编译
#!/bin/bash
set -e
set -x
if [[ -d output ]]
then
rm -rf output/*
fi
mvn install package -DskipTests -Dlicense.skip=true -Pbin-dist
mkdir -p output
cp -r pinot-distribution/target/apache-pinot-incubating-0.1.0-SNAPSHOT-bin/* output
chmod +x output/bin/*.sh
Segment 生成
一个segment的每列可能会有下面几个文件:
- .sv.unsorted.fwd single value 字典列&不排序列的正排
- .sv.sorted.fwd single value 字典列&排序列的正排
- .sv.raw.fwd single value 非字典列的正排
- .mv.fwd
- .bitmap.inv
- .bloom
SegmentIndexCreationDriverImpl
// Initialize the index creation using the per-column statistics information
indexCreator.init(config, segmentIndexCreationInfo, indexCreationInfoMap, dataSchema, tempIndexDir);
// Build the index
recordReader.rewind();
LOGGER.info("Start building IndexCreator!");
GenericRow readRow = null;
while (recordReader.hasNext()) {
long start = System.currentTimeMillis();
readRow = GenericRow.createOrReuseRow(readRow);
GenericRow transformedRow = _recordTransformer.transform(recordReader.next(readRow));
long stop = System.currentTimeMillis();
totalRecordReadTime += (stop - start);
if (transformedRow != null) {
indexCreator.indexRow(transformedRow);
long stop1 = System.currentTimeMillis();
totalIndexTime += (stop1 - stop);
}
}
Segment 读取
SegmentPreProcessor
Segment 加载
controller:
=> PinotSegmentUploadRestletResource.uploadSegment
=> ZKOperation.completeSegmentOperations
=>
1. 如果是new segment，调用 processNewSegment
2. 如果是已经存在的，调用 processExistingSegment
processExistingSegment会判断新的crc和旧的是否一致，最后调用
PinotHelixResourceManager.refreshSegment
=> PinotHelixResourceManager.refreshSegment
1. set segment zk metadata
2. sendSegmentRefreshMessage
会发送一个Helix SegmentRefreshMessage消息给Server，Server收到之后从相应的controller上拉segment
server 这边
=> 在HelixServerStarter中，注册了一个message handler
_helixManager.getMessagingService().registerMessageHandlerFactory(Message.MessageType.USER_DEFINE_MSG.toString(), messageHandlerFactory);
=> 这个message handler会根据message类型创建不同的handler
public MessageHandler createHandler(Message message, NotificationContext context) {
String msgSubType = message.getMsgSubType();
switch (msgSubType) {
case SegmentRefreshMessage.REFRESH_SEGMENT_MSG_SUB_TYPE:
return new SegmentRefreshMessageHandler(new SegmentRefreshMessage(message), _metrics, context);
case SegmentReloadMessage.RELOAD_SEGMENT_MSG_SUB_TYPE:
return new SegmentReloadMessageHandler(new SegmentReloadMessage(message), _metrics, context);
default:
throw new UnsupportedOperationException("Unsupported user defined message sub type: " + msgSubType);
}
}
=> 我们重点看下SegmentRefreshMessageHandler
@Override
public HelixTaskResult handleMessage()
throws InterruptedException {
HelixTaskResult result = new HelixTaskResult();
try {
acquireSema(_segmentName, LOGGER);
// The number of retry times depends on the retry count in SegmentOperations.
_fetcherAndLoader.addOrReplaceOfflineSegment(_tableNameWithType, _segmentName);
result.setSuccess(true);
} catch (Exception e) {
_metrics.addMeteredTableValue(_tableNameWithType, ServerMeter.REFRESH_FAILURES, 1);
Utils.rethrowException(e);
} finally {
releaseSema();
}
return result;
}
他会调用_fetcherAndLoader.addOrReplaceOfflineSegment(_tableNameWithType, _segmentName); 如果 addOrReplaceOfflineSegment失败，handleMessage会跑出异常，猜测Helix捕获到异常的话会重新发送message或者给其他server发送message。
Server
启动
=> StartServerCommand.execute
=> HelixServerStarter
1. waitForAllSegmentsLoaded
2. 初始化ServerInstance
ServerInstance 里面注册了Netty的handler，ScheduledRequestHandler。
每个server进程有一个HelixInstanceDataManager
=> HelixInstanceDataManager.addOfflineSegment
=> ImmutableSegmentLoader.load(indexDir, indexLoadingConfig, schema) 负责把segment从文件加载到内存
