On 29/01/2026 03:05, Heikki Linnakangas wrote:
> Here's a new rebased and massively re-worked patch.
>
> The patches are split differently than in the previous version:
> - Patches 0001-0005 are just refactoring around recovery conflicts which
> I posted on a separate thread [1]. Please review and comment there.
> - Patches 0006 and 0007 are small cleanups that could be applied early
> (after updating the docs, as noted in TODO comment there).
> - All the interesting bits for this thread are in the last, massive patch.
>
> To review this, I suggest starting from the new src/backend/ipc/
> README.md file. It gives a good overview of the mechanism (or if it
> doesn't, that's valuable feedback :-) ). It also contains a bunch of
> Open Questions at the bottom; I'd love to hear opinions and ideas on those.
New version attached. Lots of little cleanups, and a few more notable
changes:
- Switched to 64-bit interrupt masks. This gives more headroom for
extensions to reserve their own custom interrupts
- I mostly gave up on sending interrupts from postmaster to child
processes. So postmaster still uses kill() to tell child processes to
exit or do other things. There's a TODO section in the README about
that, but this is fine for now.
- Implemented a RequestAddinInterrupt() function for extensions to
request interrupt bits. (TODO: use it in one of the example extensions,
to show how to use it and to test that it works)
> On 15/07/2025 18:50, Andres Freund wrote:
>>> @@ -205,6 +206,8 @@ StartupProcExit(int code, Datum arg)
>>> /* Shutdown the recovery environment */
>>> if (standbyState != STANDBY_DISABLED)
>>> ShutdownRecoveryTransactionEnvironment();
>>> +
>>> + ProcGlobal->startupProc = INVALID_PROC_NUMBER;
>>> }
>>
>>
>> What if we instead had a ProcGlobal->auxProc[auxproxtype]? We have
>> different
>> versions of this for different types auf aux processes, which doesn't
>> really
>> make sense.
>
> I like that idea, but didn't try it yet.
The attached patches move a little in that direction. There's no array
like that, but I moved the responsibility of setting those
startupProc/walreceiverProc/checkpointerProc fields to
InitAuxiliaryProcess(), so there's now a clear pattern to copy if other
processes need to be advertised like that.
- Heikki
