Thread

Commits

Allow ReadStream to be consumed as raw block numbers.
- 70d38e3d8a2d 18.0 landed
- ec1d545c8f70 17.0 landed
Generalize relation analyze in table AM interface
- 27bc1772fc81 17.0 cited

Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-02-19T15:13:23Z

Hi,

I worked on using the currently proposed streaming read API [1] in ANALYZE.
The patch is attached. 0001 is the not yet merged streaming read API code
changes that can be applied to the master, 0002 is the actual code.

The blocks to analyze are obtained by using the streaming read API now.

- Since streaming read API is already doing prefetch, I removed the #ifdef
USE_PREFETCH code from acquire_sample_rows().

- Changed 'while (BlockSampler_HasMore(&bs))' to 'while (nblocks)' because
the prefetch mechanism in the streaming read API will advance 'bs' before
returning buffers.

- Removed BlockNumber and BufferAccessStrategy from the declaration of
scan_analyze_next_block(), passing pgsr (PgStreamingRead) instead of them.

I counted syscalls of analyzing ~5GB table. It can be seen that the patched
version did ~1300 less read calls.

Patched:

% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 39.67    0.012128           0     29809           pwrite64
 36.96    0.011299           0     28594           pread64
 23.24    0.007104           0     27611           fadvise64

Master (21a71648d3):

% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 38.94    0.016457           0     29816           pwrite64
 36.79    0.015549           0     29850           pread64
 23.91    0.010106           0     29848           fadvise64


Any kind of feedback would be appreciated.

[1]:
https://www.postgresql.org/message-id/CA%2BhUKGJkOiOCa%2Bmag4BF%2BzHo7qo%3Do9CFheB8%3Dg6uT5TUm2gkvA%40mail.gmail.com

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-02-28T11:42:34Z

Hi,

On Mon, 19 Feb 2024 at 18:13, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>
> I worked on using the currently proposed streaming read API [1] in ANALYZE. The patch is attached. 0001 is the not yet merged streaming read API code changes that can be applied to the master, 0002 is the actual code.
>
> The blocks to analyze are obtained by using the streaming read API now.
>
> - Since streaming read API is already doing prefetch, I removed the #ifdef USE_PREFETCH code from acquire_sample_rows().
>
> - Changed 'while (BlockSampler_HasMore(&bs))' to 'while (nblocks)' because the prefetch mechanism in the streaming read API will advance 'bs' before returning buffers.
>
> - Removed BlockNumber and BufferAccessStrategy from the declaration of scan_analyze_next_block(), passing pgsr (PgStreamingRead) instead of them.
>
> I counted syscalls of analyzing ~5GB table. It can be seen that the patched version did ~1300 less read calls.
>
> Patched:
>
> % time     seconds  usecs/call     calls    errors syscall
> ------ ----------- ----------- --------- --------- ----------------
>  39.67    0.012128           0     29809           pwrite64
>  36.96    0.011299           0     28594           pread64
>  23.24    0.007104           0     27611           fadvise64
>
> Master (21a71648d3):
>
> % time     seconds  usecs/call     calls    errors syscall
> ------ ----------- ----------- --------- --------- ----------------
>  38.94    0.016457           0     29816           pwrite64
>  36.79    0.015549           0     29850           pread64
>  23.91    0.010106           0     29848           fadvise64
>
>
> Any kind of feedback would be appreciated.
>
> [1]: https://www.postgresql.org/message-id/CA%2BhUKGJkOiOCa%2Bmag4BF%2BzHo7qo%3Do9CFheB8%3Dg6uT5TUm2gkvA%40mail.gmail.com

The new version of the streaming read API [1] is posted. I updated the
streaming read API changes patch (0001), using the streaming read API
in ANALYZE patch (0002) remains the same. This should make it easier
to review as it can be applied on top of master

[1]: https://www.postgresql.org/message-id/CA%2BhUKGJtLyxcAEvLhVUhgD4fMQkOu3PDaj8Qb9SR_UsmzgsBpQ%40mail.gmail.com

--
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-03-26T11:51:27Z

Hi,

On Wed, 28 Feb 2024 at 14:42, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>
>
> The new version of the streaming read API [1] is posted. I updated the
> streaming read API changes patch (0001), using the streaming read API
> in ANALYZE patch (0002) remains the same. This should make it easier
> to review as it can be applied on top of master
>
>

The new version of the streaming read API is posted [1]. I rebased the
patch on top of master and v9 of the streaming read API.

There is a minimal change in the 'using the streaming read API in ANALYZE
patch (0002)', I changed STREAMING_READ_FULL to STREAMING_READ_MAINTENANCE
to copy exactly the same behavior as before. Also, some benchmarking
results:

I created a 22 GB table and set the size of shared buffers to 30GB, the
rest is default.

