Hi Folks,
We have developed a simple log writer in Java that is plugged into
Apache custom log and writes log entries directly to our hadoop cluster
(50 machines, quad core, each with 16 GB Ram and 800 GB hard-disk, 1
machine as dedicated Namenode another machine as JobTracker &
TaskTracker + DataNode).
There are around 8 Apache servers dumping logs into HDFS via our writer.
Everything was working fine and we were getting around 15 - 20 MB log
data per hour from each server.
Recently we have been experiencing problems with 2-3 of our Apache
servers where a file is opened by log-writer in HDFS for writing but it
never receives any data.
Looking at apache error logs shows the following errors
08/09/22 05:02:13 INFO ipc.Client: java.io.IOException: Too many open
files
at sun.nio.ch.IOUtil.initPipe(Native Method)
at
sun.nio.ch.EPollSelectorImpl.<init>(EPollSelectorImpl.java:49)
at
sun.nio.ch.EPollSelectorProvider.openSelector(EPollSelectorProvider.java
:18)
at
org.apache.hadoop.net.SocketIOWithTimeout$SelectorPool.get(SocketIOWithT
imeout.java:312)
at
org.apache.hadoop.net.SocketIOWithTimeout$SelectorPool.select(SocketIOWi
thTimeout.java:227)
at
org.apache.hadoop.net.SocketIOWithTimeout.doIO(SocketIOWithTimeout.java:
155)
at
org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:149)
at
org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:122)
at java.io.FilterInputStream.read(FilterInputStream.java:116)
at
org.apache.hadoop.ipc.Client$Connection$1.read(Client.java:203)
at
java.io.BufferedInputStream.fill(BufferedInputStream.java:218)
at
java.io.BufferedInputStream.read(BufferedInputStream.java:237)
at java.io.DataInputStream.readInt(DataInputStream.java:370)
at org.apache.hadoop.ipc.Client$Connection.run(Client.java:289)
...
...
Followed by connection errors saying
"Retrying to connect to server: hadoop-server.com:9000. Already tried
'n' times".
(same as above) ...
....
and is retrying constantly (log-writer set up so that it waits and
retries).
Doing an lsof on the log writer java process shows that it got stuck in
a lot of pipe/event poll and eventually ran out of file handles.
Below is the part of the lsof output
lsof -p 2171
COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
....
....
java 2171 root 20r FIFO 0,7 24090207 pipe
java 2171 root 21w FIFO 0,7 24090207 pipe
java 2171 root 22r 0000 0,8 0 24090208
eventpoll
java 2171 root 23r FIFO 0,7 23323281 pipe
java 2171 root 24r FIFO 0,7 23331536 pipe
java 2171 root 25w FIFO 0,7 23306764 pipe
java 2171 root 26r 0000 0,8 0 23306765
eventpoll
java 2171 root 27r FIFO 0,7 23262160 pipe
java 2171 root 28w FIFO 0,7 23262160 pipe
java 2171 root 29r 0000 0,8 0 23262161
eventpoll
java 2171 root 30w FIFO 0,7 23299329 pipe
java 2171 root 31r 0000 0,8 0 23299330
eventpoll
java 2171 root 32w FIFO 0,7 23331536 pipe
java 2171 root 33r FIFO 0,7 23268961 pipe
java 2171 root 34w FIFO 0,7 23268961 pipe
java 2171 root 35r 0000 0,8 0 23268962
eventpoll
java 2171 root 36w FIFO 0,7 23314889 pipe
...
...
...
What in DFS client (if any) could have caused this? Could it be
something else?
Is it not ideal to use an HDFS writer to directly write logs from Apache
into HDFS?
Is 'Chukwa" (hadoop log collection and analysis framework contributed by
Yahoo) a better fit for our case?
I would highly appreciate help on any or all of the above questions.
Thanks and Regards
-Ankur
