Hi Bertrand, I reviewed and tested v2; it builds cleanly. The implementation closely mirrors the existing per-backend IO/WAL stats (same flush path, struct layout, and backend-type guard), and the hot path is untouched: the per-backend counters piggyback on the existing global counting points (pgstat_count_lock_waits / pgstat_count_lock_fastpath_exceeded), so they only fire where pg_stat_lock already counts -- fastpath_exceeded when the fast-path slot limit is exceeded, and waits/wait_time only after a wait longer than deadlock_timeout.
One tiny nit: in pgstat_count_backend_lock_waits() and pgstat_count_backend_lock_fastpath_exceeded(), the Assert() is directly followed by the counter update. A blank line after the Assert() would read a bit better and is the more usual style in this code. Otherwise LGTM. Regards, Rui
