Solr4.8.0源码分析(14) 之 SolrCloud索引深入(1)
上一章节《Solr In Action 笔记(4) 之 SolrCloud分布式索引基础》简要学习了SolrCloud的索引过程,本节开始将通过阅读源码来深入学习下SolrCloud的索引过程。
1. SolrCloud的索引过程流程图
这里借用下《solrCloud Update Request Handling 更新索引流程》流程图:
由上图可以看出,SolrCloud的索引过程主要通过一个索引链过程来实现的,那么本节主要讲述下索引链以及DistributedUpdateProcessor这个过程。
2. SolrCloud update索引链
- SolrCloud的Update索引链的类是UpdateRequestProcessorChain,这个类在Solr初始化的时候就会进行定义。
- SolrCloud的Update索引链的组成可以通过solrconfig.xml进行自定义,比较灵活,例如:
1 <updateRequestProcessorChain name="key" default="true">
2 <processor class="package.Class1" />
3 <processor class="package.Class2" >
4 <str name="someInitParam1">value</str>
5 <int name="someInitParam2">42</int>
6 </processor>
7 <processor class="solr.LogUpdateProcessorFactory" >
8 <int name="maxNumToLog">100</int>
9 </processor>
10 <processor class="solr.RunUpdateProcessorFactory" />
11 </updateRequestProcessorChain>
-
如果未自定义UpdateRequestProcessorChain,那么Solr就默认以下三个过程组成索引链,依次如下:
-
LogUpdateProcessorFactory, 它对应的处理过程为LogUpdateProcessor , 主要负责日志的记录;
-
DistributedUpdateProcessorFactory, 它对应的处理过程为DistributedUpdateProcessor ,主要负责对version的处理以及request的分发;
-
RunUpdateProcessorFactory, 它对应的处理过程为DirectUpdateHandler2 ,主要负责将记录add进本shard的lucene中;
-
如果最后一个索引链包含了RunUpdateProcessorFactory,但是没有包含DistributedUpdateProcessorFactory,那么Solr会在索引链init的时候自动在RunUpdateProcessorFactory前面加入DistributedUpdateProcessorFactory;
-
每一次update操作,都会重新创建一个索引链,即createProcessor,然后依次进行下去;
-
update操作包含add,delete,commit,所以上述UpdateProcessor都会分别包含processAdd,processDelete,processCommit,依次可以细分成add 索引链,delete索引链,commit索引链。
明白了以上几点,那么开始学习索引链的源码
- 首先来查看索引链的init代码,UpdateRequestProcessorChain.init()在初始化SolrCore的时候就调用了
1 public void init(PluginInfo info) {
2 final String infomsg = "updateRequestProcessorChain \"" +
3 (null != info.name ? info.name : "") + "\"" +
4 (info.isDefault() ? " (default)" : "");
5
6 log.info("creating " + infomsg);
7
8 // wrap in an ArrayList so we know we know we can do fast index lookups
9 // and that add(int,Object) is supported
10 //从solrcore获取索引链的成员,存放成列表形式。索引链在Solrconfig.xml中是以插件的形式加入的
11 List<UpdateRequestProcessorFactory> list = new ArrayList
12 (solrCore.initPlugins(info.getChildren("processor"),UpdateRequestProcessorFactory.class,null));
13
14 if(list.isEmpty()){
15 throw new SolrException(SolrException.ErrorCode.SERVER_ERROR,
16 infomsg + " require at least one processor");
17 }
18
19 int numDistrib = 0;
20 int runIndex = -1;
21 // hi->lo incase multiple run instances, add before first one
22 // (no idea why someone might use multiple run instances, but just in case)
23 //从后往前遍历索引链列表,寻找DistributingUpdateProcessorFactory和RunUpdateProcessorFactory
24 for (int i = list.size()-1; 0 <= i; i--) {
25 UpdateRequestProcessorFactory factory = list.get(i);
26 if (factory instanceof DistributingUpdateProcessorFactory) {
27 numDistrib++; //DistributingUpdateProcessorFactory的个数不能超过1
28 }
29 if (factory instanceof RunUpdateProcessorFactory) {
30 runIndex = i; //RunUpdateProcessorFactory的编号
31 }
32 }
33 if (1 < numDistrib) {
