Trying out read streams in pgvector (an extension)

On 11/06/2024 07:53, Thomas Munro wrote:

Someone involved in that project mentioned that it's probably not a
great topic to research in practice, because real world users of HNSW
use fully cached ie prewarmed indexes, because the performance is so
bad otherwise. (Though maybe that argument is a little circular...).

I think that's true in practice for *building* an HNSW index, but faster
*searching* when the index is not in memory seems quite useful. And of
course, faster is always better, even if it's only in a non-optimal
scenario.

So although this patch clearly speeds up cold HSNW searches to a
degree controlled by effective_io_concurrency, I'll probably look for
something else. Suggestions for interesting index types to look at
streamifying are very welcome!

GiST and GIN?

Hmm. If that's really true about HNSW though, then there may still be
an opportunity to do automatic memory prefetching[1]. But then in the
case of index building, "stream" is NULL in this patch anyway. It
surely must also be possible to find some good places to put
profitable explicit pg_mem_prefetch() calls given the predictability
and the need to get only ~60ns ahead for that usage. I didn't look
into that because I was trying to prove things about read_stream.c,
not get involved in another project :-D

Here ends my science experiment report, which I'm dropping here just
in case others see useful ideas here. The main thing I learned about
the read stream API is that it'd be nice to be able to reset the
stream but preserve the distance (something that came up on the
streaming sequential scan thread for a different reason), to deal with
cases where look-ahead opportunities come in bursts but you want a
longer lived stream than I used here. That is the reason the patch
creates and destroys temporary streams in a loop; doh. It also
provides an interesting case study for what speculative random
look-ahead support might need to look like.

This reminds me of a prototype I wrote earlier, see
https://github.com/pgvector/pgvector/pull/386, 1st commit. It
reorganizes HnswSearchLayer() so that it in iteration, it first collects
all the neighbors to visit, and then visits them, somewhat similar to
your patch.

--
Heikki Linnakangas
Neon (https://neon.tech)

On 6/11/24 12:53 AM, Thomas Munro wrote:

Hi,

I was looking around for an exotic index type to try the experience of
streamifying an extension, ie out-of-core code. I am totally new to
pgvector, but since everyone keeps talking about it, I could not avoid
picking up some basic facts in the pgconf.dev hallway track, and
understood that its scans have some degree of known-order access
predictability, and then also some degree of fuzzy-predictable
order-not-yet-determined access too. It's also quite random in the
I/O sense.

Cool! I happened to be chatting w/Andrew about this yesterday to see if
there could be some benefits for folks who are running pgvector on PG17.

Here's a toy to streamify the known-order part. I think for the fuzzy
part that links those parts together, maybe there is some way to guess
when it's a reasonable time to speculatively prefetch the lowest order
stuff in the pairing heap, and then deal with it if you're wrong, but
I didn't try that...

I would suggest submitting this at least as a draft PR to the pgvector
project[1]https://github.com/pgvector/pgvector:

https://github.com/pgvector/pgvector

Someone involved in that project mentioned that it's probably not a
great topic to research in practice, because real world users of HNSW
use fully cached ie prewarmed indexes, because the performance is so
bad otherwise.

I don't think that was me, at least in those words (and I had noted I'd
love to chat w/you about this, but we didn't find time). Stating it
differently, the "ideal" is to keep the indexes in memory, as that leads
to the best performance, but reality is more complicated. These datasets
are quite large (e.g. the 1536-dim vector is a 6KB payload, excluding
what's in the index) and if you're storing the full vector in the index
(there are now some quantization methods available[4]https://jkatz05.com/post/postgres/pgvector-scalar-binary-quantization/), you can easily
double your dataset size, and quickly exceed available memory. So I
think in the real world, you're more likely to see swapping pages
between disk and memory. Some of this was addressed in the talk @
PGConf.dev[3]https://www.pgevents.ca/events/pgconfdev2024/schedule/session/1-vectors-how-to-better-support-a-nasty-data-type-in-postgresql/ (slides here[2]https://www.pgevents.ca/events/pgconfdev2024/sessions/session/1/slides/42/pgconfdev-2024-vectors.pdf).

(Though maybe that argument is a little circular...).
So although this patch clearly speeds up cold HSNW searches to a
degree controlled by effective_io_concurrency, I'll probably look for
something else. Suggestions for interesting index types to look at
streamifying are very welcome!

Yup, so this makes sense for HNSW particularly at the higher-level
pages. But it may make more sense for IVFFlat, given how it clusters
data. With IVFFlat, you first find your lists/centers, and then you
determine how you index each vector around the lists. When those lists
are stored to disk, they're basically sequential. A lot of the struggles
with IVFFlat is both the long load from disk and ultimately some
comptuational issues for a larger set of vector comparisons (though if
you're able to build small, efficient clusters, it can be much faster
than HNSW!). HNSW behaves more like a (bear with me) typically
"tree-based" index, where you'll have hot spots at the top, but because
of the nature of vector search, the lower levels tend to be more random
in access.

