One migration did a full-table UPDATE, a blocking GIN build and the
trigger install together — long locks and a deploy that could outlast
its health-check window on a large archive.
Now two stages. 0002 only adds the nullable column and the trigger
(metadata-only; new and updated rows get a vector from then on). 0003
is non-atomic: it backfills in 5,000-id batches, each its own short
transaction and only touching rows still NULL, then builds the index
with CREATE INDEX CONCURRENTLY (dropping a stale INVALID one from a
failed earlier run first). Re-running is safe, and it's a cheap no-op on
an install that applied the earlier single-step version.
Measured on a real PostgreSQL 16 with 150,000 existing articles while a
second connection kept inserting/updating: the migration took ~20s, the
slowest concurrent write was 37ms (nothing blocked), no NULL vectors
remained, the index was valid, and unapply/re-apply worked. Full suite
passes on Postgres.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Postgres recomputed a tsvector over every article's full text on every
search. A migration now adds a stored, GIN-indexed search_vector column
kept current by a trigger (so no ingest/update path can forget it), and
the backend reads that column. Untested here: no Postgres in the dev
sandbox, so the SQL is reviewed but not executed — worth one run of the
(skipped-on-SQLite) Postgres test module against a real instance.
SQLite search built one bound parameter per subscribed feed, so enough
feeds would exceed the variable limit, and the OperationalError was
swallowed into "no results". It now scopes through a subquery on
Subscription and logs failures. SQLite connections also get WAL mode,
a 20s busy timeout, and IMMEDIATE transactions so gunicorn, Celery, and
the FTS triggers stop tripping over each other.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>