zzf created FLUME-3320:
--------------------------
Summary: Taildir source reset position bug
Key: FLUME-3320
URL: https://issues.apache.org/jira/browse/FLUME-3320
Project: Flume
Issue Type: Bug
Components: Master, Sinks+Sources
Affects Versions: 1.8.0, 1.7.0
Reporter: zzf
Attachments: error.log
1) background
we use flume to transport logs from local file to kafka
every night at 0 p.m , local file will be moved to another directory to be
compressed as backup , after that , old file's inode will be released. The
monitor directory will generate new log file with the same filename.Since flume
monitor files with inode , never seen any conflict before.
2 ) problem
2019-02-23 23:59:34,157 (Log-BackgroundWorker-c1) [INFO -
org.apache.flume.channel.file.EventQueueBackingStoreFile.beginCheckpoint(EventQueueBackingStoreFile.java:230)]
Start checkpoint for /data/flume/filechannel/checkpoint/checkpoint, elements
to sync = 3406
2019-02-23 23:59:34,167 (Log-BackgroundWorker-c1) [INFO -
org.apache.flume.channel.file.EventQueueBackingStoreFile.checkpoint(EventQueueBackingStoreFile.java:255)]
Updating checkpoint metadata: logWriteOrderID: 1550778307223, queueSize: 0,
queueHead: 1754854
2019-02-23 23:59:34,176 (Log-BackgroundWorker-c1) [INFO -
org.apache.flume.channel.file.Log.writeCheckpoint(Log.java:1065)] Updated
checkpoint for file: /data/flume/filechannel/data/log-68 position: 149727224
logWriteOrderID: 1550778307223
*2019-02-24 00:00:01,550 (PollableSourceRunner-TaildirSource-r1) [INFO -
org.apache.flume.source.taildir.ReliableTaildirEventReader.updateTailFiles(ReliableTaildirEventReader.java:257)]
Pos 15927858851 is larger than file size! Restarting from pos 0, file:
/data/logs/nginx/stm-collect.cn.miaozhen.com.log, inode: 3932190*
*2019-02-24 00:00:01,550 (PollableSourceRunner-TaildirSource-r1) [INFO -
org.apache.flume.source.taildir.TailFile.updatePos(TailFile.java:126)] Updated
position, file: /data/logs/nginx/stm-collect.cn.miaozhen.com.log, inode:
3932190, pos: 0*
2019-02-24 00:00:04,177 (Log-BackgroundWorker-c1) [INFO -
org.apache.flume.channel.file.EventQueueBackingStoreFile.beginCheckpoint(EventQueueBackingStoreFile.java:230)]
Start checkpoint for /data/flume/filechannel/checkpoint/checkpoint, elements
to sync = 33326
2019-02-24 00:00:04,189 (Log-BackgroundWorker-c1) [INFO -
org.apache.flume.channel.file.EventQueueBackingStoreFile.checkpoint(EventQueueBackingStoreFile.java:255)]
Updating checkpoint metadata: logWriteOrderID: 1550778350086, queueSize:
24452, queueHead: 1763726
2019-02-24 00:00:04,213 (Log-BackgroundWorker-c1) [INFO -
org.apache.flume.channel.file.Log.writeCheckpoint(Log.java:1065)] Updated
checkpoint for file: /data/flume/filechannel/data/log-68 position: 190948775
logWriteOrderID: 1550778350086
normally,flume will open the new file and get it's inode. if inode in
Map<inode,taildir> ,and taildir position larger than file size. the position
will be reset to 0.The system went well for a long time.
but in this time,flume failed to get the new file's inode.In the log above ,you
can see there was no 'Opening file ...' . And somehow the old log file's
position was reset and all the data in 23,Feb sent double.
3)question
Logically , flume relates the real file and object tailfile with inode ,even
though I get the old file's inode,I still can not figure out why the
tailfile.getPos() will be larger than f.length().
has this happened before? should i blame this to java NIO?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]