Re: What happens if I create new threads from within a postgresql function?

On Mon, Feb 18, 2013 at 09:56:22AM -0600, Merlin Moncure wrote:
> On Mon, Feb 18, 2013 at 5:10 AM, Seref Arikan
> <serefarikan(at)kurumsalteknoloji(dot)com> wrote:
> > Greetings,
> > What would happen if I create multiple threads from within a postgresql
> > function written in C?
> > I have the opportunity to do parallel processing on binary data, and I need
> > to create multiple threads to do that.
> > If I can ensure that all my threads complete their work before I exit my
> > function, would this cause any trouble ?
> > I am aware of postgresql's single threaded nature when executing queries,
> > but is this a limitation for custom multi threaded code use in C based
> > functions?
> > I can't see any problems other than my custom spawn threads living beyond my
> > function's execution and memory/resource allocation issues, but if I can
> > handle them, should not I be safe?
> >
> > I believe I've seen someone applying a similar principle to use GPUs with
> > postgresql, and I'm quite interested in giving this a try, unless I'm
> > missing something.
>
> Some things immediately jump to mind:
> *) backend library routines are not multi-thread safe. Notably, the
> SPI interface and the memory allocator, but potentially anything. So
> your spawned threads should avoid calling the backend API. I don't
> even know if it's safe to call malloc.
>
> *) postgres exception handling can burn you, so I'd be stricter than
> "before I exit my function"...really, you need to make sure threads
> terminate before any potentially exception throwing backend routine
> fires, which is basically all of them including palloc memory
> allocation and interrupt checking. So, we must understand that:
>
> While your threads are executing, your query can't be cancelled --
> only a hard kill will take the database down. If you're ok with that
> risk, then go for it. If you're not, then I'd thinking about
> sendinging the bytea through a protocol to a threaded processing
> server running outside of the database. More work and slower
> (protocol overhead), but much more robust.

You can see the approach of not calling any PG-specific routines from
theads here:

http://wiki.postgresql.org/wiki/Parallel_Query_Execution#Approaches

--
Bruce Momjian <bruce(at)momjian(dot)us> http://momjian.us
EnterpriseDB http://enterprisedb.com

+ It's impossible for everything to be true. +