在解决了Flume-HDFS“丢数据”的问题(http://boylook.blog.51cto.com/7934327/1308188)后还听到抱怨说Flume还丢数据,如果说数据重复是可以理解的,我一直不理解为什么还丢呢?

今天同事发现在agent端日志里一段异常:

20 Nov 2013 10:15:54,231 ERROR [pool-10-thread-1] (org.apache.flume.source.ExecSource$ExecRunnable.run:347)  - Failed while running command: xxx.sh xxx.log
org.apache.flume.ChannelException: Unable to put batch on required channel: FileChannel channel_all { dataDirs: [xxx/data] }
        at org.apache.flume.channel.ChannelProcessor.processEventBatch(ChannelProcessor.java:200)
        at org.apache.flume.source.ExecSource$ExecRunnable.flushEventBatch(ExecSource.java:376)
        at org.apache.flume.source.ExecSource$ExecRunnable.run(ExecSource.java:336)
        at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:471)
        at java.util.concurrent.FutureTask$Sync.innerRun(FutureTask.java:334)
        at java.util.concurrent.FutureTask.run(FutureTask.java:166)
        at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1110)
        at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:603)
        at java.lang.Thread.run(Thread.java:722)
Caused by: org.apache.flume.ChannelException: Failed to obtain lock for writing to the log. Try increasing the log write timeout value. [channel=xxx]
        at org.apache.flume.channel.file.FileChannel$FileBackedTransaction.doPut(FileChannel.java:478)
        at org.apache.flume.channel.BasicTransactionSemantics.put(BasicTransactionSemantics.java:93)
        at org.apache.flume.channel.BasicChannelSemantics.put(BasicChannelSemantics.java:80)
        at org.apache.flume.channel.ChannelProcessor.processEventBatch(ChannelProcessor.java:189)
        ... 8 more

之后没有报错,但是Flume没有数据流出了,同事想试试根据异常的提示增加writetimeout.

再把报错发给我后,去年关于Flume的记忆又清晰起来,心中一个声音:MD Flume的默认行为还没有更改!报错的提示还这么误导!

梳理一下流程:exec source在执行Tail的时候,put数据到channel,当一直未获取到channel lock超时后,exec source会退出,而默认行为是不自动恢复重试!因此,这个问题根本不在于timeout,而在于source在存消息失败后需要快速恢复,Flume提供快速恢复的机制,但是默认竟然是关闭的!

加上这三个参数:restartThrottlerestartlogStdErr解决

但是其实用tail的方式在类似的情况下还会有“极少量”数据的丢失,想要真做到不丢数据,推荐spooling directory的方式;

另外,今天研究了一下kafkaAPI,这里也有“丢数据”的坑,敬请期待Flume“丢数据”系列三.