╔═══════════════════════════╦═════════════════════╦════════════╗
║                           ║  Avg Timings in ms  ║            ║
╠═══════════════════════════╬══════════╦══════════╬════════════╣
║                           ║  master  ║  patched ║ percentage ║
╠═══════════════════════════╬══════════╬══════════╬════════════╣
║     Both OS cache and     ║          ║          ║            ║
║  shared buffers are clear ║ 513.9247 ║ 463.1019 ║    %9.9    ║
╠═══════════════════════════╬══════════╬══════════╬════════════╣
║   OS cache is loaded but  ║          ║          ║            ║
║  shared buffers are clear ║ 423.1097 ║ 354.3277 ║    %16.3   ║
╠═══════════════════════════╬══════════╬══════════╬════════════╣
║ Shared buffers are loaded ║          ║          ║            ║
║                           ║  89.2846 ║  84.6952 ║    %5.1    ║
╚═══════════════════════════╩══════════╩══════════╩════════════╝

Any kind of feedback would be appreciated.

[1]:
https://www.postgresql.org/message-id/CA%2BhUKGL-ONQnnnp-SONCFfLJzqcpAheuzZ%2B-yTrD9WBM-GmAcg%40mail.gmail.com

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Melanie Plageman <melanieplageman@gmail.com> — 2024-03-27T20:15:23Z

On Tue, Mar 26, 2024 at 02:51:27PM +0300, Nazir Bilal Yavuz wrote:
> Hi,
> 
> On Wed, 28 Feb 2024 at 14:42, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
> >
> >
> > The new version of the streaming read API [1] is posted. I updated the
> > streaming read API changes patch (0001), using the streaming read API
> > in ANALYZE patch (0002) remains the same. This should make it easier
> > to review as it can be applied on top of master
> >
> >
> 
> The new version of the streaming read API is posted [1]. I rebased the
> patch on top of master and v9 of the streaming read API.
> 
> There is a minimal change in the 'using the streaming read API in ANALYZE
> patch (0002)', I changed STREAMING_READ_FULL to STREAMING_READ_MAINTENANCE
> to copy exactly the same behavior as before. Also, some benchmarking
> results:
> 
> I created a 22 GB table and set the size of shared buffers to 30GB, the
> rest is default.
> 
> ╔═══════════════════════════╦═════════════════════╦════════════╗
> ║                           ║  Avg Timings in ms  ║            ║
> ╠═══════════════════════════╬══════════╦══════════╬════════════╣
> ║                           ║  master  ║  patched ║ percentage ║
> ╠═══════════════════════════╬══════════╬══════════╬════════════╣
> ║     Both OS cache and     ║          ║          ║            ║
> ║  shared buffers are clear ║ 513.9247 ║ 463.1019 ║    %9.9    ║
> ╠═══════════════════════════╬══════════╬══════════╬════════════╣
> ║   OS cache is loaded but  ║          ║          ║            ║
> ║  shared buffers are clear ║ 423.1097 ║ 354.3277 ║    %16.3   ║
> ╠═══════════════════════════╬══════════╬══════════╬════════════╣
> ║ Shared buffers are loaded ║          ║          ║            ║
> ║                           ║  89.2846 ║  84.6952 ║    %5.1    ║
> ╚═══════════════════════════╩══════════╩══════════╩════════════╝
> 
> Any kind of feedback would be appreciated.

Thanks for working on this!

A general review comment: I noticed you have the old streaming read
(pgsr) naming still in a few places (including comments) -- so I would
just make sure and update everywhere when you rebase in Thomas' latest
version of the read stream API.

> From c7500cc1b9068ff0b704181442999cd8bed58658 Mon Sep 17 00:00:00 2001
> From: Nazir Bilal Yavuz <byavuz81@gmail.com>
> Date: Mon, 19 Feb 2024 14:30:47 +0300
> Subject: [PATCH v3 2/2] Use streaming read API in ANALYZE
>
> --- a/src/backend/commands/analyze.c
> +++ b/src/backend/commands/analyze.c
> @@ -1102,6 +1102,26 @@ examine_attribute(Relation onerel, int attnum, Node *index_expr)
>  	return stats;
>  }
>  
> +/*
> + * Prefetch callback function to get next block number while using
> + * BlockSampling algorithm
> + */
> +static BlockNumber
> +pg_block_sampling_streaming_read_next(StreamingRead *stream,
> +									  void *user_data,
> +									  void *per_buffer_data)

I don't think you need the pg_ prefix

> +{
> +	BlockSamplerData *bs = user_data;
> +	BlockNumber *current_block = per_buffer_data;

Why can't you just do BufferGetBlockNumber() on the buffer returned from
the read stream API instead of allocating per_buffer_data for the block
number?

> +
> +	if (BlockSampler_HasMore(bs))
> +		*current_block = BlockSampler_Next(bs);
> +	else
> +		*current_block = InvalidBlockNumber;
> +
> +	return *current_block;


I think we'd like to keep the read stream code in heapam-specific code.
Instead of doing streaming_read_buffer_begin() here, you could put this
in heap_beginscan() or initscan() guarded by
	scan->rs_base.rs_flags & SO_TYPE_ANALYZE

same with streaming_read_buffer_end()/heap_endscan().

