perf(postgres): Optimizing feast offline Store for date-range multi-FV retrieval#6057
perf(postgres): Optimizing feast offline Store for date-range multi-FV retrieval#6057Vperiodt wants to merge 8 commits intofeast-dev:masterfrom
Conversation
Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
|
@Vperiodt thanks for the PR. if it's not too much trouble, can you give a small example of result queries for say.. 2 fv scenario and highlight differences this PR introduces? there jinja templates are hard enough to read through even when one knows what the intent is 😄 |
sure ! will add it in the description itself |
Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED
2 participants
What this PR does / why we need it:
Replaces the date-range multi-FV path in the Postgres offline store (previously base_entities + one LEFT JOIN LATERAL per feature view) with a set-based LOCF (Last Observation Carried Forward) implementation
Uses a single timeline: stack (UNION ALL of spine + feature rows), COUNT for group boundaries, FIRST_VALUE for forward-fill, then filter to spine and apply per-FV TTL.
Which issue(s) this PR fixes:
Fixes slow get_historical_features on the Postgres offline store for date-range retrieval with multiple feature views. The LATERAL/inequality joins had O(N×M) cost. This PR switches to a LOCF path: one stacked timeline of size L, sort O(L log L), and O(L) window passes, which makes total cost as O(L log L).
Misc