Unleash / unleash / 34128047932

87%

master: 91%

Build Type

push

github

Committed by web-flow

Commit Message

perf: reduce feature search segment fan-out

## Why

GET /api/admin/search/features joins feature strategies to features by
feature name before pagination. For a feature with N environments, every
strategy row can therefore be paired with all N feature-environment
rows. Segment links increase that intermediate result further even
though the response ultimately deduplicates segment names.

This became visible during the September feature-search incident, where
one instance with 1,064 flags and 17 environments saturated its database
under sustained automated search traffic.

Incident context:
https://unleash-internal.slack.com/archives/C047QAPU5U5/p1788396840135459

## What changed

- Aggregate distinct segment names once per feature before joining them
into ranked_features.
- Return that pre-aggregated segment array directly instead of
rebuilding it from repeated result rows.
- Make positive segment filters use a feature-level subquery, consistent
with the existing negative and all-of segment filters.
- Add regression coverage for segments across environments and for
combined status and segment filters that match in different
environments.

A simpler feature-strategy join on feature plus environment was tested
first, but it changed existing behavior: a development status filter
could no longer be combined with a segment used in production. The
pre-aggregation approach removes the fan-out while preserving that
behavior. It also avoids relying on a composite feature-strategy to
feature-environment constraint that the database does not currently
enforce.

## Benchmark

The benchmark harness was used locally and is intentionally not
committed in this PR. It was tied to the feature-search schema and
fixture; a reusable endpoint load-test tool should be designed
separately as an HTTP-level runner with externally prepared data.

The temporary harness called the real FeatureSearchStore against a
freshly migrated PostgreSQL database with:

- 1,064 features
- 17... (continued)

Coverage Stats

3965 of 5163 branches covered (76.8%)

3 of 3 new or added lines in 1 file covered. (100.0%)

34 existing lines in 7 files now uncovered.

16177 of 18509 relevant lines covered (87.4%)

854.41 hits per line