Regardless, the part where this is interesting (at least to me) is that
a lot of these vectors tend to take up a full page anyway, so anything
we can do to read them faster from disk will generally get a thumbs up
from me.

Hmm. If that's really true about HNSW though, then there may still be
an opportunity to do automatic memory prefetching[1]. But then in the
case of index building, "stream" is NULL in this patch anyway. It
surely must also be possible to find some good places to put
profitable explicit pg_mem_prefetch() calls given the predictability
and the need to get only ~60ns ahead for that usage. I didn't look
into that because I was trying to prove things about read_stream.c,
not get involved in another project :-D

Well, as alluded to in[2]https://www.pgevents.ca/events/pgconfdev2024/sessions/session/1/slides/42/pgconfdev-2024-vectors.pdf, thinking about how another project uses this
will certainly help, and anything we can do to continue to speed up
vector queries helps PostgreSQL ;) Some of the contributions from folks
who have focused on core have significantly helped pgvector.

Here ends my science experiment report, which I'm dropping here just
in case others see useful ideas here. The main thing I learned about
the read stream API is that it'd be nice to be able to reset the
stream but preserve the distance (something that came up on the
streaming sequential scan thread for a different reason), to deal with
cases where look-ahead opportunities come in bursts but you want a
longer lived stream than I used here. That is the reason the patch
creates and destroys temporary streams in a loop; doh. It also
provides an interesting case study for what speculative random
look-ahead support might need to look like.

If you're curious, I can fire up some of my more serious benchmarks on
this to do a before/after to see if there's anything interesting. I have
a few large datasets (10s of millions) of larger vectors (1536dim => 6KB
payloads) that could see the net effect here.

(Make sure you remember to set effective_io_concurrency to an
interesting number if you want to generate a lot of overlapping
fadvise calls.)

What would you recommend as an "interesting number?" - particularly
using the data parameters above.

Thanks,

Jonathan

[1]: https://github.com/pgvector/pgvector
[2]: https://www.pgevents.ca/events/pgconfdev2024/sessions/session/1/slides/42/pgconfdev-2024-vectors.pdf
https://www.pgevents.ca/events/pgconfdev2024/sessions/session/1/slides/42/pgconfdev-2024-vectors.pdf
[3]: https://www.pgevents.ca/events/pgconfdev2024/schedule/session/1-vectors-how-to-better-support-a-nasty-data-type-in-postgresql/
https://www.pgevents.ca/events/pgconfdev2024/schedule/session/1-vectors-how-to-better-support-a-nasty-data-type-in-postgresql/
[4]: https://jkatz05.com/post/postgres/pgvector-scalar-binary-quantization/

On Wed, Jun 12, 2024 at 3:37 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:

If you're curious, I can fire up some of my more serious benchmarks on
this to do a before/after to see if there's anything interesting. I have
a few large datasets (10s of millions) of larger vectors (1536dim => 6KB
payloads) that could see the net effect here.

(Make sure you remember to set effective_io_concurrency to an
interesting number if you want to generate a lot of overlapping
fadvise calls.)

What would you recommend as an "interesting number?" - particularly
using the data parameters above.

Hi Jonathan,

Sorry for not replying sooner (ETOOMANYPROJECTS). For HNSW, I think
the maximum useful effective_io_concurrency is bound by the number of
connections per HNSW layer ("m"). Here are some times I measured
using m=16 on two laptops:

| linux (xfs) | macos (apfs)
branch | eic | avg | speedup | stdev | avg | speedup | stdev
--------+-----+--------+---------+--------+--------+---------+--------
master | | 73.959 | 1.0 | 24.168 | 72.290 | 1.0 | 11.851
stream | 0 | 70.117 | 1.1 | 36.699 | 76.289 | 1.0 | 12.742
stream | 1 | 57.983 | 1.3 | 5.845 | 79.969 | 1.2 | 8.308
stream | 2 | 35.629 | 2.1 | 4.088 | 49.198 | 2.0 | 7.686
stream | 3 | 28.477 | 2.6 | 2.607 | 37.540 | 2.5 | 5.272
stream | 4 | 26.493 | 2.8 | 3.691 | 33.014 | 2.7 | 4.444
stream | 5 | 23.711 | 3.1 | 2.435 | 32.622 | 3.0 | 2.270
stream | 6 | 22.885 | 3.2 | 1.908 | 31.254 | 3.2 | 4.170
stream | 7 | 21.910 | 3.4 | 2.153 | 33.669 | 3.3 | 4.616
stream | 8 | 20.741 | 3.6 | 1.594 | 34.182 | 3.5 | 3.819
stream | 9 | 22.471 | 3.3 | 3.094 | 30.690 | 3.2 | 2.677
stream | 10 | 19.895 | 3.7 | 1.695 | 32.631 | 3.6 | 4.976
stream | 11 | 19.447 | 3.8 | 1.647 | 31.163 | 3.7 | 3.351
stream | 12 | 18.658 | 4.0 | 1.503 | 30.817 | 3.9 | 3.538
stream | 13 | 18.886 | 3.9 | 0.874 | 29.184 | 3.8 | 4.832
stream | 14 | 18.667 | 4.0 | 1.692 | 28.783 | 3.9 | 3.459
stream | 15 | 19.080 | 3.9 | 1.429 | 28.928 | 3.8 | 3.396
stream | 16 | 18.929 | 3.9 | 3.469 | 29.282 | 3.8 | 2.868

