Re: Parallel bitmap heap scan

On Wed, Feb 8, 2017 at 1:59 AM, Dilip Kumar <dilipbalaut(at)gmail(dot)com> wrote:
> IIUC, tbm_prepare_shared_iterate will be called only by the leader,
> for tbmiterator as well as for the prefetch_iterator. And,
> tbm_attach_shared_iterate will be called by each backend and for both
> the iterators.

That's what I had in mind.

> IMHO, tbm_attach_shared_iterate should return TBMIterator with
> reference to TBMSharedIterator inside it. The reason behind same is
> that we can not keep TBMIterateResult inside TBMSharedIterator
> otherwise, results will also go in shared memory but we want to have
> local result memory for each worker so that other worker doesn't
> disturb it.

No, I don't agree. I think TBMSharedIterator should be an unshared
structure created by tbm_attach_shared_iterate, which can internally
contain backend-private state like a TBMIterateResult, and which can
also contain a pointer to the shared-memory structure previously
created by tbm_prepare_shared_iterate. That thing needs to be called
something other than a TBMSharedIterator, like TBMSharedIterationState
or something.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company