Moving zstd and zstd_no_dict compression codecs out of experimental

[sarthakaggarwal97]

Is your feature request related to a problem? Please describe.
Currently, we have the experimental support for zstd and zstd compression codec as mentioned in #3354. The request is to move the feature out of the sandbox to enable the users to create an index using the new codecs.

Describe the solution you'd like
The idea is to introduce the new compression codecs for the users by moving the current implementation out of the box. With that, we will leverage the current index.codec settings that can be used to specify zstd and zstd_no_dict upon index creation. We will continue to support the existing zlib and lz4 codecs, with the the default as lz4 or BEST_SPEED.

There are the outcomes of the benchmarks with these new codecs:

• Move ZSTD out of sandbox
• • Version Upgrade Checks
• • Snapshot Checks
• • Plugin Compatibility (like KNN fixed over at Fix: avoid ZSTD codec from overriding service codec factory. #7037)
• • Stress Tests (100gb shard)
• Enable the index setting to accept new codec values

cc: @mulugetam @backslasht @mgodwan

[reta]

@sarthakaggarwal97 thanks a lot for publishing the compression gains, could you please share the CPU / memory profiles as well? Thank you.

[Bukhtawar]

+1 Latency is just one dimension, we need to understand if there are trade-offs with more cycles on compress/decompress

[reta]

Some benchmarks were also performed here https://issues.apache.org/jira/browse/LUCENE-8739 although the pull request never made it into Apache Lucene.

[sarthakaggarwal97]

@reta @Bukhtawar here are the percentage CPU utilization during indexing. The profiles were taken at a 5 minute interval during active indexing with the nyc_taxis dataset.

Summary of Compression Overheads across Codecs

Codecs	Compression Overhead
	%
zlib	10.48
lz4	4.2
zstd	9.22
zstd_no_dict	2.55

The numbers denote the %age CPU utilized for the compression. Please let me know if I can help with more information regarding the experiments.

[sarthakaggarwal97]

Some benchmarks were also performed here https://issues.apache.org/jira/browse/LUCENE-8739 although the pull request never made it into Apache Lucene.

@reta I think the reason why PR never made it into Lucene was because they were looking for a pure java implementation and didn't want to use libraries with JNI bindings in the lucere-core build.

[reta]

The numbers denote the %age CPU utilized for the compression. Please let me know if I can help with more information regarding the experiments.

Thanks @sarthakaggarwal97 , do you have Java heap (memory) stats?

[sarthakaggarwal97]

With respect to the implementation for this issue, there are two possible approaches I can see.

1. Move the zstd sandboxed code directly into source/server:
   This approach should have minimal changes from the already present code in the sandbox, and would be enabling new compression codecs using the index settings just by adding support for the new codecs.

2. Introduce the Codecs as a new module:
   In this approach, we will be introducing the new codecs (or even old ones) as a module, which will be plugged in to the CodecService eventually. With this, we will have to wire few more components like IndicesServices, IndexService, IndexModule, IndexShard so that we are able to fetch the filtered codecs from the newly added module using PluginsService. We should then be able to initialize the CodecService with the filtered/new codecs. This approach might also enable us in the future to extend support to more codecs.

I would request the community to review the implementation approaches and which one would be preferable. Please share if there are any other approaches that can be taken into consideration.

cc: @reta @dblock @backslasht @mgodwan

[dblock]

I like (2) of course because it's more generic and extensibly, but I think I'd also merge (1) if it gives the feature to users earlier.

[reta]

I would go with 2nd option, Introduce the Codecs as a new module

[mulugetam]

The quickest approach to make it accessible to users is to move it from sandbox/plugins to modules. However, the second option is more generic and can hopefully be done soon.

[backslasht]

I agree option 2 is the right long term solution. Can it be done in two steps where in option 1 is done first followed by option 2?