On Fri, Mar 22, 2024 at 11:25 AM Fred Habash <fmhabash@gmail.com> wrote:
Facing an issue where sometimes humans login to a database and run DDL statements causing a long locking tree of over 1000 waiters. As a workaround, we asked developers to always start their DDL sessions with 'SET lock_timeout = 'Xs'.
I reviewed the native lock timeout parameter in Postgres and found 7. None seem to be related to blocker timeouts directly.
idle_in_transaction_session_timeout idle_session_timeout lock_timeout: How long a session waits for a lock statement_timeout authentication_timeout deadlock_timeout log_lock_waits
Instead, I put together a quick procedure that counts waiter sessions for a given blocker and terminates it if waiter count exceeds a threshold.
Is there not a native way to ...
1. Automatically time out a blocker
2. A metric that shows how many waiters for a blocker?
I guess this probably does not belong in the native codebase because in most real world scenarios with contention you would end up with priority inversion or a situation where no work gets done. With current locking rules, theoretically the work queue would always clear (assuming the locker doesn't hold the transaction indefinitely), where with your setting enabled it might not always assume the locker retries.
In your case, a hand written 'unblocker' script might be the way to go, or (probably better) encourage patterns where critical tables are not blocked, say by building up a scratch table and swapping in on a separate transaction. Reducing contention rather than mitigating the symptoms of it, is *always* a good thing.