## Summary
pgschema doesn't compare partition bounds. Re-applying
`schema/schema.sql` to a database a relay has used therefore plans `DROP
TABLE IF EXISTS ... CASCADE` for every `events` / `delivery_log` monthly
the partition manager created at runtime, and those tables' rows are
deleted with them. Since #8464 the manager advances the catch-all by
default, so any local dev or test database a relay has started against
now hits this. One example is the persistent `buzz` database that
`scripts/start-relay-for-tests.sh` re-applies.
This adds a root `.pgschemaignore` that hides dated monthlies
(`events_p2*`, `delivery_log_p2*`) from pgschema. The monthlies declared
in `schema/partitions/` are exempted with `!` patterns, so pgschema
still manages them. pgschema reads the file from its working directory,
and every caller already runs it from the repo root: the test-relay
launchers, `postgres-test-setup.sh`, the desktop release smoke and
`_ci-relay.yml`. Both partition files now carry a note pointing at it,
and AGENTS.md gotcha #7 adds "run pgschema from the repo root" next to
the reconcile rule.
New-install layout is unchanged.
**Why the drops haven't deleted data yet:** on a reconciled database a
re-apply also plans to drop and re-add the `events` primary key on the
parent and on each declared child. The apply then fails with `cannot
drop inherited constraint "events_p2026_01_pkey"` (42P16). That happens
with or without this PR. The planned drops run in the same single
transaction, so they roll back, and the runtime monthlies have survived
so far. Once that primary-key drift is fixed, re-applies will succeed
and this ignore file is what keeps them from dropping the runtime
monthlies. The primary-key fix is a separate follow-up.
### Related issue
Follow-up to #8464. Salvaged from an earlier branch; no existing issue
found.
### Testing
- `pgschema_reapply_keeps_runtime_partitions` (Postgres) follows the
real bootstrap path in a dedicated database:
- `bin/pgschema apply` from the repo root, then
`scripts/reconcile-schema-after-pgschema.sql`;
- the relay's default maintenance pass (create plus advance);
- asserts that both catch-alls were attached before the pass and that
their bounds moved after it;
- asserts that each parent's `p2026_07` monthly exists. The catch-all
advanced past the declared bounds but kept its name.
- asserts that `pgschema plan` contains no `DROP TABLE` and no line
mentioning any runtime monthly.
- Falsified both ways:
- With `.pgschemaignore` removed, the test fails and prints the planned
drops.
- Without the reconcile step, it fails with `reconcile did not attach
both catch-alls`.
- `pgschemaignore_exempts_exactly_the_declared_monthlies` checks that
the `!` exemptions exactly match the dated monthlies in
`schema/partitions/`.
- The admin parity test now shares the `pgschema` helper, so it also
runs from the repo root like the real callers.
- All `runtime::migration` and `store::partition` tests pass against
local Postgres after merging current main, including the ignored ones:
104 passed, run with `--test-threads=1` because the full nextest
Postgres lane needs local `dropdb`. `clippy -D warnings` and the
pre-push lanes pass.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Signed-off-by: Ian Oberst <ioberst@block.xyz>
Co-authored-by: Ian Oberst <ioberst@block.xyz>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