You'd also then need to save the reference to the read stream in the
HeapScanDescData.

> +	stream = streaming_read_buffer_begin(STREAMING_READ_MAINTENANCE,
> +										 vac_strategy,
> +										 BMR_REL(scan->rs_rd),
> +										 MAIN_FORKNUM,
> +										 pg_block_sampling_streaming_read_next,
> +										 &bs,
> +										 sizeof(BlockSamplerData));
>  
>  	/* Outer loop over blocks to sample */

In fact, I think you could use this opportunity to get rid of the block
dependency in acquire_sample_rows() altogether.

Looking at the code now, it seems like you could just invoke
heapam_scan_analyze_next_block() (maybe rename it to
heapam_scan_analyze_next_buffer() or something) from
heapam_scan_analyze_next_tuple() and remove
table_scan_analyze_next_block() entirely.

Then table AMs can figure out how they want to return tuples from
table_scan_analyze_next_tuple().

If you do all this, note that you'll need to update the comments above
acquire_sample_rows() accordingly.

> -	while (BlockSampler_HasMore(&bs))
> +	while (nblocks)
>  	{
>  		bool		block_accepted;
> -		BlockNumber targblock = BlockSampler_Next(&bs);
> -#ifdef USE_PREFETCH
> -		BlockNumber prefetch_targblock = InvalidBlockNumber;
> -
> -		/*
> -		 * Make sure that every time the main BlockSampler is moved forward
> -		 * that our prefetch BlockSampler also gets moved forward, so that we
> -		 * always stay out ahead.
> -		 */
> -		if (prefetch_maximum && BlockSampler_HasMore(&prefetch_bs))
> -			prefetch_targblock = BlockSampler_Next(&prefetch_bs);
> -#endif
>  
>  		vacuum_delay_point();
>  
> -		block_accepted = table_scan_analyze_next_block(scan, targblock, vac_strategy);
> +		block_accepted = table_scan_analyze_next_block(scan, stream);

- Melanie

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-04-02T07:23:55Z

Hi,

Thanks for the review!

On Wed, 27 Mar 2024 at 23:15, Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> On Tue, Mar 26, 2024 at 02:51:27PM +0300, Nazir Bilal Yavuz wrote:
> > Hi,
> >
> > On Wed, 28 Feb 2024 at 14:42, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
> > >
> > >
> > > The new version of the streaming read API [1] is posted. I updated the
> > > streaming read API changes patch (0001), using the streaming read API
> > > in ANALYZE patch (0002) remains the same. This should make it easier
> > > to review as it can be applied on top of master
> > >
> > >
> >
> > The new version of the streaming read API is posted [1]. I rebased the
> > patch on top of master and v9 of the streaming read API.
> >
> > There is a minimal change in the 'using the streaming read API in ANALYZE
> > patch (0002)', I changed STREAMING_READ_FULL to STREAMING_READ_MAINTENANCE
> > to copy exactly the same behavior as before. Also, some benchmarking
> > results:
> >
> > I created a 22 GB table and set the size of shared buffers to 30GB, the
> > rest is default.
> >
> > ╔═══════════════════════════╦═════════════════════╦════════════╗
> > ║                           ║  Avg Timings in ms  ║            ║
> > ╠═══════════════════════════╬══════════╦══════════╬════════════╣
> > ║                           ║  master  ║  patched ║ percentage ║
> > ╠═══════════════════════════╬══════════╬══════════╬════════════╣
> > ║     Both OS cache and     ║          ║          ║            ║
> > ║  shared buffers are clear ║ 513.9247 ║ 463.1019 ║    %9.9    ║
> > ╠═══════════════════════════╬══════════╬══════════╬════════════╣
> > ║   OS cache is loaded but  ║          ║          ║            ║
> > ║  shared buffers are clear ║ 423.1097 ║ 354.3277 ║    %16.3   ║
> > ╠═══════════════════════════╬══════════╬══════════╬════════════╣
> > ║ Shared buffers are loaded ║          ║          ║            ║
> > ║                           ║  89.2846 ║  84.6952 ║    %5.1    ║
> > ╚═══════════════════════════╩══════════╩══════════╩════════════╝
> >
> > Any kind of feedback would be appreciated.
>
> Thanks for working on this!
>
> A general review comment: I noticed you have the old streaming read
> (pgsr) naming still in a few places (including comments) -- so I would
> just make sure and update everywhere when you rebase in Thomas' latest
> version of the read stream API.

Done.

>
> > From c7500cc1b9068ff0b704181442999cd8bed58658 Mon Sep 17 00:00:00 2001
> > From: Nazir Bilal Yavuz <byavuz81@gmail.com>
> > Date: Mon, 19 Feb 2024 14:30:47 +0300
> > Subject: [PATCH v3 2/2] Use streaming read API in ANALYZE
> >
> > --- a/src/backend/commands/analyze.c
> > +++ b/src/backend/commands/analyze.c
> > @@ -1102,6 +1102,26 @@ examine_attribute(Relation onerel, int attnum, Node *index_expr)
> >       return stats;
> >  }
> >
> > +/*
> > + * Prefetch callback function to get next block number while using
> > + * BlockSampling algorithm
> > + */
> > +static BlockNumber
> > +pg_block_sampling_streaming_read_next(StreamingRead *stream,
> > +                                                                       void *user_data,
> > +                                                                       void *per_buffer_data)
>
> I don't think you need the pg_ prefix

Done.

