Fluentd + Hadoop: Instant Big Data Collection
Last updated November 13, 2012This post describes how to use Fluentd's newly released WebHDFS plugin to aggregate semi-structured logs into Hadoop HDFS.
Background
Fluentd is a JSON-based, open-source log collector originally written at Treasure Data. Fluentd is specifically designed for solving big data collection problem.
Many companies choose Hadoop Distributed Filesystem (HDFS) for big data storage. [1] Until recently, however, the only API interface was Java. This changed with the new WebHDFS interface, which allows users to interact with HDFS via HTTP. [2]
This post shows you how to set up Fluentd to receive data over HTTP and upload it to HDFS via WebHDFS.
Mechanism
The figure below shows the high-level architecture.
Install
For simplicity, this post shows the one-node configuration. You should have the following software installed on the same node.
- Fluentd with WebHDFS Plugin
- HDFS
Fluentd’s most recent version of deb/rpm package (v1.1.10 or later) includes the WebHDFS plugin. If you want to use Ruby Gems to install the plugin, gem install fluent-plugin-webhdfs
does the job.
- Debian Package
- RPM Package
- For CDH, please refer to the downloads page (CDH3u5 and CDH4 or later)
Fluentd Configuration
Let’s configure Fluentd. If you use deb/rpm, the Fluentd’s config file is located at /etc/td-agent/td-agent.conf
. Otherwise, it is located at /etc/fluentd/fluentd.conf
.
HTTP Input
For input, let’s set up Fluentd to accept data from HTTP. This is what the Fluentd configuration looks like.
type http port 8080
WebHDFS Output
The output configuration should look like this:
type webhdfs host namenode.your.cluster.local port 50070 path /log/%Y%m%d_%H/access.log.${hostname} flush_interval 10s
The match section specifies the regexp to match the tags. If the tag is matched, then the config inside it is used.
flush_internal
indicates how often data is written to HDFS. Append operation is used to append the incoming data to the file specified by the path
parameter.
For the value of path
, you can use the placeholders for time and hostname (notice how %Y%m%d_%H
and ${hostname}
are used above). This prevents multiple Fluentd instances to append the data into the same file, which must be avoided for append operation.
The other two options, host
and port
, specify HDFS’s NameNode host and port respectively.
HDFS Configuration
Append is disabled by default. Please put these configurations into your hdfs-site.xml
and restart the whole cluster.
dfs.webhdfs.enabled true dfs.support.append true dfs.support.broken.append true
Also, please make sure that path
specified in Fluentd’s WebHDFS output is configured to be writable by hdfs user.
Test
To test the setup, just post a JSON to Fluentd. This example users curl command to do so.
$ curl -X POST -d 'json={"action":"login","user":2}' http://localhost:8080/hdfs.access.test
Then, let’s access HDFS and see the stored data.
$ sudo -u hdfs hadoop fs -lsr /log/ drwxr-xr-x - 1 supergroup 0 2012-10-22 09:40 /log/20121022_14/access.log.dev
Success!
Conclusion
Fluentd + WebHDFS make real-time log collection easy, robust and scalable! @tagomoris has been using this plugin to collect 100,000 msgs/sec for a couple of months to help NHN Japan analyze big data.
Further Readings
- Fluentd + WebHDFS Tutorial
- Fluentd WebHDFS Plugin
- Fluentd Documentation
- Fluentd Plugins List
- Fluentd Source Code
Acknowledgement
Satoshi Tagomori contributed the WebHDFS plugin and battle-tested it in a super large-scale production environment. Thanks Satoshi!
- Some of Fluentd users have been using fluent-plugin-mongo with MongoDB quite successfully.
- WebHDFS is supported for Apache 1.0.0 (and later), CDH3u5 (and later) and CDH4 (and later).