[jemiahw]

Summary

The relay applied #a (NIP-01 address) filters only after the SQL LIMIT. A selective read like {kinds:[7], "#a":["30175:<pubkey>:<d>"], limit:n} fetched the newest kind:7 rows across the whole community and then dropped the ones that didn't match. When older matches existed, it could return fewer than n of them, or none. COUNT also could not take its exact fast path. Clients therefore cannot count reactions to an addressable event by address, which is the identity that persists when the event is edited.

What changed:

• EventQuery.a_tags (buzz-db) works like e_tags. It is an exact positional tag match (tag->>0 = 'a' AND tag->>1 = ANY(..)) with JSONB @> as the prefilter on the existing GIN index idx_events_tags_gin.
  • No schema change, migration or backfill.
  • It gets the same MATERIALIZED CTE fence as #e, so the planner cannot walk created_at backward looking for a rare match.
  • It reads the table whole rather than recent partitions first.
  • It is counted in count_events_sql.
  • An empty set matches nothing, as the NIP-01 post-filter does.
• filter_to_query_params fills a_tags for every filter, and filter_fully_pushable treats #a as pushed. REQ, COUNT and the HTTP bridge /query and /count all go through these two functions.
• per_channel_scope now skips per-channel reads for every fenced query, through a shared reference_tag_fence helper. Before this, it excluded #e reads explicitly, and #a with a multi-value #h would have wrapped the fence around the per-channel lateral, which is invalid SQL.

Not included: #k and other low-cardinality tags. Postgres's estimate for JSONB @> assumes any match is rare, so pushing a common value could switch to a slower plan. They remain post-filters.

Related issue

None found.

Testing

• req.rs unit test: #a maps to a_tags (including empty and absent) and is fully pushable; #a + #k is not.
• buzz-db shape tests: an #a read is fenced, COUNT applies the match, an #a read is never split into recent and older passes, and #a makes a read ineligible for per-channel scope. An empty #a is answered without a database read.
• buzz-db Postgres: an #a case added to pushed_down_tag_filters_reject_positional_decoys, covering query and COUNT.
• New handlers/tag_pushdown_postgres_tests.rs:
  • REQ returns an older #a match ahead of a full page of newer reactions to another address. COUNT is exact, and an empty #a counts 0. This test fails when a_tags is removed from filter_to_query_params.
  • REQ and COUNT with #a + a multi-value #h. This test fails, with CLOSED database error, without the per_channel_scope fix.
• Full buzz-db and buzz-relay suites, unit and Postgres lanes, run locally.
• Live check against a local relay built from this branch, over WebSocket with NIP-42 auth. Two reactions to one address (each on a different head event, as after an edit) and six newer reactions to another address were published. REQ {kinds:[7], #a:[needle], limit:2} returned exactly the two needle reactions, newest first. COUNT returned 2 for the needle, 6 for the hay and 0 for an empty #a. #a + #e with limit:1 returned the newest needle reaction.

🤖 Generated with Claude Code