Those are millisecond times to run the test() function shown earlier,
with empty kernel cache and PostgreSQL cache (see below) for maximum
physical I/O. I ran the master test 30 times, and each
effective_io_concurrency level 10 times, to show that the variance
decreases even at the default effective_io_concurency = 1, so we're
not only talking about the avg speed improving.

The all-cached performance also seems to improve, ~8.9ms -> ~6.9ms on
Linux, but I can't fully explain why that is, maybe just some random
stuff about memory layout run-to-run in my quick and dirty test or
something like that, so I'm not claiming that is significant. It
certainly didn't get slower, anyway.

I think you would get very different numbers on a high latency storage
system (say, non-local cloud storage) and potentially much more
speedup with your large test indexes. Also my 6d random number test
may not be very representative and you may be able to come up with
much better tests.

Here's a new version with a TODO tidied up. I also understood that we
need to tweak the read_stream_reset() function, so that it doesn't
forget its current readhead distance when it hops between HNSW nodes
(which is something that comes up in several other potential uses
cases including another one I am working in in core). Without this
patch for PostgreSQL, it reads 1, 2, 4, 7 blocks (= 16 in total)
before it has to take a break to hop to a new page, and then it start
again at 1. Oops. With this patch, it is less forgetful, and reaches
the full possible I/O concurrency of 16 (or whatever the minimum of
HNSW's m parameter and effective_io_concurrency is for you).

PSA two patches, one for PostgreSQL and one for pgvector.

I am not actively working on this right now. If someone wants to try
to develop it further, please feel free! I haven't looked at IVFFlat
at all.

--- function to let you do SELECT uncache('t_embedding_idx'),
--- which is the opposite of SELECT pg_prewarm('t_embedding_idx')
--- see also "echo 1 | sudo tee /proc/sys/vm/drop_caches" (Linux)
--- "sudo purge" (macOS)
create extension pg_buffercache;
create or replace function uncache(name text) returns bool
begin atomic;
  select bool_and(pg_buffercache_evict(bufferid))
    from pg_buffercache where relfilenode = name::regclass;
end;

On Fri, Sep 6, 2024 at 4:28 PM Thomas Munro <thomas.munro@gmail.com> wrote:

Without this
patch for PostgreSQL, it reads 1, 2, 4, 7 blocks (= 16 in total)
before it has to take a break to hop to a new page, and then it start
again at 1. Oops.

Erm, correction: 1, 2, 4, 8, 1 (because it runs out due to m == 16 and resets).

There was a mistake in my query, so the macOS speedup column was wrong
(was accidentally comparing Linux number with macOS master, sorry for
the noise). I also forgot to mention that you don't actually get the
speedup on PostgreSQL 17 on a Mac, because Peter only recently
implemented the needed read-ahead support for macOS in master/18, but
it doesn't get slower. Here's the corrected table:

| linux (xfs) | macos (apfs)
branch | eic | avg | speedup | stdev | avg | speedup | stdev
--------+-----+--------+---------+--------+--------+---------+--------
master | | 73.959 | 1.0 | 24.168 | 72.290 | 1.0 | 11.851
stream | 0 | 70.117 | 1.1 | 36.699 | 76.289 | 0.9 | 12.742
stream | 1 | 57.983 | 1.3 | 5.845 | 79.969 | 0.9 | 8.308
stream | 2 | 35.629 | 2.1 | 4.088 | 49.198 | 1.5 | 7.686
stream | 3 | 28.477 | 2.6 | 2.607 | 37.540 | 1.9 | 5.272
stream | 4 | 26.493 | 2.8 | 3.691 | 33.014 | 2.2 | 4.444
stream | 5 | 23.711 | 3.1 | 2.435 | 32.622 | 2.2 | 2.270
stream | 6 | 22.885 | 3.2 | 1.908 | 31.254 | 2.3 | 4.170
stream | 7 | 21.910 | 3.4 | 2.153 | 33.669 | 2.1 | 4.616
stream | 8 | 20.741 | 3.6 | 1.594 | 34.182 | 2.1 | 3.819
stream | 9 | 22.471 | 3.3 | 3.094 | 30.690 | 2.4 | 2.677
stream | 10 | 19.895 | 3.7 | 1.695 | 32.631 | 2.2 | 4.976
stream | 11 | 19.447 | 3.8 | 1.647 | 31.163 | 2.3 | 3.351
stream | 12 | 18.658 | 4.0 | 1.503 | 30.817 | 2.3 | 3.538
stream | 13 | 18.886 | 3.9 | 0.874 | 29.184 | 2.5 | 4.832
stream | 14 | 18.667 | 4.0 | 1.692 | 28.783 | 2.5 | 3.459
stream | 15 | 19.080 | 3.9 | 1.429 | 28.928 | 2.5 | 3.396
stream | 16 | 18.929 | 3.9 | 3.469 | 29.282 | 2.5 | 2.868