>
> > +{
> > +     BlockSamplerData *bs = user_data;
> > +     BlockNumber *current_block = per_buffer_data;
>
> Why can't you just do BufferGetBlockNumber() on the buffer returned from
> the read stream API instead of allocating per_buffer_data for the block
> number?

Done.

>
> > +
> > +     if (BlockSampler_HasMore(bs))
> > +             *current_block = BlockSampler_Next(bs);
> > +     else
> > +             *current_block = InvalidBlockNumber;
> > +
> > +     return *current_block;
>
>
> I think we'd like to keep the read stream code in heapam-specific code.
> Instead of doing streaming_read_buffer_begin() here, you could put this
> in heap_beginscan() or initscan() guarded by
>         scan->rs_base.rs_flags & SO_TYPE_ANALYZE

In the recent changes [1], heapam_scan_analyze_next_[block | tuple]
are removed from tableam. They are directly called from
heapam-specific code now. So, IMO, no need to do this now.

v4 is rebased on top of v14 streaming read API changes.

[1] 27bc1772fc814946918a5ac8ccb9b5c5ad0380aa

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Jakub Wartak <jakub.wartak@enterprisedb.com> — 2024-04-03T08:41:42Z

On Tue, Apr 2, 2024 at 9:24 AM Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
[..]
> v4 is rebased on top of v14 streaming read API changes.

Hi Nazir, so with streaming API committed, I gave a try to this patch.
With autovacuum=off and 30GB table on NVMe (with standard readahead of
256kb and ext4, Debian 12, kernel 6.1.0, shared_buffers = 128MB
default) created using: create table t as select repeat('a', 100) || i
|| repeat('b', 500) as filler from generate_series(1, 45000000) as i;

on master, effect of mainteance_io_concurency [default 10] is like
that (when resetting the fs cache after each ANALYZE):

    m_io_c = 0:
    Time: 3137.914 ms (00:03.138)
    Time: 3094.540 ms (00:03.095)
    Time: 3452.513 ms (00:03.453)

    m_io_c = 1:
    Time: 2972.751 ms (00:02.973)
    Time: 2939.551 ms (00:02.940)
    Time: 2904.428 ms (00:02.904)

    m_io_c = 2:
    Time: 1580.260 ms (00:01.580)
    Time: 1572.132 ms (00:01.572)
    Time: 1558.334 ms (00:01.558)

    m_io_c = 4:
    Time: 938.304 ms
    Time: 931.772 ms
    Time: 920.044 ms

    m_io_c = 8:
    Time: 666.025 ms
    Time: 660.241 ms
    Time: 648.848 ms

    m_io_c = 16:
    Time: 542.450 ms
    Time: 561.155 ms
    Time: 539.683 ms

    m_io_c = 32:
    Time: 538.487 ms
    Time: 541.705 ms
    Time: 538.101 ms

with patch applied:

    m_io_c = 0:
    Time: 3106.469 ms (00:03.106)
    Time: 3140.343 ms (00:03.140)
    Time: 3044.133 ms (00:03.044)

    m_io_c = 1:
    Time: 2959.817 ms (00:02.960)
    Time: 2920.265 ms (00:02.920)
    Time: 2911.745 ms (00:02.912)

    m_io_c = 2:
    Time: 1581.912 ms (00:01.582)
    Time: 1561.444 ms (00:01.561)
    Time: 1558.251 ms (00:01.558)

    m_io_c = 4:
    Time: 908.116 ms
    Time: 901.245 ms
    Time: 901.071 ms

    m_io_c = 8:
    Time: 619.870 ms
    Time: 620.327 ms
    Time: 614.266 ms

    m_io_c = 16:
    Time: 529.885 ms
    Time: 526.958 ms
    Time: 528.474 ms

    m_io_c = 32:
    Time: 521.185 ms
    Time: 520.713 ms
    Time: 517.729 ms

No difference to me, which seems to be good. I've double checked and
patch used the new way

acquire_sample_rows -> heapam_scan_analyze_next_block ->
ReadBufferExtended -> ReadBuffer_common (inlined) -> WaitReadBuffers
-> mdreadv -> FileReadV -> pg_preadv (inlined)
acquire_sample_rows -> heapam_scan_analyze_next_block ->
ReadBufferExtended -> ReadBuffer_common (inlined) -> StartReadBuffer
-> ...