34 throw new SolrException(SolrException.ErrorCode.SERVER_ERROR,
35 infomsg + " may not contain more then one " +
36 "instance of DistributingUpdateProcessorFactory");
37 }
38 //如果存在RunUpdateProcessorFactory且没有DistributingUpdateProcessorFactory,
39 //那么会在RunUpdateProcessorFactory之前加入DistributingUpdateProcessorFactory
40 if (0 <= runIndex && 0 == numDistrib) {
41 // by default, add distrib processor immediately before run
42 DistributedUpdateProcessorFactory distrib
43 = new DistributedUpdateProcessorFactory();
44 distrib.init(new NamedList());
45 list.add(runIndex, distrib);
46
47 log.info("inserting DistributedUpdateProcessorFactory into " + infomsg);
48 }
49
50 chain = list.toArray(new UpdateRequestProcessorFactory[list.size()]);
51 }
- 每当有update请求时候就会触发索引链的创建,UpdateRequestProcessorChain**.**createProcessor。
1 public UpdateRequestProcessor createProcessor(SolrQueryRequest req,
2 SolrQueryResponse rsp)
3 {
4 UpdateRequestProcessor processor = null;
5 UpdateRequestProcessor last = null;
6 //获取distribPhase 是否需要跳过DistributingUpdateProcessorFactory前面那一过程(即LogUpdateProcessor),该参数不是由客户端生成
7 final String distribPhase = req.getParams().get(DistributingUpdateProcessorFactory.DISTRIB_UPDATE_PARAM);
8 final boolean skipToDistrib = distribPhase != null;
9 boolean afterDistrib = true; // we iterate backwards, so true to start
10 //从后往前组件索引链即
11 //LogUpdateProcessor.next = DistributedUpdateProcessor
12 //DistributedUpdateProcessor.next = DirectUpdateHandler2
13 //DirectUpdateHandler2.next = null
14 for (int i = chain.length-1; i>=0; i--) {
15 UpdateRequestProcessorFactory factory = chain[i];
16
17 if (skipToDistrib) {
18 if (afterDistrib) {
19 // 跳过DistributingUpdateProcessorFactory前面的索引链过程
20 if (factory instanceof DistributingUpdateProcessorFactory) {
21 afterDistrib = false;
22 }
23 } else if (!(factory instanceof UpdateRequestProcessorFactory.RunAlways)) {
24 // skip anything that doesn't have the marker interface
25 continue;
26 }
27 }
28 //创建UpdateRequestProcessorFactory对应的UpdateProcessor,并进行连接
29 processor = factory.getInstance(req, rsp, last);
30 last = processor == null ? last : processor;
31 }
32
33 return last;
34 }
-
在UpdateRequestProcessorChain.createProcessor会调用UpdateRequestProcessorFactory的getInstance创建对应的updateProcessor,
-
LogUpdateProcessor
1 public UpdateRequestProcessor getInstance(SolrQueryRequest req, SolrQueryResponse rsp, UpdateRequestProcessor next) {
2 return LogUpdateProcessor.log.isInfoEnabled() ? new LogUpdateProcessor(req, rsp, this, next) : null;
3 }
-
- DistributedUpdateProcessor
1 public DistributedUpdateProcessor getInstance(SolrQueryRequest req,
2 SolrQueryResponse rsp, UpdateRequestProcessor next) {
3
4 return new DistributedUpdateProcessor(req, rsp, next);
5 }
-
- RunUpdateProcessorFactory
1 public UpdateRequestProcessor getInstance(SolrQueryRequest req, SolrQueryResponse rsp, UpdateRequestProcessor next)
2 {
3 return new RunUpdateProcessor(req, next);
4 }
总结:本节主要学习了SolrCloud分布式索引的整体流程,以及SolrCloud建索引时候索引链的情况,下一节将详细介绍索引链的具体过程。
版权声明:本文不是「本站」原创文章,版权归原作者所有 | 原文地址: