http://boylook.itpub.net/post/43144/531407
对于Flume来说主要有两个Channel:Memory,File;对于线上环境主要以FileChannel为主,因此这里主要讨论它的实现:
在FileChannel里主要由一个WAL的log和一个内存队列组成:
FileChannel的Queue主要又以下几个部分组成:
privatefinal EventQueueBackingStore backingStore;
privatefinal InflightEventWrapper inflightTakes;
privatefinal InflightEventWrapper inflightPuts;
其中backingStore代表了queue在持久化存在,使用了内存映射文件的方式;每次对queue的读写操作都记录在backingStore的overwritemap(update in place)中,当进行checkpoint的时候合并到elementsBuffer并持久化到磁盘;所有未提交的正在读写数据都分别保存在inflight结构中,当checkpoint时一并进行持久化,为回滚时使用;
在inflight中存储了transactionid->fileid以及transactionid->eventptr的映射,具体存储在backingStore里的则是eventptr(fileid,offset);
Checkpoint file的文件结构如下:
File Header:1029 bytes
Eventptr;
在File header里前8个字节存储了版本号,接下来24个字节是sequeuece no.(类似rdbms的scn),接下来4个字节存储了checkpoint的状态;
作为WAL的Log主要存储了(transactionid,sequenceNo,Event),每次读写都先在log里写入event,对于写操作会拿到eventptr放入queue中;而commit和rollback操作在log中的记录形式是(transactionid,sequenceNo,OP={commit,rollback});
这两个结构主要是体现在FileBackedTransaction中如下:
FileBackedTransaction extends BasicTransactionSemantics
......
LinkedBlockingDeque<FlumeEventPointer>takeList;
LinkedBlockingDeque<FlumeEventPointer> putList;
longtransactionID;
Log log;
FlumeEventQueue queue: EventQueueBackingStoreFile
其中queue = log.getFlumeEventQueue();
首先看put/take path以及commit:
1. doPut(Eventevent)->
queue.addWithoutCommit(ptr, transactionID)
log.put(transactionID, event)->
synchronized LogFile.Writer.put(ByteBufferbuffer)
putList.offer(ptr)
2. doTake()->
FlumeEventPointer ptr = queue.removeHead(transactionID);
takeList.offer(ptr),
log.take(transactionID, ptr); ->
synchronizedLogFile.Writer.take(ByteBuffer buffer)
Event event = log.get(ptr);
3. doCommit()->
if(puts > 0) {
log.commitPut(transactionID);
synchronized (queue) {
while(!putList.isEmpty()) {
queue.addTail(putList.removeFirst())
queue.completeTransaction(transactionID);
}
}
elseif (takes > 0) {
log.commitTake(transactionID);->
logFileWriter.commit(buffer);
logFileWriter.sync();
queue.completeTransaction(transactionID);
queueRemaining.release(takes);
}
}
从上面的代码可以看出,对于每一个put/take都会记录一条oplog到log里,当commit的时候会对log进行sync到磁盘持久化,同时会把event指针存放到queue上;这里的log就类似于mysql里的binlog(binlog_format=statement),而这里的queue存放的是指向event的指针;
简例:FileChannel如下,对FileChannel put了2个消息,a,b;则在log,queue里的存储状态如下,Log里存储了(transactionid,sequenceNo,Event),queue则存储了eventptr;
Queue:ptr->a,ptr->b
WAL log:(1,1,put a),(1,2,put b),(1,3,commit)
当实例crash时,通过log来恢复queue的状态,类似rdbms一样,replay是很耗时的操作,因此会定期对queue进行checkpoint:
Log在初始化的时候会启动一个调度线程workerExecutor,由调度线程定期(checkpoint interval)调度一个backgroupWorkder来进行非强制性checkpoint;
Log.writeCheckpoint(Boolean force):tryLockExclusive->
synchronized queue.checkpoint->
backingStore.beginCheckpoint();//检查是否checkpoint正在进行;同时进行标记checkpoint开始,并同步MMAP file;
inflightPuts.serializeAndWrite();//
inflightTakes.serializeAndWrite();//将inflightputs/takes序列化并写到相应文件
backingStore.checkpoint();->
setLogWriteOrderID(WriteOrderOracle.next());
writeCheckpointMetaData();
//copy from overwriteMap toelementsBuffer(MMAP)
//标记checkpoint结束,并同步文件
简例:接上例,在a,b提交后,这时进行了一次checkpoint(存储在磁盘上的checkpoint则是2个指针ptr->a,ptr->b),此时scn=4;之后,又完成了一个take transaction ,ptr to a 也同时被删除;如果这时Flume crash,queue从checkpoint中重建,并且取得checkpoint scn=4,则replay这之后的log进行crash recovery;在恢复后,立刻执行一次checkpoint.
queue:ptr->b
WAL log:(1,1,put a),(1,2,put b),(1,3,commit),(2,5,take a),(2,6,commit)