Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
1.6.3
-
None
-
None
Description
We're writing to Hadoop via DataFileWriter (FSDataOutputStream). We're doing this with two threads per node, on 8 nodes. Some of the nodes share the same path. For example, our: TimestampedWriter class, takes a path argument and appends the timestamp to it (ex: SomePath/2012/05/14). Thus, two threads or two nodes can access the same path. The "race" condition when these streams are written, is resolved with a check to see if the file exists (has been created) by a faster thread. If that's so, it appends, instead of creating the file on the HDFS.
The problem is that DataFileWriter, generates a 16-byte, random string for each instance. So, two threads with 2 different writer instances, have a different sync marker. That means that data, when trying to read it back, will get an IOException ("Invalid sync!").
There's a big performance penalty here. Because only one writer can write at once to one given path, it becomes a bottleneck. For 1B (billion) rows, it took us 4 hours to generate & load. With 20 concurrent threads, it took only 12.5 minutes.
If DataFileWriter would expose the "sync" marker, a developer could read that and make sure that the next thread that appends to the file, uses the same sync marker. Don't know if it's even possible to expose the sync marker so as other instances of "DataFileWriter" can share the sync marker, from the file. We have a fix for this, making sure each writer is an "unique" instance and generating a path based on that uniqueness. But instead of having "SomePath/2012/05/14/Shard.avro" we'd now have "SomePath/2012/05/14/Shard-some-random-UUID.avro" for each of the writers that write the data in.
If it can be done, it would be a huge fix for a bottleneck problem. The bottleneck being the single writer that can write to a single path.
THIS HAS ALSO been requested on the avro-user thread: http://grokbase.com/t/avro/user/122m4sjm1y/is-it-possible-to-append-to-an-already-existing-avro-file
I just could not find the JIRA ticket for this request.