(tmpIdxDirName);
              deleteTmpIdxDir = false;
            } else {
              successfulInstall = moveIndexFiles(tmpIndexDir, indexDir);
            }

      接下来要重点介绍下modifyIndexProps和moveIndexFiles的实现。前文讲到，同步索引文件时候，下载下来的数据会存放在data目录下，以index. 加上同步开始时间的时间戳结构的目录下。当下载数据完成后，Replication会在同级目录下新建index.property文件。该文件内只会放入一句内容，index= index.2014XXXXXXXXXX,这样做的目的就是将索引目录index重定向到index.2014XXXXXXXXXX上，这个时候相当于index.2014XXXXXXXXXX成为了index目录。然后就可以删除原来的index目录了。

     而moveIndexFiles则比较简单，即将临时文件下的索引文件都拷贝到正在用的index目录上，其中segment_N文件最后复制。

• 将Replication的统计信息存于Replication.properties文件当中。统计信息较多，这里就不介绍了。
• 如果配置文件发生变化，需要进行reloadcore操作才能使得配置生效。
• 最后进行一次openNewSearcherAndUpdateCommitPoint，重新打开searcher以及更新commit信息。

      Replication的一次同步过程就这么结束了，但是有个问题需要搞清楚，那就是在进行Replication的时候即shard的状态recoverying时候，分片是可以建索引的但是不能进行查询。在同步的时候，新进来的数据会进入到ulog中，但是这些数据是否会进入索引文件中？源码上我还没有发现可以证明新进来的数据的只会进入ulog中，但是不会进入索引文件。

     目前我认为，当shard变为recoverying时候，新进来的请求只会进入ulog中，而不会进入索引文件中，原因有3：

1. 因为一旦有新数据写入旧索引文件中，索引文件发送变化了，那么下载好后的数据(索引文件)就不好替换旧的索引文件。

2. 在同步过程中，如果isFullCopyNeeded是false，那么就会close indexwriter，既然关闭了indexwriter就无法写入新的数据。而如果isFullCopyNeeded是true的话，因为整个index目录替换，所以是否关闭indexeriter也没啥意义。

3. 在recovery过程中，当同步replication结束后，会进行replay过程，该过程就是将ulog中的请求重新进行一遍。

      以上是我目前的猜测，待我搞明白了再来修改这部分内容，或者是否有网友能指导下。

      补上最近看的一些内容，主要是关于第二点的证明：在Replication的时候如果isFullCopyNeeded是false，那么在closeIndexWriter时候会对indexWriter进行回滚回到上次commit刚结束时候，即清空上次commit之后update操作。

      
      if (indexWriter != null) {
        if (!rollback) {
          try {
            log.info("Closing old IndexWriter... core=" + coreName);
            indexWriter.close();
          } catch (Exception e) {
            SolrException.log(log, "Error closing old IndexWriter. core="
                + coreName, e);
          }
        } else {
          try {
            log.info("Rollback old IndexWriter... core=" + coreName);
            indexWriter.rollback();
          } catch (Exception e) {
            SolrException.log(log, "Error rolling back old IndexWriter. core="
                + coreName, e);
          }
        }
      }

        而且在Replication结束的时候，Solr会重新打开IndexWriter,而重新打开IndexWriter任然会进行一次回滚，清除update记录。所以在一次Replication过程中update数据是无效的。

      try {
        if (indexWriter != null) {
          if (!rollback) {
            try {
              log.info("Closing old IndexWriter... core=" + coreName);
              indexWriter.close();
            } catch (Exception e) {
              SolrException.log(log, "Error closing old IndexWriter. core="
                  + coreName, e);
            }
          } else {
            try {
              log.info("Rollback old IndexWriter... core=" + coreName);
              indexWriter.rollback();
            } catch (Exception e) {
              SolrException.log(log, "Error rolling back old IndexWriter. core="
                  + coreName, e);
            }
          }
        }
        indexWriter = createMainIndexWriter(core, "DirectUpdateHandler2");
        log.info("New IndexWriter is ready to be used.");
        // we need to null this so it picks up the new writer next get call
        refCntWriter = null;
      } finally {
        
        pauseWriter = false;
        writerPauseLock.notifyAll();
      }

二. Replay过程

       在整个recovery过程中，当replication结束后，会调用replay的来将ulog的请求重新刷入索引文件中。replay过程的本质是调用ulog的LogReplayer线程。

• LogReplayer是以transactionlog为单位的。

        for(;;) {
          TransactionLog translog = translogs.pollFirst();
          if (translog == null) break;
          doReplay(translog);
        }

• doReplay会重新获取索引链，读取transctionlog的update命令然后重新走一遍索引链三步骤，这些内容在<Solr4.8.0源码分析(14)之SolrCloud索引深入(1)>已经介绍过了，这里就不再介绍了。需要指出的是在进行doReplay时候会设置updatecmd为replay，而一旦cmd=UpdateCmd.Replay，因为无法获取到nodes所以不会分发给其他分片包括leader，所以doReplay只会对本分片有效，且不会记录ulog中。

        tlogReader = translog.getReader(recoveryInfo.positionOfStart);

        // NOTE: we don't currently handle a core reload during recovery.  This would cause the core
        // to change underneath us.

        UpdateRequestProcessorChain processorChain = req.getCore().getUpdateProcessingChain(null);
        UpdateRequestProcessor proc = processorChain.createProcessor(req, rsp);

      if ((updateCommand.getFlags() & (UpdateCommand.REPLAY | UpdateCommand.PEER_SYNC)) != 0) {
        isLeader = false;     // we actually might be the leader, but we don't want leader-logic for these types of updates anyway.
        forwardToLeader = false;
        return nodes;
      }

• LogReplayer主要用于applyBufferedUpdates(replication策略中)以及recoverFromLog(单机模式下的recovery，即从ulog进行recovery)。

总结：

        本文主要介绍了SolrCloud中Replication的原理以及过程，同时简要讲述LogReplayer的doReplay线程对ulog的日志进行recovery。下文将要重点介绍主从模式下的Replication的配置以及使用。