Hbase split region代码阅读笔记

在服务器端进行Region的分割，其分割的代码在org.apache.hadoop.hbase.regionserver包下的SplitRequest.java

其中的run()函数：

SplitTransaction st = new SplitTransaction(parent, midKey);

.......

st.execute(this.server, this.server);

.......

进行分割的函数实现。

调用的execute函数如下：

public PairOfSameType<HRegion> execute(final Server server,
      final RegionServerServices services)
  throws IOException {
    useZKForAssignment = server == null ? true :
      ConfigUtil.useZKForAssignment(server.getConfiguration());
    if (useCoordinatedStateManager(server)) {
      std =
          ((BaseCoordinatedStateManager) server.getCoordinatedStateManager())
              .getSplitTransactionCoordination().getDefaultDetails();//分割Region之前的一些协调工作
    }
    PairOfSameType<HRegion> regions = createDaughters(server, services);//创建子Region函数
    if (this.parent.getCoprocessorHost() != null) {
      this.parent.getCoprocessorHost().preSplitAfterPONR();
    }
    return stepsAfterPONR(server, services, regions);
  }

所有的分裂操作都在SplitTransaction类中进行，execute函数本身就在SplitTransaction类中。

其中的createDaughters()函数实现创建子Region的工作。

PairOfSameType<HRegion> createDaughters(final Server server,
      final RegionServerServices services) throws IOException {
    LOG.info("Starting split of region " + this.parent);
    if ((server != null && server.isStopped()) ||
        (services != null && services.isStopping())) {
      throw new IOException("Server is stopped or stopping");
    }
    assert !this.parent.lock.writeLock().isHeldByCurrentThread():
      "Unsafe to hold write lock while performing RPCs";

    journal.add(new JournalEntry(JournalEntryType.BEFORE_PRE_SPLIT_HOOK));

    // Coprocessor callback
    if (this.parent.getCoprocessorHost() != null) {
      // TODO: Remove one of these
      this.parent.getCoprocessorHost().preSplit();
      this.parent.getCoprocessorHost().preSplit(this.splitrow);
    }

    journal.add(new JournalEntry(JournalEntryType.AFTER_PRE_SPLIT_HOOK));

    // If true, no cluster to write meta edits to or to update znodes in.
    boolean testing = server == null? true:
        server.getConfiguration().getBoolean("hbase.testing.nocluster", false);
    this.fileSplitTimeout = testing ? this.fileSplitTimeout :
        server.getConfiguration().getLong("hbase.regionserver.fileSplitTimeout",
          this.fileSplitTimeout);

    PairOfSameType<HRegion> daughterRegions = stepsBeforePONR(server, services, testing);//此处进行Region的划分

    List<Mutation> metaEntries = new ArrayList<Mutation>();
    if (this.parent.getCoprocessorHost() != null) {
      if (this.parent.getCoprocessorHost().
          preSplitBeforePONR(this.splitrow, metaEntries)) {
        throw new IOException("Coprocessor bypassing region "
            + this.parent.getRegionNameAsString() + " split.");
      }
      try {
        for (Mutation p : metaEntries) {
          HRegionInfo.parseRegionName(p.getRow());
        }
      } catch (IOException e) {
        LOG.error("Row key of mutation from coprossor is not parsable as region name."
            + "Mutations from coprocessor should only for hbase:meta table.");
        throw e;
      }
    }

    // This is the point of no return.  Adding subsequent edits to .META. as we
    // do below when we do the daughter opens adding each to .META. can fail in
    // various interesting ways the most interesting of which is a timeout
    // BUT the edits all go through (See HBASE-3872).  IF we reach the PONR
    // then subsequent failures need to crash out this regionserver; the
    // server shutdown processing should be able to fix-up the incomplete split.
    // The offlined parent will have the daughters as extra columns.  If
    // we leave the daughter regions in place and do not remove them when we
    // crash out, then they will have their references to the parent in place
    // still and the server shutdown fixup of .META. will point to these
    // regions.
    // We should add PONR JournalEntry before offlineParentInMeta,so even if
    // OfflineParentInMeta timeout,this will cause regionserver exit,and then
    // master ServerShutdownHandler will fix daughter & avoid data loss. (See
    // HBase-4562).
    this.journal.add(new JournalEntry(JournalEntryType.PONR));

    // Edit parent in meta.  Offlines parent region and adds splita and splitb
    // as an atomic update. See HBASE-7721. This update to META makes the region
    // will determine whether the region is split or not in case of failures.
    // If it is successful, master will roll-forward, if not, master will rollback
    // and assign the parent region.
    if (!testing && useZKForAssignment) {
      if (metaEntries == null || metaEntries.isEmpty()) {
        MetaTableAccessor.splitRegion(server.getConnection(),
          parent.getRegionInfo(), daughterRegions.getFirst().getRegionInfo(),
          daughterRegions.getSecond().getRegionInfo(), server.getServerName());
      } else {
        offlineParentInMetaAndputMetaEntries(server.getConnection(),
          parent.getRegionInfo(), daughterRegions.getFirst().getRegionInfo(), daughterRegions
              .getSecond().getRegionInfo(), server.getServerName(), metaEntries);
      }
    } else if (services != null && !useZKForAssignment) {
      if (!services.reportRegionStateTransition(TransitionCode.SPLIT_PONR,
          parent.getRegionInfo(), hri_a, hri_b)) {
        // Passed PONR, let SSH clean it up
        throw new IOException("Failed to notify master that split passed PONR: "
          + parent.getRegionInfo().getRegionNameAsString());
      }
    }
    return daughterRegions;
  }

在创建子Region结束后，完成父Region的下线工作，此外还需要对meta表进行更新操作。至此整体的Region工作分析完毕。