I gave also io_combine_limit to 32 (max, 256kB) a try and got those
slightly better results:

[..]
m_io_c = 16:
Time: 494.599 ms
Time: 496.345 ms
Time: 973.500 ms

m_io_c = 32:
Time: 461.031 ms
Time: 449.037 ms
Time: 443.375 ms

and that (last one) apparently was able to push it to ~50-60k still
random IOPS range, the rareq-sz was still ~8 (9.9) kB as analyze was
still reading random , so I assume no merging was done:

Device            r/s     rMB/s   rrqm/s  %rrqm r_await rareq-sz
w/s     wMB/s   wrqm/s  %wrqm w_await wareq-sz     d/s     dMB/s
drqm/s  %drqm d_await dareq-sz     f/s f_await  aqu-sz  %util
nvme0n1       61212.00    591.82     0.00   0.00    0.10     9.90
2.00      0.02     0.00   0.00    0.00    12.00    0.00      0.00
0.00   0.00    0.00     0.00    0.00    0.00    6.28  85.20

So in short it looks good to me.

-Jakub Wartak.

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-04-03T10:31:00Z

Hi,

On Tue, 2 Apr 2024 at 10:23, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>
> v4 is rebased on top of v14 streaming read API changes.

Streaming API has been committed but the committed version has a minor
change, the read_stream_begin_relation function takes Relation instead
of BufferManagerRelation now. So, here is a v5 which addresses this
change.

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Heikki Linnakangas <hlinnaka@iki.fi> — 2024-04-03T17:17:31Z

On 03/04/2024 13:31, Nazir Bilal Yavuz wrote:
> Streaming API has been committed but the committed version has a minor
> change, the read_stream_begin_relation function takes Relation instead
> of BufferManagerRelation now. So, here is a v5 which addresses this
> change.

I'm getting a repeatable segfault / assertion failure with this:

postgres=# CREATE TABLE tengiga (i int, filler text) with (fillfactor=10);
CREATE TABLE
postgres=# insert into tengiga select g, repeat('x', 900) from 
generate_series(1, 1400000) g;
INSERT 0 1400000
postgres=# set default_statistics_target = 10; ANALYZE tengiga;
SET
ANALYZE
postgres=# set default_statistics_target = 100; ANALYZE tengiga;
SET
ANALYZE
postgres=# set default_statistics_target =1000; ANALYZE tengiga;
SET
server closed the connection unexpectedly
	This probably means the server terminated abnormally
	before or while processing the request.

TRAP: failed Assert("BufferIsValid(hscan->rs_cbuf)"), File: 
"heapam_handler.c", Line: 1079, PID: 262232
postgres: heikki postgres [local] 
ANALYZE(ExceptionalCondition+0xa8)[0x56488a0de9d8]
postgres: heikki postgres [local] 
ANALYZE(heapam_scan_analyze_next_block+0x63)[0x5648899ece34]
postgres: heikki postgres [local] ANALYZE(+0x2d3f34)[0x564889b6af34]
postgres: heikki postgres [local] ANALYZE(+0x2d2a3a)[0x564889b69a3a]
postgres: heikki postgres [local] ANALYZE(analyze_rel+0x33e)[0x564889b68fa9]
postgres: heikki postgres [local] ANALYZE(vacuum+0x4b3)[0x564889c2dcc0]
postgres: heikki postgres [local] ANALYZE(ExecVacuum+0xd6f)[0x564889c2d7fe]
postgres: heikki postgres [local] 
ANALYZE(standard_ProcessUtility+0x901)[0x564889f0b8b9]
postgres: heikki postgres [local] 
ANALYZE(ProcessUtility+0x136)[0x564889f0afb1]
postgres: heikki postgres [local] ANALYZE(+0x6728c8)[0x564889f098c8]
postgres: heikki postgres [local] ANALYZE(+0x672b3b)[0x564889f09b3b]
postgres: heikki postgres [local] ANALYZE(PortalRun+0x320)[0x564889f09015]
postgres: heikki postgres [local] ANALYZE(+0x66b2c6)[0x564889f022c6]
postgres: heikki postgres [local] 
ANALYZE(PostgresMain+0x80c)[0x564889f06fd7]
postgres: heikki postgres [local] ANALYZE(+0x667876)[0x564889efe876]
postgres: heikki postgres [local] 
ANALYZE(postmaster_child_launch+0xe6)[0x564889e1f4b3]
postgres: heikki postgres [local] ANALYZE(+0x58e68e)[0x564889e2568e]
postgres: heikki postgres [local] ANALYZE(+0x58b7f0)[0x564889e227f0]
postgres: heikki postgres [local] 
ANALYZE(PostmasterMain+0x152b)[0x564889e2214d]
postgres: heikki postgres [local] ANALYZE(+0x4444b4)[0x564889cdb4b4]
/lib/x86_64-linux-gnu/libc.so.6(+0x2724a)[0x7f7d83b6724a]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x85)[0x7f7d83b67305]
postgres: heikki postgres [local] ANALYZE(_start+0x21)[0x564889971a61]
2024-04-03 20:15:49.157 EEST [262101] LOG:  server process (PID 262232) 
was terminated by signal 6: Aborted

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

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-04-03T18:59:32Z

Hi Jakub,

Thank you for looking into this and doing a performance analysis.

