Re: Auxiliary Processes and MyAuxProc

Hi,

On 2020-03-19 11:35:41 +0100, Peter Eisentraut wrote:
> On 2020-03-18 17:07, Mike Palmiotto wrote:
> > On Wed, Mar 18, 2020 at 11:26 AM Mike Palmiotto
> > <mike(dot)palmiotto(at)crunchydata(dot)com> wrote:
> > >
> > > On Wed, Mar 18, 2020 at 10:17 AM Justin Pryzby <pryzby(at)telsasoft(dot)com> wrote:
> > > > Also, I saw this was failing tests both before and after my rebase.
> > > >
> > > > http://cfbot.cputube.org/
> > > > https://ci.appveyor.com/project/postgresql-cfbot/postgresql/builds/31535161
> > > > https://ci.appveyor.com/project/postgresql-cfbot/postgresql/builds/31386446
> > >
> > > Good catch, thanks. Will address this as well in the next round. Just
> > > need to set up a Windows dev environment to see if I can
> > > reproduce/fix.
> >
> > While I track this down, here is a rebased patchset, which drops
> > MySubprocessType and makes use of the MyBackendType.
>
> While working on (My)BackendType, I was attempting to get rid of
> (My)AuxProcType altogether. This would mostly work except that it's sort of
> wired into the pgstats subsystem (see NumBackendStatSlots). This can
> probably be reorganized, but I didn't pursue it further.

Hm. Why does the number of stat slots prevent dropping AuxProcType?

> Now, I'm a sucker for refactoring, but I feel this proposal is going into a
> direction I don't understand. I'd prefer that we focus around building out
> background workers as the preferred subprocess mechanism. Building out a
> second generic mechanism, again, I don't understand the direction. Are we
> hoping to add more of these processes? Make it extensible? The net lines
> added/removed by this patch series seems pretty neutral. What are we
> getting at the end of this?

I think background workers for internal processes are the *wrong*
direction to go. They were used as a shortcut for parallelism, and then
that was extended for logical replication. In my opinion that was done,
to a significant degree, because the aux proc stuff is/was too painful
to deal with, but that's something that should be fixed, instead of
building more and more parallel infrastructure.

Bgworkers are imo not actually a very good fit for internal
processes. We have to be able to guarantee that there's a free "slot" to
start internal processes, we should be able to efficiently reference
their pids (especially from postmaster, but also other processes), we
want to precisely know which shared PGPROC is being used, etc.

We now have somewhat different systems for at least: non-shmem
postmaster children, aux processes, autovacuum workers, internal
bgworkers, extension bgworkers. That's just insane.

We should merge those as much as possible. There's obviously going to be
some differences, but it needs to be less than now. I think we're
mostly on the same page on that, I just don't see bgworkers getting us
there.

The worst part about the current situation imo is that there's way too
many places that one needs to modify / check to create / understand a
process. Moving towards having a single c file w/ associated header that
defines 95% of that seems like a good direction. I've not looked at
recent versions of the patch, but there was some movement towards that
in earlier versions.

On a green field, I would say that there should be one or two C
arrays-of-structs defining subprocesses. And most behaviour should be
channeled through that.

struct PgProcessType
{
const char* name;
PgProcessBeforeShmem before_shmem;
PgProcessEntry entrypoint;
uint8:1 only_one_exists;
uint8:1 uses_shared_memory;
uint8:1 client_connected;
uint8:1 database_connected;
...
};

PgProcessType ProcessTypes[] = {
[StartupType] = {
.name = "startup",
.entrypoint = StartupProcessMain,
.only_one_exists = true,
.uses_shared_memory = true,
.client_connected = false,
.database_connected = false,
...
},
...
[UserBackendType] = {
.name = "backend",
.before_shmem = BackendInitialize,
.entrypoint = BackendRun, // fixme
.only_one_exists = false,
.uses_shared_memory = true,
.client_connected = true,
.database_connected = true,
...
}
...

Then there should be a single startup routine for all postmaster
children. Since most of the startup is actually shared between all the
different types of processes, we can declutter it a lot.

When starting a process postmaster would just specify the process type,
and if relevant, an argument (struct Port for backends, whatever
relevant for bgworkers etc) . Generic code should handle all the work
until the process type entry point - and likely we should move more work
from the individual process types into generic code.

If a process is 'only_one_exists' (to be renamed), the generic code
would also (in postmaster) register the pid as
subprocess_pids[type] = pid;
which would make it easy to only have per-type code in the few locations
that need to be aware, instead of many locations in
postmaster.c. Perhaps also some shared memory location.

Coming back to earth, from green field imagination land: I think the
patchset does go in that direction (to be fair, I think I outlined
something like the above elsewhere in this thread in a discussion with
Mike). And that's good.

Greetings,

Andres Freund