Force exhaustive search when faceting is requested

[dureuill]

Pull Request

Related issue

Related to #3426.

What does this PR do?

When not using pagination mode, the search is not exhaustive. This can result in incomplete results for facet distribution and wrong results for facet stats (as implemented in #3423).

This PR forces the search to be as exhaustive as possible¹, even when not using pagination mode, as soon as a facet distribution (or facet stats) is required.

Tests

Unfortunately, I was not able to come up with a test case that demonstrates how the search is not exhaustive by default. I'm still opening the PR should the issue pop up.

Footnotes

1. My understanding is that we can still have differences in distribution/stats when distinct attributes are used in the query. ↩

[dureuill]

cc @ManyTheFish

[ManyTheFish]

Hello @dureuill,

thank you for the PR, I'd personally wait for proof or an issue showing that the facet distribution is not sufficiently exhaustive to merge this PR, I'm a bit concerned by the performance impact. 😕

Did you see any significant difference in the search performances?

[dureuill]

Did you see any significant difference in the search performances?

Testing on my dataset was insufficient to show anything one way or the other. We'd need tests or some more realistic use cases.

thank you for the PR, I'd personally wait for proof or an issue showing that the facet distribution is not sufficiently exhaustive to merge this PR, I'm a bit concerned by the performance impact. 😕

Sure! Let's keep this PR as draft for now

[gmourier]

@dureuill We could test it on the dataset @Strift will use to make an e-commerce demo

[Strift]

Hey there. Here is the link to the aforementioned ecommerce dataset.

To be honest, I haven't taken much time to challenge the relevance of the data. So if you see anything weird, you can let me know!

Cheers,

[Kerollmops]

I am also very concerned about the performance impact it can have. I want to avoid falling into this kind of performance issue. Having estimated counts is enough for some of our users that search performance over accuracy, but it can be an issue for other users.