On Mon, Oct 20, 2025 at 01:01:31PM +1100, Peter Smith wrote:
> Some comments for the latest v8 patch.
The comments of Peter apply to comments and parameters. I am not
going down to these details in this message, these can be infinitely
tuned.
The injection point integration looks correct. You are checking the
compile flag and if the extension is available in the installation
path, which should be enough.
+ if (IS_INJECTION_POINT_ATTACHED("reduce-ncounts"))
+ ntries = 10;
1s is much faster than the default of 5s, still I am wondering if this
cannot be brought down a bit more. Dropping the worker still around
after the first test with CREATE DATABASE works here.
+# Confirm a background worker is still running
+$node->safe_psql(
+ "postgres", qq(
+ SELECT count(1) FROM pg_stat_activity
+ WHERE backend_type = 'worker_spi dynamic';));
This does not check that the worker that does not have the flag set is
still running: you are not feeding the output of this query to an is()
test.
+ is($result, 't', "dynamic bgworker launched");
In launch_bgworker(), this uses the same test description for all the
callers of this subroutine. Let's prefix it with $testcase.
+void
+TerminateBgWorkersByDbOid(Oid oid)
FWIW, while reading this code, I was wondering about one improvement
that could show benefits for more extension code than only what we are
discussing here because external code has no access to
BackgroundWorkerSlot while holding the LWLock BackgroundWorkerLock in
a single loop, by rewriting this new routine with something like that:
void TerminateBackgroundWorkerMatchin(
bool (*do_terminate) (int pid, BackgroundWorker *, Datum))
Then the per-database termination would be a custom routine, defined
also in bgworker.c. Other extension code could define their own
filtering callback routine. Just an idea in passing, to let extension
code take more actions on bgworker slots in use-based on a PGPROC
entry, like a role ID for example, or it could be a different factor.
Feel free to dislike such a funky idea if you do not like it and say
so, of course.
--
Michael