On Fri, Sep 6, 2024 at 4:28 PM Thomas Munro <thomas.munro@gmail.com> wrote:

Here's a new version with a TODO tidied up. I also understood that we
need to tweak the read_stream_reset() function, so that it doesn't
forget its current readhead distance when it hops between HNSW nodes
(which is something that comes up in several other potential uses
cases including another one I am working in in core). Without this
patch for PostgreSQL, it reads 1, 2, 4, 7 blocks (= 16 in total)
before it has to take a break to hop to a new page, and then it start
again at 1. Oops. With this patch, it is less forgetful, and reaches
the full possible I/O concurrency of 16 (or whatever the minimum of
HNSW's m parameter and effective_io_concurrency is for you).

I heard that the pgvector project is now trying to do this for real,
and (surprise!) running into this problem. It causes streamified HNSW
search to regress in performance on some queries when the overheads of
streaming are not outweighed by the (bogusly constrained) gains in
concurrency. We just don't generate enough concurrency to win. I
probably should have been more opinionated and just committed a
version of that distance-reset policy change, but I guess at the time
I wrote the above, streaming and AIO were a little too abstract to
attract reviews relating to hypothetical external projects.

We definitely want to fix that for v19, because it also affects the
streamified index scan project and doubtless many other things. I
wrote about that with patches[1]/messages/by-id/CA+hUKGL6hCd40Dh1AcFcoiw5zJXK2T1dRKO3oe8RkPExqA5zoQ@mail.gmail.com and will start a new thread soon with
a new collection of rebased heuristics improvements.

But for now, to fix pgvector's woes, I wonder if it might make sense
to call this a bug in v18, and back-patch the tiniest possible change.
Something like what I posted[2]/messages/by-id/CA+hUKG+x2BcqWzBC77cN0ewhzMF0kYhC6c4G_T2gJLPbqYQ6Ow@mail.gmail.com in this thread almost two years ago.
I don't think it really affects any core code: we use
read_stream_reset() only in very minimal ways there (I could
elaborate), and it's quite arguable that the existing policy is wrong
for them too, but we'd need to confirm that and perhaps think about
other extensions that might be using it.

Better ideas?

[1]: /messages/by-id/CA+hUKGL6hCd40Dh1AcFcoiw5zJXK2T1dRKO3oe8RkPExqA5zoQ@mail.gmail.com
[2]: /messages/by-id/CA+hUKG+x2BcqWzBC77cN0ewhzMF0kYhC6c4G_T2gJLPbqYQ6Ow@mail.gmail.com

On Tue, Nov 11, 2025 at 4:22 PM Thomas Munro <thomas.munro@gmail.com> wrote:

But for now, to fix pgvector's woes, I wonder if it might make sense
to call this a bug in v18, and back-patch the tiniest possible change.
Something like what I posted[2] in this thread almost two years ago.
I don't think it really affects any core code: we use
read_stream_reset() only in very minimal ways there (I could
elaborate), and it's quite arguable that the existing policy is wrong
for them too, but we'd need to confirm that and perhaps think about
other extensions that might be using it.

If we are worried about regressing other extensions using
read_stream_reset(), we could make the read stream reset which
preserves the distance a different function in backbranches.

- Melanie

On Wed, Nov 12, 2025 at 11:52 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:

On Tue, Nov 11, 2025 at 4:22 PM Thomas Munro <thomas.munro@gmail.com> wrote:

But for now, to fix pgvector's woes, I wonder if it might make sense
to call this a bug in v18, and back-patch the tiniest possible change.
Something like what I posted[2] in this thread almost two years ago.
I don't think it really affects any core code: we use
read_stream_reset() only in very minimal ways there (I could
elaborate), and it's quite arguable that the existing policy is wrong
for them too, but we'd need to confirm that and perhaps think about
other extensions that might be using it.

If we are worried about regressing other extensions using
read_stream_reset(), we could make the read stream reset which
preserves the distance a different function in backbranches.

