David Mollitor created AVRO-3167:
------------------------------------
Summary: Simplify DeflateCodec
Key: AVRO-3167
URL: https://issues.apache.org/jira/browse/AVRO-3167
Project: Apache Avro
Issue Type: Improvement
Components: java
Reporter: David Mollitor
Assignee: David Mollitor
Some performance testing of another product highlighted some weirdness to me in
Avro library. In particular the way that blocks are compressed/decompressed in
{{DeflateCodec}}.
For each block of raw data, it is compressed/decompressed into a new buffer.
That new buffer is then immediately written out to disk. Well, that buffer is
requested by the caller with a requested size, but it's a bit odd because the
buffer is cached, so only the first call has any affect. Also, that buffer is
expanded as needed, but then is maintained at that size for the life of the
application, it is never resized smaller, so it could hold that large
(underutilized) buffer for awhile.
Finally, even if the requested size was working as expected, the "requested"
size is quite dubious. Right now, the requested size is equal to the size of
the raw block, which means that the buffer requested for a decompress will
always be too small and the buffer created for a compression will always be too
big. Instead, I propose that we just fix a sensible default value for all
buffers.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
