It is possible for the directory size to increase even if no writes are going in to the channel. If the channel size is non-zero and the sink is still writing events to HDFS, the takes get written to disk as well (so we know what events in the files were removed when the channel/agent restarts). Eventually the channel will clean up the files which have all events taken (though it will keep at least 2 files per data directory, just to be safe).

Hari Shreedharan

On Tuesday, March 19, 2013 at 10:32 AM, Alexander Alten-Lorenz wrote:


what says debug? Do you can gather logs and attach them?

- Alex

On Mar 19, 2013, at 5:27 PM, "Kenison, Matt" <> wrote:

Check the JMX counter first, to make sure you really are not sending new events. If not, is it your checkpoint directory or data directory that is increasing in size?

From: Zhiwen Sun <>
Date: Tue, 19 Mar 2013 01:19:19 -0700
Subject: Why used space of flie channel buffer directory increase?

hi all:

I test flume-ng in my local machine. The data flow is :

tail -F file | nc 4444 > flume agent > hdfs

My configuration file is here :

a1.sources = r1
a1.channels = c2

a1.sources.r1.type = netcat
a1.sources.r1.bind =
a1.sources.r1.port = 44444
a1.sources.r1.max-line-length = 1000000

a1.sinks.k1.type = logger

a1.channels.c1.type = memory
a1.channels.c1.capacity = 10000
a1.channels.c1.transactionCapacity = 10000

a1.channels.c2.type = file
a1.sources.r1.channels = c2

a1.sources.r1.interceptors = i1
a1.sources.r1.interceptors.i1.type = timestamp

a1.sinks = k2
a1.sinks.k2.type = hdfs = c2
a1.sinks.k2.hdfs.path = hdfs://
a1.sinks.k2.hdfs.writeFormat = Text
a1.sinks.k2.hdfs.rollInterval = 10
a1.sinks.k2.hdfs.rollSize = 10000000
a1.sinks.k2.hdfs.rollCount = 0

a1.sinks.k2.hdfs.filePrefix = app
a1.sinks.k2.hdfs.fileType = DataStream

it seems that events were collected correctly.

But there is a problem boring me: Used space of file channel (~/.flume) has always increased, even there is no new event.

Is my configuration wrong or other problem?


Best regards.

Zhiwen Sun

Alexander Alten-Lorenz
German Hadoop LinkedIn Group: