代码改变世界

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

2014-12-06 00:08  追风的蓝宝  阅读(1535)  评论(1)  编辑  收藏  举报

  • 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)。
 1     // Fire off the requests before getting our own recent updates (for better concurrency)
 2     // This also allows us to avoid getting updates we don't need... if we got our updates and then got their updates, they would
 3     // have newer stuff that we also had (assuming updates are going on and are being forwarded).
 4     for (String replica : replicas) {
 5       requestVersions(replica);
 6     }
 7 
 8   private void requestVersions(String replica) {
 9     SyncShardRequest sreq = new SyncShardRequest();
10     sreq.purpose = 1;
11     sreq.shards = new String[]{replica};
12     sreq.actualShards = sreq.shards;
13     sreq.params = new ModifiableSolrParams();
14     sreq.params.set("qt","/get");
15     sreq.params.set("distrib",false);
16     sreq.params.set("getVersions",nUpdates);
17     shardHandler.submit(sreq, replica, sreq.params);
18   }
  • 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。
1     recentUpdates = ulog.getRecentUpdates();
2     try {
3       ourUpdates = recentUpdates.getVersions(nUpdates);
4     } finally {
5       recentUpdates.close();
6     }
7 
8     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。
1       // now make sure that the starting updates overlap our updates
2       // there shouldn't be reorders, so any overlap will do.
3 
4       long smallestNewUpdate = Math.abs(ourUpdates.get(ourUpdates.size()-1));
5 
6       if (Math.abs(startingVersions.get(0)) < smallestNewUpdate) {
7         log.warn(msg() + "too many updates received since start - startingUpdates no longer overlaps with our currentUpdates");
8         return false;
9       }
  • 如果ourUpdates和startingversions有交集，则合并两个列表，即求并集。
1       // let's merge the lists
2       List<Long> newList = new ArrayList<>(ourUpdates);
3       for (Long ver : startingVersions) {
4         if (Math.abs(ver) < smallestNewUpdate) {
5           newList.add(ver);
6         }
7       }
8 
9       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。
 1     long otherHigh = percentile(otherVersions, .2f);
 2     long otherLow = percentile(otherVersions, .8f);
 3 
 4     if (ourHighThreshold < otherLow) {
 5       // Small overlap between version windows and ours is older
 6       // This means that we might miss updates if we attempted to use this method.
 7       // Since there exists just one replica that is so much newer, we must
 8       // fail the sync.
 9       log.info(msg() + " Our versions are too old. ourHighThreshold="+ourHighThreshold + " otherLowThreshold="+otherLow);
10       return false;
11     }
  • 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。
1     if (ourLowThreshold > otherHigh) {
2       // Small overlap between windows and ours is newer.
3       // Using this list to sync would result in requesting/replaying results we don't need
4       // and possibly bringing deleted docs back to life.
5       log.info(msg() + " Our versions are newer. ourLowThreshold="+ourLowThreshold + " otherHigh="+otherHigh);
6       return true;
7     }

 

  • 对本分片的version和其他分片的version求差，获取本分片缺少的version。
 1     for (Long otherVersion : otherVersions) {
 2       // stop when the entries get old enough that reorders may lead us to see updates we don't need
 3       if (!completeList && Math.abs(otherVersion) < ourLowThreshold) break;
 4 
 5       if (ourUpdateSet.contains(otherVersion) || requestedUpdateSet.contains(otherVersion)) {
 6         // we either have this update, or already requested it
 7         // TODO: what if the shard we previously requested this from returns failure (because it goes
 8         // down)
 9         continue;
10       }
11 
12       toRequest.add(otherVersion);
13       requestedUpdateSet.add(otherVersion);
14     }
  • 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
 1   private boolean requestUpdates(ShardResponse srsp, List<Long> toRequest) {
 2     String replica = srsp.getShardRequest().shards[0];
 3 
 4     log.info(msg() + "Requesting updates from " + replica + "n=" + toRequest.size() + " versions=" + toRequest);
 5 
 6     // reuse our original request object
 7     ShardRequest sreq = srsp.getShardRequest();
 8 
 9     sreq.purpose = 0;
10     sreq.params = new ModifiableSolrParams();
11     sreq.params.set("qt", "/get");
12     sreq.params.set("distrib", false);
13     sreq.params.set("getUpdates", StrUtils.join(toRequest, ','));
14     sreq.params.set("onlyIfActive", onlyIfActive);
15     sreq.responses.clear();  // needs to be zeroed for correct correlation to occur
16 
17     shardHandler.submit(sreq, sreq.shards[0], sreq.params);
18 
19     return true;
20   }

 

 

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。