Redshift historical retrieval query performance regression

[adchia]

There seems to have been some regression in historical retrieval for Redshift that happened somewhere between 0.14.1 and 0.16.1. A quick summary of one particular user's issue:

• currently the point-in-time joins on big entity dataframes are slow in Redshift. The entity_df is generated right now by creating e.g. a user_id CROSS JOIN item_id
• The tables in question have usually have 2-4 columns per feature view, and models use between 10 - 30 feature views.

@woop had done some quick investigation and the main shift seemed to happen in #1911 introduced by @MattDelac

[MattDelac]

the main shift seemed to happen in #1911 introduced by @MattDelac

This PR does not seem to have impact on performance. It could have impact of the columns selected at the end though ...

@woop Could you post your investigation please 🙏

[adchia]

I think it wasn't any deep investigation. Was mainly trying to find changes to the query that happened in that timeframe and that PR being the main change

[MattDelac]

I think it wasn't any deep investigation. Was mainly trying to find changes to the query that happened in that timeframe and that PR being the main change

Ok ok. Can you explain what the regression is ? Is it performance wise (like the title suggests it) or "feature" wise (like the set of columns being part of the output) ?

If it's the former, I have no clue, if it's the latter, I have an idea 🙂

[adchia]

Was a slowness (performance regression).

In digging deeper, looks like 0.14.1 was cut Oct 28 and 0.16.1 was cut Dec 11, but this PR was before Oct 28, so shouldn't have caused the issue.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.