On Wed, 3 Apr 2024 at 11:42, Jakub Wartak <jakub.wartak@enterprisedb.com> wrote:
>
> On Tue, Apr 2, 2024 at 9:24 AM Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
> [..]
> > v4 is rebased on top of v14 streaming read API changes.
>
> Hi Nazir, so with streaming API committed, I gave a try to this patch.
> With autovacuum=off and 30GB table on NVMe (with standard readahead of
> 256kb and ext4, Debian 12, kernel 6.1.0, shared_buffers = 128MB
> default) created using: create table t as select repeat('a', 100) || i
> || repeat('b', 500) as filler from generate_series(1, 45000000) as i;
>
> on master, effect of mainteance_io_concurency [default 10] is like
> that (when resetting the fs cache after each ANALYZE):
>
>     m_io_c = 0:
>     Time: 3137.914 ms (00:03.138)
>     Time: 3094.540 ms (00:03.095)
>     Time: 3452.513 ms (00:03.453)
>
>     m_io_c = 1:
>     Time: 2972.751 ms (00:02.973)
>     Time: 2939.551 ms (00:02.940)
>     Time: 2904.428 ms (00:02.904)
>
>     m_io_c = 2:
>     Time: 1580.260 ms (00:01.580)
>     Time: 1572.132 ms (00:01.572)
>     Time: 1558.334 ms (00:01.558)
>
>     m_io_c = 4:
>     Time: 938.304 ms
>     Time: 931.772 ms
>     Time: 920.044 ms
>
>     m_io_c = 8:
>     Time: 666.025 ms
>     Time: 660.241 ms
>     Time: 648.848 ms
>
>     m_io_c = 16:
>     Time: 542.450 ms
>     Time: 561.155 ms
>     Time: 539.683 ms
>
>     m_io_c = 32:
>     Time: 538.487 ms
>     Time: 541.705 ms
>     Time: 538.101 ms
>
> with patch applied:
>
>     m_io_c = 0:
>     Time: 3106.469 ms (00:03.106)
>     Time: 3140.343 ms (00:03.140)
>     Time: 3044.133 ms (00:03.044)
>
>     m_io_c = 1:
>     Time: 2959.817 ms (00:02.960)
>     Time: 2920.265 ms (00:02.920)
>     Time: 2911.745 ms (00:02.912)
>
>     m_io_c = 2:
>     Time: 1581.912 ms (00:01.582)
>     Time: 1561.444 ms (00:01.561)
>     Time: 1558.251 ms (00:01.558)
>
>     m_io_c = 4:
>     Time: 908.116 ms
>     Time: 901.245 ms
>     Time: 901.071 ms
>
>     m_io_c = 8:
>     Time: 619.870 ms
>     Time: 620.327 ms
>     Time: 614.266 ms
>
>     m_io_c = 16:
>     Time: 529.885 ms
>     Time: 526.958 ms
>     Time: 528.474 ms
>
>     m_io_c = 32:
>     Time: 521.185 ms
>     Time: 520.713 ms
>     Time: 517.729 ms
>
> No difference to me, which seems to be good. I've double checked and
> patch used the new way
>
> acquire_sample_rows -> heapam_scan_analyze_next_block ->
> ReadBufferExtended -> ReadBuffer_common (inlined) -> WaitReadBuffers
> -> mdreadv -> FileReadV -> pg_preadv (inlined)
> acquire_sample_rows -> heapam_scan_analyze_next_block ->
> ReadBufferExtended -> ReadBuffer_common (inlined) -> StartReadBuffer
> -> ...
>
> I gave also io_combine_limit to 32 (max, 256kB) a try and got those
> slightly better results:
>
> [..]
> m_io_c = 16:
> Time: 494.599 ms
> Time: 496.345 ms
> Time: 973.500 ms
>
> m_io_c = 32:
> Time: 461.031 ms
> Time: 449.037 ms
> Time: 443.375 ms
>
> and that (last one) apparently was able to push it to ~50-60k still
> random IOPS range, the rareq-sz was still ~8 (9.9) kB as analyze was
> still reading random , so I assume no merging was done:
>
> Device            r/s     rMB/s   rrqm/s  %rrqm r_await rareq-sz
> w/s     wMB/s   wrqm/s  %wrqm w_await wareq-sz     d/s     dMB/s
> drqm/s  %drqm d_await dareq-sz     f/s f_await  aqu-sz  %util
> nvme0n1       61212.00    591.82     0.00   0.00    0.10     9.90
> 2.00      0.02     0.00   0.00    0.00    12.00    0.00      0.00
> 0.00   0.00    0.00     0.00    0.00    0.00    6.28  85.20
>
> So in short it looks good to me.

My results are similar to yours, also I realized a bug while working
on your benchmarking cases. I will share the cause and the fix soon.

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-04-03T19:25:01Z

Hi,

Thank you for looking into this!