Hmm, yeah, interesting idea. Candidate names might include
read_stream_restart() and read_stream_continue(). The point being
that the block number callback reported end-of-stream, but that was
only temporary, and now it has more information and would like to
continue. Those are some of the names I bounced around for a new
read_stream_reset() flag argument for v19 (I rather liked "continue"),
but I also like this separate function idea. Back-patching a new
function would certainly remove all doubt about unintended
consequences for existing callers of read_stream_reset(), so yeah,
that wins on pure conservative safety grounds. As for the future,
hmm, it might even be better to have an explicit separate API for this
operation in master too, as it is turning out to be quite a common
requirement and the naming is much clearer like that. We don't
usually design new APIs while back-patching though, that's probably
why I didn't think of that, but if we view this as a design bug that
folded too many jobs into read_stream_reset() that we now want to fix
by splitting one off, maybe that's OK? Seems pretty risk-free,
anyway.

On Wed, Nov 12, 2025 at 12:19 PM Thomas Munro <thomas.munro@gmail.com> wrote:

On Wed, Nov 12, 2025 at 11:52 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:

If we are worried about regressing other extensions using
read_stream_reset(), we could make the read stream reset which
preserves the distance a different function in backbranches.

Here is a draft patch like that, that tries to be as small as
possible. Trying out the name read_stream_resume().

Hi,

Thank you for working on this!

On Wed, 12 Nov 2025 at 07:12, Thomas Munro <thomas.munro@gmail.com> wrote:

On Wed, Nov 12, 2025 at 12:19 PM Thomas Munro <thomas.munro@gmail.com> wrote:

On Wed, Nov 12, 2025 at 11:52 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:

If we are worried about regressing other extensions using
read_stream_reset(), we could make the read stream reset which
preserves the distance a different function in backbranches.

Here is a draft patch like that, that tries to be as small as
possible. Trying out the name read_stream_resume().

I liked the idea of having a different function named
read_stream_resume for this purpose.

0001 looks good to me.

0002:

+        /* End-of-stream. */
+        buf = read_stream_next_buffer(stream, NULL);
+        Assert(buf == InvalidBuffer);
+        buf = read_stream_next_buffer(stream, NULL);
+        Assert(buf == InvalidBuffer);

I noticed there are two 'read_stream_next_buffer()' and
'InvalidBuffer' checks. Does having both provide any additional
validation? I tried removing one of them, and the test still passed.

Also, there is one thing I wanted to clarify about the
'read_stream_resume()'. If 'read_stream_next_buffer()' returns an
'InvalidBuffer', then we can use 'read_stream_resume()' alone because
we know that we already consumed all buffers in the stream. For the
rest, we need to use 'read_stream_resume()' together with the
'read_stream_reset()', right?

--
Regards,
Nazir Bilal Yavuz
Microsoft

On Wed, Nov 12, 2025 at 9:04 PM Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:

On Wed, 12 Nov 2025 at 07:12, Thomas Munro <thomas.munro@gmail.com> wrote:
0002:

+        /* End-of-stream. */
+        buf = read_stream_next_buffer(stream, NULL);
+        Assert(buf == InvalidBuffer);
+        buf = read_stream_next_buffer(stream, NULL);
+        Assert(buf == InvalidBuffer);

I noticed there are two 'read_stream_next_buffer()' and
'InvalidBuffer' checks. Does having both provide any additional
validation? I tried removing one of them, and the test still passed.

I wanted to demonstrate that this is a state that the stream is stuck
in until you call _resume().

I suppose an alternative design would be that _next_buffer() returns
InvalidBuffer only once (= the block number callback returns
InvalidBlock once) and then automatically resumes (= it restores the
distance) and then you can call read_stream_next_buffer() again (= the
block number callback will be called again to fill the stream up with
new buffers before waiting for the first one to be ready to give to
you if it isn't already). That would have the advantage of not
requiring a new function at all and make the patch even shorter, but I
don't know, I guess I thought that would be a bit more fragile in some
way, less explicit. Hmm, would it actually be better if it worked
like that?

Also, there is one thing I wanted to clarify about the
'read_stream_resume()'. If 'read_stream_next_buffer()' returns an
'InvalidBuffer', then we can use 'read_stream_resume()' alone because
we know that we already consumed all buffers in the stream. For the
rest, we need to use 'read_stream_resume()' together with the
'read_stream_reset()', right?

For the rest, there would be no need to call read_stream_resume().

To recap the uses of read_stream_reset(): the original purpose was to
release any buffers (pins) that the stream is holding internally
because of look-ahead, and put it back to its original state, ready to
be reused. It is called (1) by read_stream_end() as an implementation
detail (eg if a LIMIT or anything else except ERROR/FATAL ends your
query early, we need to unpin buffers queued in the stream before we
pfree it), (2) explicitly by rescans, (3) in hypothetical code I
thought about that would want to stream blocks speculatively and then
change its mind after predicting incorrectly (I had a few patches like
that, abandoned for now), and then (4) in this case, by places that
temporarily ran out of block numbers, but will have some more again
soon and want to resume the stream.

