Hey folks,
I've been experimenting with Flume for a few weeks now, trying to
determine an approach to designing a reliable, highly available,
scalable system to store logs from various sources, including syslog.
Ideally, this system will meet the following requirements:
1. Logs from syslog across all servers make their way into HDFS.
2. Logs are stored in HDFS in a manner that is available for
post-processing:
* Example: HIVE partitions - with HDFS Flume Sink, can set
hdfs.path to
hdfs://namenode/flume/syslog/server=%{host}/facility=%{Facility}
* Example: Custom map reduce jobs...
3. Logs are stored in HDFS in a manner that is available for "reading"
by sysadmins:
* During troubleshooting/firefighting, it is quite helpful to be
able to login to a central logging system and tail -f / grep logs.
* We need to be able to see the logs "live".
Some folks may be wondering why are we choosing Flume for syslog,
instead of something like Graylog2 or Logstash? The answer is we will
be using Flume + Hadoop for the transport and processing of other types
of data in addition to syslog. For example, webserver access logs for
post processing and statistical analysis. So, we would like to make the
most use of the Hadoop cluster, keeping all logs of all types in one
redundant/scalable solution. Additionally, by keeping both syslog and
webserver access logs in Hadoop/HDFS, we can begin to correlate events.
I've run into some snags while attempting to implement Flume in a manner
that satisfies the requirements listed in the top of this message:
1. Logs to HDFS:
* I can indeed use the Flume HDFS Sink to reliably write logs into
HDFS.
* Needed to write custom serializer to add Hostname and Timestamp
fields back to syslog messages.
* See: https://issues.apache.org/jira/browse/FLUME-1666
2. Logs to HDFS in manner available for
reading/firefighting/troubleshooting by sysadmins:
* Flume HDFS Sink uses the BucketWriter for recording flume events
to HDFS.
* Creates data files like:
/flume/syslog/server=%{host}/facility=%{Facility}/FlumeData.1350997160213
* Each file is format of FlumeData (or custom prefix) followed by
. followed by unix timestamp of when the file was created.
o This is somewhat necessary... As you could have multiple
Flume writers, writing to the same HDFS, the files cannot be
opened by more than one writer. So each writer should write
to its own file.
* Latest file, currently being written to, is suffixed with ".tmp".
* This approach is not very sysadmin-friendly....
o You have to find the latest (ie. the .tmp files) and hadoop
fs -tail -f /path/to/file.tmp
o Hadoop's fs -tail -f command first prints the entire file's
contents, then begins tailing.
So the sum of it all is Flume is awesome for getting syslog (and other)
data into HDFS for post processing, but not the best at getting it into
HDFS in a sysadmin troubleshooting/firefighting format. In an ideal
world, I have syslog data coming into Flume via one transport (i.e.
SyslogTcp Source or SyslogUDP Source) and being written into HDFS in a
manner that is both post-processable and sysadmin-friendly, but it looks
like this isn't going to happen.
I've thus investigated some alternative approaches to meet the
requirements. One of these approaches is to have all of my servers send
their syslog messages to a central box running rsyslog. Then, rsyslog
would perform one of the following actions:
1. Write logs to HDFS directly using 'omhdfs' module, in a format that
is both post-processable and sysadmin-friendly :-)
2. Write logs to HDFS directly using 'hadoop-fuse-dfs' utility, which
has HDFS mounted as a filesystem.
3. Write logs to a local filesystem and also replicate logs into a
flume agent, configured with a SyslogSource and HDFS sink.
Option #1 sounds great. But unfortunately the 'omhdfs' module for
rsyslog isn't working very well. I've gotten it to login to Hadoop/HDFS
but it has issues creating/appending files. Additionally, templating is
somewhat suspect (ie. making directories /syslog/someserver/somefacility
dynamically).
Option #2 sounds reasonable, but either the HDFS FUSE module doesn't
support append mode (yet) or rsyslog is trying to create/open the files
in a manner not compliant with HDFS. No surprise, as we all know HDFS
can be somewhat "special" at times ;-) It's actually no matter
anyways... Trying to "tail -f" a file mounted via HDFS FUSE is rather
useless. The data is only and finally fed to the tail command once a
full 64MB (or whatever you use) block size of data has been written to
the file. One would only be able to use "hadoop fs -tail -f
/path/to/log" which has its own issues mentioned previously.
Option #3 would definitely work. However, now I'm storing my logs
twice. Once on some local filesystem and another time in HDFS. It
works but its not ideal as it's a waste of space. And you've probably
noticed from this email so far, I'd prefer the *ideal* solution :-)
*Note*: Astute flumers would probably look at option #3 and recommend
making use of the RollingFileSink in addition to the HDFSSink.
Unfortunately, the RollingFileSink doesn't support templated/dynamic
directory creation like the HDFSSink with its hdfs.path setting of
"hdfs://namenode/flume/syslog/server=%{host}/facility=%{Facility}".
So what exactly am I asking here? Well, I'd like to know first how
others are doing this. A hybrid of rsyslog and Flume? All and only
Flume? With custom serializers/interceptors/sinks? Or perhaps... how
would you recommend I handle this?
Thanks for any and all thoughts you can provide.
--
Josh West
Lead Systems Administrator
One.com, [email protected]
