Committed to the flume-728 branch. Thanks Prasad!

Committed to flume-728 branch with modifications.

• Renamed classes to be camel case.
• Moved HDFS* related classes into o.a.flume.sink.hdfs.
• Added Hadoop as a dependency to the pom so we properly build / test.

Compression and parametric serializer interface

HDFS sink with roll, batch and bucketing support.
compression part is not fully tested yet ..

First cut with tests
Limited functionality, no bucketing support yet.

First cut.
Limited functionality. not fully tested. Additional changes and tests will follow

Assigning to Prasad.

Port the Flume HDFS sink functionality over to Flume NG.

The interesting features are file rotation, output bucketing, and support for append (flush).

A minimal implementation would support file rotation. Rotation should be configurable based on both time interval (specified in seconds) and size. Ideally, we do not create files unless there are events output (i.e. lazy file creation). It should be possible to specify rotation for time and size together, meaning rotate on whichever happens first.

Output bucketing is a feature support by Flume today that allows interpolation of event attributes in output paths. For instance, an output path of /logs/%

/%

/%

/ should become /logs/2011/01/01/ for an event with the atributes year=2011, month=01, day=01. This implies we must keep multiple writers open concurrently, each with separate bookkeeping on rotation time and output size.

Support for append should be orthogonal to file rotation. In other words we should still allow the user to specify a rotation policy (time and size) but we should call flush with a given frequency, probably specified in terms of the number of events. A fully durable configuration would flush after each event (i.e. flushInterval=1). We should only enable append support if the underlying HDFS install supports it. If the user specifies a flush policy and HDFS doesn't support append, we should warn, but continue.