It was already debatable whether heuristic state like lookahead
distance should survive acoss rescans, or in other words, whether the
expected I/O requirements of the previous scan are a useful prediction
of the requirements of the next scan, but the answer is clearer in
case (4), hence the desire to find a way to separate that use case
from the others.

On Tue, Nov 11, 2025 at 11:12 PM Thomas Munro <thomas.munro@gmail.com> wrote:

Here is a draft patch like that, that tries to be as small as
possible. Trying out the name read_stream_resume().

I like read_stream_resume(). Tested out 0001 with pgvector and can
confirm it works.

In the test, I would initialize test_read_stream_resume_state.count to 0

+ test_read_stream_resume_state state = {.blkno = blkno};

- Melanie

On Wed, Nov 12, 2025 at 5:47 AM Thomas Munro <thomas.munro@gmail.com> wrote:

I suppose an alternative design would be that _next_buffer() returns
InvalidBuffer only once (= the block number callback returns
InvalidBlock once) and then automatically resumes (= it restores the
distance) and then you can call read_stream_next_buffer() again (= the
block number callback will be called again to fill the stream up with
new buffers before waiting for the first one to be ready to give to
you if it isn't already). That would have the advantage of not
requiring a new function at all and make the patch even shorter, but I
don't know, I guess I thought that would be a bit more fragile in some
way, less explicit. Hmm, would it actually be better if it worked
like that?

We discussed off-list and decided that changing existing functionality
in an unexpected way is undesirable. So, it is better we stick with
adding read_stream_resume. However, in talking about
read_stream_resume() further, Thomas and I also thought of potential
issues with it:

If read_stream_resume() is called before the read stream user callback
has ever returned InvalidBlockNumber,
1) The value of resume_distance will be the original value of distance
from read_stream_begin_relation(). You don't want to reset the
distance to that value.
2) There may be inflight or completed buffers that have yet to be
yielded which will be returned the next time read_stream_next_buffer()
is invoked. If the user resets the state the callback is using to
return blocks and expects the next invocation of
read_stream_next_buffer() to return buffers with those blocks, they
will be disappointed.

If we try to address this by requiring that stream->distance is 0 when
read_stream_resume() is called, that won't work because while it is
set to 0 when the callback returns InvalidBlockNumber, there may still
be unreturned buffers in the stream.

If the user wants to use read_stream_reset() to exhaust the stream
before calling read_stream_resume(), read_stream_reset() sets
stream->distance to 1 at the end, so read_stream_resume() couldn't
detect if reset() was correctly called first or if the distance is > 0
because the stream is still in progress.

To make sure 1) distance isn't reset to a resume_distance from
read_stream_begin_relation() and 2) unexpected buffers aren't returned
from the read stream, we could error out in read_stream_resume() if
pinned_buffers > 0. And in read_stream_reset(), we would save distance
in resume_distance before clearing distance. That would allow calling
read_stream_resume() either if you called read_stream_reset() or if
you exhausted the stream yourself. See rough attached patch for a
sketch of this.

It would be nicer if we could error out if read_stream_next_buffer()
didn't return InvalidBuffer, but we can't do that if we want to allow
calling read_stream_reset() followed by read_stream_resume() because
distance won't be 0.

I tried this with a modified pgvector with an hnsw read stream user
and it seemed to work correctly.

- Melanie

Hi,

On Wed, 19 Nov 2025 at 00:17, Melanie Plageman
<melanieplageman@gmail.com> wrote:

To make sure 1) distance isn't reset to a resume_distance from
read_stream_begin_relation() and 2) unexpected buffers aren't returned
from the read stream, we could error out in read_stream_resume() if
pinned_buffers > 0. And in read_stream_reset(), we would save distance
in resume_distance before clearing distance. That would allow calling
read_stream_resume() either if you called read_stream_reset() or if
you exhausted the stream yourself. See rough attached patch for a
sketch of this.

This looks correct to me. What do you think about using an assert
instead of erroring out?

--
Regards,
Nazir Bilal Yavuz
Microsoft

On Wed, Nov 19, 2025 at 2:28 AM Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:

To make sure 1) distance isn't reset to a resume_distance from
read_stream_begin_relation() and 2) unexpected buffers aren't returned
from the read stream, we could error out in read_stream_resume() if
pinned_buffers > 0. And in read_stream_reset(), we would save distance
in resume_distance before clearing distance. That would allow calling
read_stream_resume() either if you called read_stream_reset() or if
you exhausted the stream yourself. See rough attached patch for a
sketch of this.

This looks correct to me. What do you think about using an assert
instead of erroring out?

