Freegle / Iznik / 27762

85%

Build Type

push

circleci

Committed by edwh

Commit Message

perf(search): cache the browse-search reach arm for 60s per member

On top of the UNION split (1abb23824) and the restored sandwich prefilter
(366fd88d2), the remaining cold cost of a browse-scoped search is the exact
ST_Contains pass over the prefiltered reach candidates. That answer only
changes on the rippling engine's minutes-cadence ticks, but members search
repeatedly in quick succession (the reporting member tried 'headboard',
'wall tiles', 'table' within minutes), re-paying it every time.

Split the universe into its two arms in Go and cache ONLY the reach arm,
keyed by member + location (4dp, so moving invalidates immediately), 60s
TTL, bounded size. The own-posts arm stays uncached - it is a cheap indexed
query, and keeping it fresh means a just-posted item is searchable
immediately rather than after the TTL. An inner_bound fast-accept variant
was measured and rejected (decides ~23% of candidates, consistently slower).

Measured on apiv2-live against production data: 1.34s cold, 0.16s warm,
result sets byte-identical. All browse modes re-verified (nearby, group,
distance cap + Closest sort, Newest sort, legacy no-flag path unchanged).
Go suite 3516/3516.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CQZ56NwFYEuhrtaoU7oYoM

Coverage Stats

13216 of 17412 branches covered (75.9%)

Branch coverage included in aggregate %.

18 of 65 new or added lines in 1 file covered. (27.69%)

17 existing lines in 2 files now uncovered.

138114 of 191637 relevant lines covered (72.07%)

37.97 hits per line