Enabling Compression Level configuration of ZSTD and ZSTD (No Dict)

[sarthakaggarwal97]

Is your feature request related to a problem? Please describe.
In the present implementation, zstd and zstd_no_dict uses default compression level i.e. 6 as mentioned here, here and here. According to documentation, these algorithms can support the compression level range of 1-22. Theoretically, configurable compression levels can help customers tune the tradeoff between storage and throughput according to their requirements.

Describe the solution you'd like
An introduction of configurable compression level as an index setting for zstd and zstd-no-dict

Additional context
#3354

cc: @mgodwan @backslasht @shwetathareja

[shwetathareja]

@sarthakaggarwal97 OpenSearch users can use some of the existing benchmarks for their reference when choosing the compression level. It would be great if you can share some of the benchmark runs as well across different levels.

[dblock]

We've been collaborating with the Intel team and benchmarking compression levels in additional real world scenarios with @backslasht. Stay tuned @shwetathareja!

[dblock]

Related, in #7475 we are proposing to make the block size configurable as well.

[sarthakaggarwal97]

Currently, we have set the default compression level for zstd and zstd_no_dict as level 6 over here.

In the experiments, I'm observing that there is a improvement of roughly 5-6% in the average indexing throughput with level 3 as the default compression level.

Benchmarks

NYC Taxis Dataset

HTTP Logs Dataset

---

I think we should switch to level 3 as the default compression level. Moreover, zstd also internally uses level 3 as the default compression level as mentioned over here

cc: @mgodwan @shwetathareja @backslasht @dblock @reta

[mgodwan]

Thanks for sharing these numbers.
Given level 3 is the default level used in zstd implementation as well and yields better results in terms of throughput in majority of the cases, it should be alright to proceed with the same.

Is there any comparison of storage used as well across these to get insights around the trade-off?

[sarthakaggarwal97]

Thanks for sharing these numbers. Given level 3 is the default level used in zstd implementation as well and yields better results in terms of throughput in majority of the cases, it should be alright to proceed with the same.

Is there any comparison of storage used as well across these to get insights around the trade-off?

@mgodwan Adding comparison for store size along with indexing throughput. We roughly see 4% increase in the store size with level 3 having level 6 as baseline.

NYC Taxis

---

HTTP Logs

[backslasht]

Thanks @sarthakaggarwal97. This seems like a safe bet as the increase in indexing throughput and storage size are proportional.

Have we run tests for sufficiently large time? wondering if it will also increase the background merge time as the segments might be large with compression level 3?

[sarthakaggarwal97]

@backslasht yes, I observed background merges were happening for the runs. The segments with level 3 should not be large enough to start affecting background merge time when compared to level 6. Level 3 looks like a good sweet spot.
I will raise the PR to change the default compression level to 3.

[dblock]

Can this be closed with #8471?