I'm not totally opposed to this. My rationale for making it an error
is that the developer could have test cases where all the buffers are
consumed but the code is written such that that won't always happen.
Then if a real production query doesn't consume all the buffers, it
could return wrong results (I think). That will mean the user can't
complete their query until the extension author releases a new version
of their code. But I'm not sure what the right answer is here.

- Melanie

On Fri, Nov 21, 2025 at 4:28 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:

I'm not totally opposed to this. My rationale for making it an error
is that the developer could have test cases where all the buffers are
consumed but the code is written such that that won't always happen.
Then if a real production query doesn't consume all the buffers, it
could return wrong results (I think). That will mean the user can't
complete their query until the extension author releases a new version
of their code. But I'm not sure what the right answer is here.

Focusing on making sure v19 has a good interface for this, and
abandoning thoughts of back-patching a bandaid, and the constraints
that leads to, for now...

I think it'd be better if that were the consumer's choice. I don't
want the consumer to be required to drain the stream before resuming,
as that'd be an unprincipled stall. For example, if new WAL arrives
over the network then I think it should be possible for recovery's
WAL-powered stream of heap pages to resume looking ahead even if
recovery hasn't drained the existing stream completely.

Peter G (CC'd) and I discussed some problems he had in the index
prefetching work, and I tried to extend this a bit to give the
semantics he wanted, in point 2 below. It's simple itself, but might
lead to some tricky questions higher up. Posted for experimentation.
It'll be interesting to see if this goes somewhere.

1. read_stream_resume() as before, but with a new explicit
read_stream_pause(): if a block number callback would like to report a
temporary lack of information, it should return
read_stream_pause(stream), not InvalidBlockNumber. Then after
read_stream_resume(stream) is called, the next
read_stream_next_buffer() enters the lookahead loop again. While
paused, if the consumer drains all the existing buffers in the stream
and then one more, it will receive InvalidBuffer, but if the _resume()
call is made sooner, the consumer won't ever know about the temporary
lack of buffers in the stream.

2. read_stream_yield(): while streaming heap pages that come from
TIDs on index pages, Peter didn't like that the executor lost control
of how much work was done by the lookahead loop underneath
read_stream_next_buffer(). The consumer might have a heap page with
some tuples that could be emitted right now, but the block number
callback might be evaluating arbitrarily expensive filter qual
expressions far ahead, and they might prefer to emit more tuples now
before doing an unbounded amount of work finding more. This interface
allows some limited coroutine-like multitasking, where the block
number callback can return read_stream_yield(stream) to return control
back to the consumer periodically if it knows the consumer could
already do something else. It works by pausing the stream and
resuming it in the next read_stream_next_buffer() call, but that's an
internal detail.

Some half-baked thoughts about the resulting flow control:

Yielding control periodically just when it happens to be possible
within the constraints of the volcano executor is an interesting thing
to think about. You can only yield if you already have a tuple to
emit. There is no saying when control will return to you, and the
node you yield to might immediately block on I/O and yet you could
have been doing useful CPU work. You probably need an event-driven
node-hopping executor to fix that in general, but on the flip side, I
can think of one bet that I'd take: if you already have a tuple to
emit AND if index scans themselves (not only referenced heap pages)
were also streamed AND if a hypothetical
read_stream_next_buffer_no_wait(btree_stream) said the next index page
you need is not ready yet, then you should yield. You're gambling
that other plan nodes will have better luck running without an I/O
stall, but you have ~0% chance.

Yielding just because you've scanned N index pages/tuples/whatever is
harder to think about. The stream shouldn't get far ahead unless it's
recently been useful for I/O concurrency (though optimal distance
heuristics are an open problem), but in this case a single invocation
of the block number callback can call ReadBuffer() an arbitrary number
of times, filtering out all the index tuples as it rampages through
the whole index IIUC. I see why you might want to yield periodically
if you can, but I also wonder how much that can really help if you
still have to pick up where you left off next time. I guess it
depends on the distribution of matches. It's also clear that any
cold-cache testing done with direct I/O enabled will stall abominably
as long as that level calls ReadBuffer(), possibly confusing matters.

On Mon, Dec 8, 2025 at 10:47 PM Thomas Munro <thomas.munro@gmail.com> wrote:

I think it'd be better if that were the consumer's choice. I don't
want the consumer to be required to drain the stream before resuming,
as that'd be an unprincipled stall. For example, if new WAL arrives
over the network then I think it should be possible for recovery's
WAL-powered stream of heap pages to resume looking ahead even if
recovery hasn't drained the existing stream completely.

1. read_stream_resume() as before, but with a new explicit
read_stream_pause(): if a block number callback would like to report a
temporary lack of information, it should return
read_stream_pause(stream), not InvalidBlockNumber. Then after
read_stream_resume(stream) is called, the next
read_stream_next_buffer() enters the lookahead loop again. While
paused, if the consumer drains all the existing buffers in the stream
and then one more, it will receive InvalidBuffer, but if the _resume()
call is made sooner, the consumer won't ever know about the temporary
lack of buffers in the stream.

