Migrations on large tables never take a lock that stops writes

OPS-36: a migration built an index on access_requests (3.4 million rows in the EU) without CONCURRENTLY and added a foreign key in the same step. Writes to access_requests waited for 9 minutes: no new requests, no decisions.

• CI fails a migration on people, access_requests, grants or sms_messages that builds an index without CONCURRENTLY, adds a validated foreign key, or changes a column type.
• Those take the slow path: nullable column, concurrent index, foreign key NOT VALID then validated, backfill in a job.
• Release checklist in Confluence ENG › Release process.