fluent-plugin-webhdfs

Fluentd output plugin to write data into Hadoop HDFS over WebHDFS/HttpFs.

WebHDFSOutput slices data by time (specified unit), and store these data as hdfs file of plain text. You can specify to:

  • format whole data as serialized JSON, single attribute or separated multi attributes
  • include time as line header, or not
  • include tag as line header, or not
  • change field separator (default: TAB)
  • add new line as termination, or not

And you can specify output file path as 'path /path/to/dir/access.%Y%m%d.log', then got '/path/to/dir/access.20120316.log' on HDFS.

Configuration

WebHDFSOutput

To store data by time,tag,json (same with 'type file') over WebHDFS:

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /path/on/hdfs/access.log.%Y%m%d_%H.log
</match>

If you want JSON object only (without time or tag or both on header of lines), specify it by output_include_time or output_include_tag (default true):

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /path/on/hdfs/access.log.%Y%m%d_%H.log
  output_include_time false
  output_include_tag  false
</match>

To specify namenode, namenode is also available:

<match access.**>
  type     webhdfs
  namenode master.your.cluster.local:50070
  path     /path/on/hdfs/access.log.%Y%m%d_%H.log
</match>

To store data as LTSV without time and tag over WebHDFS:

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /path/on/hdfs/access.log.%Y%m%d_%H.log
  output_data_type ltsv
</match>

With username of pseudo authentication:

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /path/on/hdfs/access.log.%Y%m%d_%H.log
  username hdfsuser
</match>

Store data over HttpFs (instead of WebHDFS):

<match access.**>
  type webhdfs
  host httpfs.node.your.cluster.local
  port 14000
  path /path/on/hdfs/access.log.%Y%m%d_%H.log
  httpfs true
</match>

Store data as TSV (TAB separated values) of specified keys, without time, with tag (removed prefix 'access'):

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /path/on/hdfs/access.log.%Y%m%d_%H.log

  field_separator TAB        # or 'SPACE', 'COMMA' or 'SOH'(Start Of Heading: \001)
  output_include_time false
  output_include_tag true
  remove_prefix access

  output_data_type attr:path,status,referer,agent,bytes
</match>

If message doesn't have specified attribute, fluent-plugin-webhdfs outputs 'NULL' instead of values.

Namenode HA / Auto retry for WebHDFS known errors

fluent-plugin-webhdfs (v0.2.0 or later) accepts 2 namenodes for Namenode HA (active/standby). Use standby_namenode like this:

<match access.**>
  type             webhdfs
  namenode         master1.your.cluster.local:50070
  standby_namenode master2.your.cluster.local:50070
  path             /path/on/hdfs/access.log.%Y%m%d_%H.log
</match>

And you can also specify to retry known hdfs errors (such like LeaseExpiredException) automatically. With this configuration, fluentd doesn't write logs for this errors if retry successed.

<match access.**>
  type               webhdfs
  namenode           master1.your.cluster.local:50070
  path               /path/on/hdfs/access.log.%Y%m%d_%H.log
  retry_known_errors yes
  retry_times        1 # default 1
  retry_interval     1 # [sec] default 1
</match>

Performance notifications

Writing data on HDFS single file from 2 or more fluentd nodes, makes many bad blocks of HDFS. If you want to run 2 or more fluentd nodes with fluent-plugin-webhdfs, you should configure 'path' for each node. You can use '$hostname' or '$uuid:random' placeholders in configuration for this purpose.

For hostname:

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /log/access/%Y%m%d/${hostname}.log
</match>

Or with random filename (to avoid duplicated file name only):

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /log/access/%Y%m%d/${uuid:random}.log
</match>

With configurations above, you can handle all of files of '/log/access/20120820/*' as specified timeslice access logs.

For high load cluster nodes, you can specify timeouts for HTTP requests.

<match access.**>
  type webhdfs
  namenode master.your.cluster.local:50070
  path /log/access/%Y%m%d/${hostname}.log
  open_timeout 180 # [sec] default: 30
  read_timeout 180 # [sec] default: 60
</match>

For unstable Namenodes

With default configuration, fluent-plugin-webhdfs checks HDFS filesystem status and raise error for inacive NameNodes.

If you were usging unstable NameNodes and have wanted to ignore NameNode errors on startup of fluentd, enable ignore_start_check_error option like below:

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070
  path /log/access/%Y%m%d/${hostname}.log
  ignore_start_check_error true
</match>

For unstable Datanodes

With unstable datanodes that frequently downs, appending over WebHDFS may produce broken files. In such cases, specify append no and ${chunk_id} parameter.

<match access.**>
  type webhdfs
  host namenode.your.cluster.local
  port 50070

  append no
  path   /log/access/%Y%m%d/${hostname}.${chunk_id}.log
</match>

out_webhdfs creates new files on hdfs per flush of fluentd, with chunk id. You shouldn't care broken files from append operations.

TODO

  • configuration example for Hadoop Namenode HA
    • here, or docs.fluentd.org ?
  • patches welcome!
  • Copyright (c) 2012- TAGOMORI Satoshi (tagomoris)
  • License
    • Apache License, Version 2.0