I like this new interface. If the user does want to exhaust the stream
(as was the case with earlier pgvector read stream user code), I
assume you would want to do:

read_stream_pause()
read_stream_reset()
read_stream_resume()

- Melanie

On Mon, Dec 8, 2025 at 10:47 PM Thomas Munro <thomas.munro@gmail.com> wrote:

Yielding just because you've scanned N index pages/tuples/whatever is
harder to think about. The stream shouldn't get far ahead unless it's
recently been useful for I/O concurrency (though optimal distance
heuristics are an open problem), but in this case a single invocation
of the block number callback can call ReadBuffer() an arbitrary number
of times, filtering out all the index tuples as it rampages through
the whole index IIUC. I see why you might want to yield periodically
if you can, but I also wonder how much that can really help if you
still have to pick up where you left off next time.

I think of it as a necessary precaution against pathological behavior
where the amount of memory used to cache matching tuples/TIDs gets out
of hand. There's no specific reason to expect that to happen (or no
good reason). But I'm pretty sure that it'll prove necessary to pay
non-zero attention to how much work has been done since the last time
we returned a tuple (when there's a tuple available to return).

I guess it
depends on the distribution of matches.

To be clear, I haven't done any kind of modelling of the problems in
this area. Once I do that (in 2026), I'll be able to say more about
the requirements. Maybe Tomas could take a look sooner?

Right now my focus is on getting the basic interfaces/API revisions in
better shape. And avoiding regressions while doing so.

--
Peter Geoghegan

On 12/9/25 23:38, Peter Geoghegan wrote:

On Mon, Dec 8, 2025 at 10:47 PM Thomas Munro <thomas.munro@gmail.com> wrote:

Yielding just because you've scanned N index pages/tuples/whatever is
harder to think about. The stream shouldn't get far ahead unless it's
recently been useful for I/O concurrency (though optimal distance
heuristics are an open problem), but in this case a single invocation
of the block number callback can call ReadBuffer() an arbitrary number
of times, filtering out all the index tuples as it rampages through
the whole index IIUC. I see why you might want to yield periodically
if you can, but I also wonder how much that can really help if you
still have to pick up where you left off next time.

I think of it as a necessary precaution against pathological behavior
where the amount of memory used to cache matching tuples/TIDs gets out
of hand. There's no specific reason to expect that to happen (or no
good reason). But I'm pretty sure that it'll prove necessary to pay
non-zero attention to how much work has been done since the last time
we returned a tuple (when there's a tuple available to return).

I think it's not all that difficult to hit such runaway cases, loading
arbitrary number of batches ahead. That's exactly why I had to come up
with the read_stream_reset approach in the first place, actually.

The simplest example is an index scan on a correlated index. We skip
duplicate blocks, so to find the next block to pass to the stream, we
may have to load multiple leaf pages. And we may need multiple such
blocks, to satisfy the prefetch distance (= number of IOs).

An even more extreme example is IOS, with just a couple not-allvisible
pages (that need prefetching). You hit the first one, then the distance
ramps up, and at some point there are no more not-allvisible pages. But
we try to maintain the distance, and we end up reading all remaining
leaf pages.

I'm sure we need to protect against these issues, which is why we have
INDEX_SCAN_MAX_BATCHES.

IIRC you also suggested we will need some internal limit to keep the
number of buffer pins under control, right?

I agree there may be other important reasons to "pause", e.g. based on
how much work was done since the last time the index scan returned a
tuple. But I'm not sure what exactly to look at, because most of the
"work" is happening outside the index scan, no?

I guess it
depends on the distribution of matches.

To be clear, I haven't done any kind of modelling of the problems in
this area. Once I do that (in 2026), I'll be able to say more about
the requirements. Maybe Tomas could take a look sooner?

I don't have any explicit formal model of the problems, but it should
not be difficult to come up with "adversary cases" for the prefetching
patches, for example. That is, ways to construct datasets / indexes that
end up looking very far (many leafs) ahead.

Is that what you meant by modelling, or did you mean some sort of a
formal model of how far to prefetch for a given dataset?

AFAICS the "ideal" prefetch distance would be such that a page gets
loaded into shared buffers right before the scan actually needs it
(after processing through the earlier index entries).

Let's say we know

* T1 - cost of "processing" one index tuple (average time between calls
to table_index_fetch_tuple?)

* T2 - cost of I/O, i.e. how long does it take to read a block

We want keep the look-ahead distance at least K, such that

K * T1 > T2

but not much more than that. But I'm not sure where to get T1 and T2, as
T1 depends on what happens outside the index scan, and T2 is hardware
dependent (and not actually linear / additive).

Or am I confused and you asked about something else?

regards

--
Tomas Vondra