Обсуждение: Configuring bgw_restart_time
I am working on an extension that uses background workers, and interested in adding code for it to auto-restart after a crash or server restart (similar to as-coded in worker_spi).
But I'm also interested in being able to configure bgw_restart_time using a GUC without having to restart the server, using only SIGHUP. For example, I want to normally have the worker restart after 10 seconds. But if I am doing maintenance on the server (without a db restart), I perhaps want to change this to -1 (BGW_NEVER_RESTART), kill the worker, do my business, then restart the worker. Or another reason would be my background worker has some bug and I want to disable it without having to restart my db server. For us as for many, a small outage for a db restart is expensive.
I have played around with this and done some digging around the codebase in bgworker.c (with my limited knowledge thus far of the pg codebase), and so far as I can tell, it isn't possible to change bgw_restart_time without a server restart. But I'm not sure if that's just because I don't know how this code works, or if the current libraries actually don't support modifying this part of the background worker. I am setting the GUC in _PG_init, but I can see that changing it after it has been registered has no effect unless I restart the server.
If indeed this is possible, I'd be very grateful for some insight on how to do it. I may even try to add such an example to worker_spi.
Thanks!
Jeremy
On Tue, 13 Aug 2019 at 20:37, Jeremy Finzel <finzelj@gmail.com> wrote:
I am working on an extension that uses background workers, and interested in adding code for it to auto-restart after a crash or server restart (similar to as-coded in worker_spi).
What pglogical does for this is use dynamic background workers with restart turned off. It does its own worker exit and restart handling from a manager worker that's an always-running static bgworker.
It's not ideal as it involves a considerable amount of extra work, but with BDR we rapidly found that letting postgres itself restart bgworkers was much too inflexible and hard to control. Especially given the issues around soft-crash restarts and worker registration (see thread https://www.postgresql.org/message-id/flat/534E6569.1080506%402ndquadrant.com) .
But I'm also interested in being able to configure bgw_restart_time using a GUC without having to restart the server, using only SIGHUP. For example, I want to normally have the worker restart after 10 seconds. But if I am doing maintenance on the server (without a db restart), I perhaps want to change this to -1 (BGW_NEVER_RESTART), kill the worker, do my business, then restart the worker.
Instead of doing that I suggest having a SQL-callable function that sets a flag in a shared memory segment used by the worker then sets the worker's ProcLatch to wake it up if it's sleeping. The flag can *ask* it to exit cleanly. If its exit code is 0 it will not be restarted.
You could also choose to have the worker exit with code 0 on SIGTERM, again causing itself to be unregistered and not restarted.