Re: Streaming read-ready sequential scan code - Mailing list pgsql-hackers

From Melanie Plageman
Subject Re: Streaming read-ready sequential scan code
Date
Msg-id 20240401195848.qdlgt6u4dlyikm5k@liskov
Whole thread Raw
In response to Re: Streaming read-ready sequential scan code  (Melanie Plageman <melanieplageman@gmail.com>)
Responses Re: Streaming read-ready sequential scan code
List pgsql-hackers
On Wed, Mar 27, 2024 at 08:47:03PM -0400, Melanie Plageman wrote:
> On Fri, Mar 8, 2024 at 4:56 PM Melanie Plageman
> <melanieplageman@gmail.com> wrote:
> >
> > On Sat, Mar 02, 2024 at 06:07:48PM -0500, Melanie Plageman wrote:
> > > On Wed, Feb 28, 2024 at 12:30 PM Melanie Plageman
> > > <melanieplageman@gmail.com> wrote:
> > > >
> > > > On Mon, Feb 26, 2024 at 03:56:57PM -0500, Melanie Plageman wrote:
> > > > > On Mon, Feb 19, 2024 at 6:05 PM Melanie Plageman
> > > > > <melanieplageman@gmail.com> wrote:
> > > > > >
> > > > > > On Mon, Jan 29, 2024 at 4:17 PM Melanie Plageman
> > > > > > <melanieplageman@gmail.com> wrote:
> > > > > > >
> > > > > > > There is an outstanding question about where to allocate the
> > > > > > > PgStreamingRead object for sequential scans
> > > > > >
> > > > > > I've written three alternative implementations of the actual streaming
> > > > > > read user for sequential scan which handle the question of where to
> > > > > > allocate the streaming read object and how to handle changing scan
> > > > > > direction in different ways.
> > > > > >
> > > > > > Option A) https://github.com/melanieplageman/postgres/tree/seqscan_pgsr_initscan_allocation
> > > > > > - Allocates the streaming read object in initscan(). Since we do not
> > > > > > know the scan direction at this time, if the scan ends up not being a
> > > > > > forwards scan, the streaming read object must later be freed -- so
> > > > > > this will sometimes allocate a streaming read object it never uses.
> > > > > > - Only supports ForwardScanDirection and once the scan direction
> > > > > > changes, streaming is never supported again -- even if we return to
> > > > > > ForwardScanDirection
> > > > > > - Must maintain a "fallback" codepath that does not use the streaming read API
> > > > >
> > > > > Attached is a version of this patch which implements a "reset"
> > > > > function for the streaming read API which should be cheaper than the
> > > > > full pg_streaming_read_free() on rescan. This can easily be ported to
> > > > > work on any of my proposed implementations (A/B/C). I implemented it
> > > > > on A as an example.
> > > >
> > > > Attached is the latest version of this patchset -- rebased in light of
> > > > Thomas' updatees to the streaming read API [1]. I have chosen the
> > > > approach I think we should go with. It is a hybrid of my previously
> > > > proposed approaches.
> > >
> > > While investigating some performance concerns, Andres pointed out that
> > > the members I add to HeapScanDescData in this patch push rs_cindex and
> > > rs_ntuples to another cacheline and introduce a 4-byte hole. Attached
> > > v4's HeapScanDescData is as well-packed as on master and its members
> > > are reordered so that rs_cindex and rs_ntuples are back on the second
> > > cacheline of the struct's data.
> >
> > I did some additional profiling and realized that dropping the
> > unlikely() from the places we check rs_inited frequently was negatively
> > impacting performance. v5 adds those back and also makes a few other
> > very minor cleanups.
> >
> > Note that this patch set has a not yet released version of Thomas
> > Munro's Streaming Read API with a new ramp-up logic which seems to fix a
> > performance issue I saw with my test case when all of the sequential
> > scan's blocks are in shared buffers. Once he sends the official new
> > version, I will rebase this and point to his explanation in that thread.
> 
> Attached v6 has the version of the streaming read API mentioned here
> [1]. This resolved the fully-in-shared-buffers regressions
> investigated in that thread by Andres, Bilal, and Thomas.

Attached v7 has version 14 of the streaming read API as well as a few
small tweaks to comments and code.

I noticed that 0001 in the set posed a small regression from master for
a sequential scan of a relation already in shared buffers. While
investigating this, I saw that heapfetchbuf() was still not being
inlined (compiled at -O2) and when I promoted heapfetchbuf() from static
inline to static pg_attribute_always_inline, most of the very small
regression I saw went away. I don't know if I squashed the issue
entirely, though.

- Melanie

Attachment

pgsql-hackers by date:

Previous
From: Corey Huinker
Date:
Subject: Re: Statistics Import and Export
Next
From: Nathan Bossart
Date:
Subject: Re: Allow non-superuser to cancel superuser tasks.