[Feature Request] Add a cluster setting for memory threshold to pick OrdinalsCollector in Cardinality aggregation

[rishabhmaurya]

Is your feature request related to a problem? Please describe

Currently the logic to pick Ordinals vs DirectCollector is dynamic and based on the number of ordinals to collect and memory overhead. If memory overhead is high for OrdinalsCollector, then DirectCollector is used. Due to https://issues.apache.org/jira/browse/LUCENE-9663 few users are reporting regression in Cardinality aggregation because of slower DirectCollector after replacing prefix compression with LZ4 compression for terms dictionary in lucene 8.9.
Fix proposed here is to use OrdinalsCollector more often which will collect the ordinals into a bitset first and then performs term lookup in postCollect() of segment and that's a lot faster.
However, we don't have a way to control picking up OrdinalsCollector in OpenSearch.

Describe the solution you'd like

Introduce a memory threshold dynamic setting which OrdinalsCollector can use and if its usage is under this threshold, always use OrdinalsCollector. This logic can be added here.

Related component

Search:Aggregations

Describe alternatives you've considered

Use of eager_global_ordinals or murmur hash, but its an index time setting and can have impact on indexing performance. Its discussed in more detail here.

Additional context

No response

[mch2]

Thanks @rishabhmaurya seems like something we should do, am removing untriaged.

[maitreya2954]

@mch2 I would like to work on this. I am new to contributing to OpenSearch. So, would appreciate any guidance.

[jainankitk]

Thanks @maitreya2954 for volunteering, assigned it to you!

[rishabhmaurya]

@maitreya2954 Appreciate you picking this up. Here is what we can do -

1. Introduce execution_hint for cardinality aggregation similar to term aggregation, with values: direct, ordinal to start with.
2. The default behaviour should be current behaviour. When execution_hint is provided, override the collector to be picked here and don't honour this condition here.
3. Add unit tests for all 3 cases.
4. Add documentation here by opening a documentation issue on this repo.

If you're blocked don't hesitate to ask for pointers. Thank you.

[maitreya2954]

@rishabhmaurya I added execution hint field to the cardinality agg. However, I am stuck at adding unit tests.

Add unit tests for all 3 cases.

3 cases being: direct, ordinal and when no value is provided. right?

Here's my question: What type of query should I test after setting the execution hint. I can see many different type of queries being tested here. Should I just do MatchAllDocsQuery?

[rishabhmaurya]

@maitreya2954 That's great that you were able to add execution hint.

3 cases being: direct, ordinal and when no value is provided. right?

yes.

What type of query should I test after setting the execution hint. I can see many different type of queries being tested here. Should I just do MatchAllDocsQuery?

You can test for both match_all query and a filter using term query on a field other than cardinality aggregation field.