bug-2 (Postgres) — replace_oidc_roles read existing rows under default
READ COMMITTED with no row lock. Two concurrent OIDC callbacks for the
same user_id (racing token refreshes with differing claim sets) could
both observe the same baseline and produce a final role state matching
neither caller's intent. Adds .with_for_update() to the SELECT so the
existing rows for this user are locked for the duration of the
transaction.
The lock is per-user_id, not table-wide; unrelated user writes are
unaffected. Empty result sets acquire no locks, so a brand-new user
with no rows yet still allows two callers to proceed and merge via
ON CONFLICT DO NOTHING — that's a permissive race that self-heals on
the next reconciliation cycle, documented in code.
perf-1 (SQLite) — replace_oidc_roles took the SQLite global write
lock unconditionally via BEGIN IMMEDIATE before reading. Steady-state
re-logins (claims unchanged, no INSERT/DELETE needed) paid the lock
cost for nothing and serialised against unrelated writers.
Replaces with a double-check pattern: phase 1 reads under the default
deferred transaction (no write lock), computes the diff, and returns
(set(), set()) on no-op. Phase 2, only when mutation is needed,
commits the read txn, escalates to BEGIN IMMEDIATE, RE-READS, and
re-computes the diff under the lock before writing. The returned
(added, removed) reflects what was actually written, so caller logging
in apply_role_mapping stays truthful even when concurrent writers
shifted state between the two reads.
The OR IGNORE on insert is now defense-in-depth (the lock makes it
unnecessary) but kept as a safety net.
Patrick Buckley
d5087ef3b9