On Wed, 3 Apr 2024 at 20:17, Heikki Linnakangas <hlinnaka@iki.fi> wrote:
>
> On 03/04/2024 13:31, Nazir Bilal Yavuz wrote:
> > Streaming API has been committed but the committed version has a minor
> > change, the read_stream_begin_relation function takes Relation instead
> > of BufferManagerRelation now. So, here is a v5 which addresses this
> > change.
>
> I'm getting a repeatable segfault / assertion failure with this:
>
> postgres=# CREATE TABLE tengiga (i int, filler text) with (fillfactor=10);
> CREATE TABLE
> postgres=# insert into tengiga select g, repeat('x', 900) from
> generate_series(1, 1400000) g;
> INSERT 0 1400000
> postgres=# set default_statistics_target = 10; ANALYZE tengiga;
> SET
> ANALYZE
> postgres=# set default_statistics_target = 100; ANALYZE tengiga;
> SET
> ANALYZE
> postgres=# set default_statistics_target =1000; ANALYZE tengiga;
> SET
> server closed the connection unexpectedly
>         This probably means the server terminated abnormally
>         before or while processing the request.
>
> TRAP: failed Assert("BufferIsValid(hscan->rs_cbuf)"), File:
> "heapam_handler.c", Line: 1079, PID: 262232
> postgres: heikki postgres [local]
> ANALYZE(ExceptionalCondition+0xa8)[0x56488a0de9d8]
> postgres: heikki postgres [local]
> ANALYZE(heapam_scan_analyze_next_block+0x63)[0x5648899ece34]
> postgres: heikki postgres [local] ANALYZE(+0x2d3f34)[0x564889b6af34]
> postgres: heikki postgres [local] ANALYZE(+0x2d2a3a)[0x564889b69a3a]
> postgres: heikki postgres [local] ANALYZE(analyze_rel+0x33e)[0x564889b68fa9]
> postgres: heikki postgres [local] ANALYZE(vacuum+0x4b3)[0x564889c2dcc0]
> postgres: heikki postgres [local] ANALYZE(ExecVacuum+0xd6f)[0x564889c2d7fe]
> postgres: heikki postgres [local]
> ANALYZE(standard_ProcessUtility+0x901)[0x564889f0b8b9]
> postgres: heikki postgres [local]
> ANALYZE(ProcessUtility+0x136)[0x564889f0afb1]
> postgres: heikki postgres [local] ANALYZE(+0x6728c8)[0x564889f098c8]
> postgres: heikki postgres [local] ANALYZE(+0x672b3b)[0x564889f09b3b]
> postgres: heikki postgres [local] ANALYZE(PortalRun+0x320)[0x564889f09015]
> postgres: heikki postgres [local] ANALYZE(+0x66b2c6)[0x564889f022c6]
> postgres: heikki postgres [local]
> ANALYZE(PostgresMain+0x80c)[0x564889f06fd7]
> postgres: heikki postgres [local] ANALYZE(+0x667876)[0x564889efe876]
> postgres: heikki postgres [local]
> ANALYZE(postmaster_child_launch+0xe6)[0x564889e1f4b3]
> postgres: heikki postgres [local] ANALYZE(+0x58e68e)[0x564889e2568e]
> postgres: heikki postgres [local] ANALYZE(+0x58b7f0)[0x564889e227f0]
> postgres: heikki postgres [local]
> ANALYZE(PostmasterMain+0x152b)[0x564889e2214d]
> postgres: heikki postgres [local] ANALYZE(+0x4444b4)[0x564889cdb4b4]
> /lib/x86_64-linux-gnu/libc.so.6(+0x2724a)[0x7f7d83b6724a]
> /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x85)[0x7f7d83b67305]
> postgres: heikki postgres [local] ANALYZE(_start+0x21)[0x564889971a61]
> 2024-04-03 20:15:49.157 EEST [262101] LOG:  server process (PID 262232)
> was terminated by signal 6: Aborted

I realized the same error while working on Jakub's benchmarking results.

Cause: I was using the nblocks variable to check how many blocks will
be returned from the streaming API. But I realized that sometimes the
number returned from BlockSampler_Init() is not equal to the number of
blocks that BlockSampler_Next() will return as BlockSampling algorithm
decides how many blocks to return on the fly by using some random
seeds.

There are a couple of solutions I thought of:

1- Use BlockSampler_HasMore() instead of nblocks in the main loop in
the acquire_sample_rows():

Streaming API uses this function to prefetch block numbers.
BlockSampler_HasMore() will reach to the end first as it is used while
prefetching, so it will start to return false while there are still
buffers to return from the streaming API. That will cause some buffers
at the end to not be processed.

2- Expose something (function, variable etc.) from the streaming API
to understand if the read is finished and there is no buffer to
return:

I think this works but I am not sure if the streaming API allows
something like that.

3- Check every buffer returned from the streaming API, if it is
invalid stop the main loop in the acquire_sample_rows():

This solves the problem but there will be two if checks for each
buffer returned,
- in heapam_scan_analyze_next_block() to check if the returned buffer is invalid
- to break main loop in acquire_sample_rows() if
heapam_scan_analyze_next_block() returns false
One of the if cases can be bypassed by moving
heapam_scan_analyze_next_block()'s code to the main loop in the
acquire_sample_rows().

