Putting it here because it isn't well known and bit me once.

NOTIFY payloads are deduped over the course of a transaction, with a O(n^2) algorithm if all payloads are different.

If you do a NOTIFY on delete or update and hit a few million rows, the query can go from seconds to hours.

The solution would be to write your payloads to a temp table and send them in batch when the transaction ends. Notifying in chunks of 100 will make the process 10'000 faster.

The best part of this is that the notifications are transactional! If other work in the transaction causes a rollback then the notification is never sent.