Current state: "Under Discussion"
Discussion thread: here
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
This proposal suggests adding compression level and compression buffer size options to the producer, topic, and broker config.
Basically, CPU (running time) and I/O (compressed size) are trade-offs in compression. Since the best is use case dependent, lots of compression algorithms provide a way to control the compression level with reasonable default level, which results in a good performance in general. Add to this, the compression ratio is also affected by the buffer size - some users may want to trade compression for less/more memory.
However, Kafka does not provide a way to configure those options. Although it shows good performance with default settings, there are some cases which don't fit. For example:
- zstd supports a wide range of compression levels to reach the Pareto frontier, which means it decompresses faster than any other algorithm with similar or better compression ratio. In other words, disallowing users to adjust compression level means abandoning much of potential of zstd. In fact, Kafka's go client (sarama) already supports compression level feature for gzip and zstd.
- The default buffer size for lz4 is a little bit small (64kb). By changing up this value, the compression ratio can be improved.
This feature introduces the following new options to producer, topic, and broker configuration.
|Compression level to be used.|
|compression.buffer.size||The size of compression buffer to be used.|
The valid range and default value of compression level and buffer size are entirely up to the compression library, so they may be changed in the future. Their current values are like following:
|Compression Codec||availability||Valid Range||Default|
|gzip||Yes||1 (Deflater.BEST_SPEED) ~ 9 (Deflater.BEST_COMPRESSION)||6|
|lz4||Yes||1 ~ 17||9|
|zstd||Yes||-131072 ~ 22||3|
Compression buffer size
|Compression Codec||availability||Valid Range||Default||Note|
|gzip||Yes||Positive Integer||8192 (8kb)||Kafka's own default.|
|snappy||Yes||Positive Integer||32768 (32kb)||Library default.|
|lz4||Yes||4 ~ 7 (4=64kb, 5=256kb, 6=1mb, 7=4mb)||4 (64kb)||Kafka's own default.|
The compression will be done with the specified level using the specified compression buffer size, like following:
- If the specified option is not supported for the codec ('compression.type'), it is ignored (e.g., compression level for snappy or buffer size for zstd.)
- If the specified value is not available (or invalid), it raises an error.
- If there is no specified value, it falls back to the default one.
Compatibility, Deprecation, and Migration Plan
Since this update follows the default compression level and current buffer size if they are not set, there is no backward compatibility problem.
Can we support compression level feature only?
We can, but after some discussion, we decided that supporting both options at once is better - since both of them impacts the compression. So we decided to expand the initial proposal, which handles compression level only.
Can we support universal 'default compression level' value for producer config?
Impossible. Currently, most compression codecs allow to adjust the compression level with int type value, and it seems like it never changes. However, not all of these codecs support a value to denote 'default compression level' or the assigned value to default level differs; For example, gzip uses '-1' for default level but zstd used 0 for default level; Since the latest release of zstd allows negative compression level, the meaning of 0 level is also changing.
For these reasons, we can't provide a universal int value to denote default compression level.
Can we use external dictionary feature?
This feature requires an option to specify the dictionary for the supported codec, e.g., snappy, lz4, and zstd. It obviously over the scope of this KIP.