Safeguard: limit number of terms in TermsQuery

[jonaf]

Describe the feature: I have seen a lot of cases where an application will generate some long list of terms for a TermsQuery and pass it off to Elasticsearch. In general, TermsQuery performance degrades quickly as more terms are added, and once you get into hundreds (or thousands), you can pretty much forget it. (Unless there's a way to do a terms query with hundreds or thousands of terms and I just don't know the secret.) I think we should consider short-circuiting on the number of terms if it goes beyond some threshold unless some overriding argument is given. Additionally, it might be a good idea for Elastic.co documentation to warn users that many terms can have a substantial impact on latency, and of course point out the default limit/safeguard (supposing it is decided that the number of terms should be limited).

Relates #11511

[jpountz]

+1 on adding a soft limit to discourage having too many terms in terms queries. However I don't think we should let the query execute with a partial list of terms in that case, we should fail the entire request.

[jonaf]

Sorry, by "short-circuiting," I meant abandoning the entire query, not truncating the list of terms.

[jpountz]

Cool, thanks for clarifying. +1 then!

[jpountz]

@jonaf In case you would like to work on a PR that fixes this issue, feel free to ping me for review.

[hluu]

+1

[jonaf]

@jpountz Happy to. I just had to get authorization to sign the CLA. Should this contribution be destined for master only, or 2.x as well? (Based on reading contribution guidelines, it looks like I need to open a separate PR for each branch and reference this issue from the PR's?)

Edit: just noticed the #11511 is tagged for v5, so I'll just make these changes from master.

[jpountz]

@jonaf Awesome! In general we build PRs for master first and only consider backporting when the change is ready for master.

[toniov]

In general, TermsQuery performance degrades quickly as more terms are added, and once you get into hundreds (or thousands), you can pretty much forget it.

Would this be applied in case of a Terms Query using the lookup mechanism as well?

In the above link it says:

When it’s needed to specify a terms filter with a lot of terms it can be beneficial to fetch those term values from a document in an index.

But not sure how much is that a lot.