跳到主要内容

21、Solr4.8.0源码分析(21)之SolrCloud的Recovery策略(二)

题记：前文<Solr4.8.0源码分析(20)之SolrCloud的Recovery策略(一)>中提到Recovery有两种策略，一是PeerSync和Replication。本节将具体介绍下PeerSync策略。

PeeySync是Solr的优先选择策略，每当需要进行recovery了，Solr总是会先去判断是否需要进入PeerSync，只有当PeerSync被设置为跳过或者PeerSync时候发现没符合条件才会进入到Replication。这是由PeeySync的特性决定的，PeeySync是面向中断时间短，需要recovery的document个数较少时使用的策略，因此它Recovery的速度较快，对Solr的影响较小。而Replication则是对中断时间长，需要recovery数量多的情况下进行的，耗时较长。

前文已经介绍了Recovery的总体流程，那么本文就直接来介绍PeerSync的流程了，请看下图所示：

 

  • 首先 Solr会向所有Replica发送getversion的请求，来获取最新的nupdate个version(默认是100个)。
// Fire off the requests before getting our own recent updates (for better concurrency)
// This also allows us to avoid getting updates we don't need... if we got our updates and then got their updates, they would
// have newer stuff that we also had (assuming updates are going on and are being forwarded).
for (String replica : replicas) {
    requestVersions(replica);
}

private void requestVersions(String replica) {
  SyncShardRequest sreq = new SyncShardRequest();
  sreq.purpose = 1;
  sreq.shards = new String[]{replica};
  sreq.actualShards = sreq.shards;
  sreq.params = new ModifiableSolrParams();
  sreq.params.set("qt","/get");
  sreq.params.set("distrib",false);
  sreq.params.set("getVersions",nUpdates);
  shardHandler.submit(sreq, replica, sreq.params);
}
  • 获取本分片最新的nupdate个version(默认是100个)，并对这些version进行排序。
recentUpdates = ulog.getRecentUpdates();
try {
  ourUpdates = recentUpdates.getVersions(nUpdates);
} finally {
  recentUpdates.close();
}

Collections.sort(ourUpdates, absComparator);
  • 获取recovery之前的version信息startingversions。通过比较startingversions与ourUpdates可以来比较recovery期间是否有索引更新。
  • 检查ourUpdates和startingversions是否有交集，由于ourUpdates和startingversions的version个数是限制为nUpdates的，也就是判断索引更新的个数是否大于nUpdate。如果需要更新的索引太多即ourUpdates和startingversions无交集，则进入Replication。
// now make sure that the starting updates overlap our updates
// there shouldn't be reorders, so any overlap will do.

long smallestNewUpdate = Math.abs(ourUpdates.get(ourUpdates.size()-1));

if (Math.abs(startingVersions.get(0)) < smallestNewUpdate) {
  log.warn(msg() + "too many updates received since start - startingUpdates no longer overlaps with our currentUpdates");
  return false;
}
  • 如果ourUpdates和startingversions有交集，则合并两个列表，即求并集。
// let's merge the lists
List<Long> newList = new ArrayList<>(ourUpdates);
for (Long ver : startingVersions) {
  if (Math.abs(ver) < smallestNewUpdate) {
    newList.add(ver);
  }
}

ourUpdates = newList;
  • 本分片的version比别的分片低，则进入Replication策略。这里进行分片version的比较，并没有按version的最大或者最小值，而是比较0.8和0.2比例处的version。
long otherHigh = percentile(otherVersions, .2f);
long otherLow = percentile(otherVersions, .8f);

if (ourHighThreshold < otherLow) {
  // Small overlap between version windows and ours is older
  // This means that we might miss updates if we attempted to use this method.
  // Since there exists just one replica that is so much newer, we must
  // fail the sync.
  log.info(msg() + " Our versions are too old. ourHighThreshold="+ourHighThreshold + " otherLowThreshold="+otherLow);
  return false;
}
  • 如果本分片的version比其他分片高，则说明不需要进行recovery直接退出peersync。
if (ourLowThreshold > otherHigh) {
  // Small overlap between windows and ours is newer.
  // Using this list to sync would result in requesting/replaying results we don't need
  // and possibly bringing deleted docs back to life.
  log.info(msg() + " Our versions are newer. ourLowThreshold="+ourLowThreshold + " otherHigh="+otherHigh);
  return true;
}
  • 对本分片的version和其他分片的version求差，获取本分片缺少的version。
for (Long otherVersion : otherVersions) {
  // stop when the entries get old enough that reorders may lead us to see updates we don't need
  if (!completeList && Math.abs(otherVersion) < ourLowThreshold) break;

  if (ourUpdateSet.contains(otherVersion) || requestedUpdateSet.contains(otherVersion)) {
    // we either have this update, or already requested it
    // TODO: what if the shard we previously requested this from returns failure (because it goes
    // down)
    continue;
  }

  toRequest.add(otherVersion);
  requestedUpdateSet.add(otherVersion);
}

最后向其他分片发送getupdate命令，根据处理后的version获取相应的document，至此完成peersync过程

private boolean requestUpdates(ShardResponse srsp, List<Long> toRequest) {
  String replica = srsp.getShardRequest().shards[0];

  log.info(msg() + "Requesting updates from " + replica + "n=" + toRequest.size() + " versions=" + toRequest);

  // reuse our original request object
  ShardRequest sreq = srsp.getShardRequest();

  sreq.purpose = 0;
  sreq.params = new ModifiableSolrParams();
  sreq.params.set("qt", "/get");
  sreq.params.set("distrib", false);
  sreq.params.set("getUpdates", StrUtils.join(toRequest, ','));
  sreq.params.set("onlyIfActive", onlyIfActive);
  sreq.responses.clear();  // needs to be zeroed for correct correlation to occur

  shardHandler.submit(sreq, sreq.shards[0], sreq.params);

  return true;
}

总结：

本文具体介绍PeerSync的过程，由此可见PeerSync策略的recovery过程还是比较简单的，下一节将具体介绍Replication策略，这个较PeerSync复杂。