I implemented the third solution, here is v6.

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Melanie Plageman <melanieplageman@gmail.com> — 2024-04-03T20:44:20Z

On Wed, Apr 03, 2024 at 10:25:01PM +0300, Nazir Bilal Yavuz wrote:
>
> I realized the same error while working on Jakub's benchmarking results.
> 
> Cause: I was using the nblocks variable to check how many blocks will
> be returned from the streaming API. But I realized that sometimes the
> number returned from BlockSampler_Init() is not equal to the number of
> blocks that BlockSampler_Next() will return as BlockSampling algorithm
> decides how many blocks to return on the fly by using some random
> seeds.
> 
> There are a couple of solutions I thought of:
> 
> 1- Use BlockSampler_HasMore() instead of nblocks in the main loop in
> the acquire_sample_rows():
> 
> Streaming API uses this function to prefetch block numbers.
> BlockSampler_HasMore() will reach to the end first as it is used while
> prefetching, so it will start to return false while there are still
> buffers to return from the streaming API. That will cause some buffers
> at the end to not be processed.
> 
> 2- Expose something (function, variable etc.) from the streaming API
> to understand if the read is finished and there is no buffer to
> return:
> 
> I think this works but I am not sure if the streaming API allows
> something like that.
> 
> 3- Check every buffer returned from the streaming API, if it is
> invalid stop the main loop in the acquire_sample_rows():
> 
> This solves the problem but there will be two if checks for each
> buffer returned,
> - in heapam_scan_analyze_next_block() to check if the returned buffer is invalid
> - to break main loop in acquire_sample_rows() if
> heapam_scan_analyze_next_block() returns false
> One of the if cases can be bypassed by moving
> heapam_scan_analyze_next_block()'s code to the main loop in the
> acquire_sample_rows().
> 
> I implemented the third solution, here is v6.

I've reviewed the patches inline below and attached a patch that has
some of my ideas on top of your patch.

> From 8d396a42186325f920d5a05e7092d8e1b66f3cdf Mon Sep 17 00:00:00 2001
> From: Nazir Bilal Yavuz <byavuz81@gmail.com>
> Date: Wed, 3 Apr 2024 15:14:15 +0300
> Subject: [PATCH v6] Use streaming read API in ANALYZE
> 
> ANALYZE command gets random tuples using BlockSampler algorithm. Use
> streaming reads to get these tuples by using BlockSampler algorithm in
> streaming read API prefetch logic.
> ---
>  src/include/access/heapam.h              |  6 +-
>  src/backend/access/heap/heapam_handler.c | 22 +++---
>  src/backend/commands/analyze.c           | 85 ++++++++----------------
>  3 files changed, 42 insertions(+), 71 deletions(-)
> 
> diff --git a/src/include/access/heapam.h b/src/include/access/heapam.h
> index a307fb5f245..633caee9d95 100644
> --- a/src/include/access/heapam.h
> +++ b/src/include/access/heapam.h
> @@ -25,6 +25,7 @@
>  #include "storage/bufpage.h"
>  #include "storage/dsm.h"
>  #include "storage/lockdefs.h"
> +#include "storage/read_stream.h"
>  #include "storage/shm_toc.h"
>  #include "utils/relcache.h"
>  #include "utils/snapshot.h"
> @@ -388,9 +389,8 @@ extern bool HeapTupleIsSurelyDead(HeapTuple htup,
>  								  struct GlobalVisState *vistest);
>  
>  /* in heap/heapam_handler.c*/
> -extern void heapam_scan_analyze_next_block(TableScanDesc scan,
> -										   BlockNumber blockno,
> -										   BufferAccessStrategy bstrategy);
> +extern bool heapam_scan_analyze_next_block(TableScanDesc scan,
> +										   ReadStream *stream);
>  extern bool heapam_scan_analyze_next_tuple(TableScanDesc scan,
>  										   TransactionId OldestXmin,
>  										   double *liverows, double *deadrows,
> diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
> index 0952d4a98eb..d83fbbe6af3 100644
> --- a/src/backend/access/heap/heapam_handler.c
> +++ b/src/backend/access/heap/heapam_handler.c
> @@ -1054,16 +1054,16 @@ heapam_relation_copy_for_cluster(Relation OldHeap, Relation NewHeap,
>  }
>  
>  /*
> - * Prepare to analyze block `blockno` of `scan`.  The scan has been started
> - * with SO_TYPE_ANALYZE option.
> + * Prepare to analyze block returned from streaming object.  If the block returned
> + * from streaming object is valid, true is returned; otherwise false is returned.
> + * The scan has been started with SO_TYPE_ANALYZE option.
>   *
>   * This routine holds a buffer pin and lock on the heap page.  They are held
>   * until heapam_scan_analyze_next_tuple() returns false.  That is until all the
>   * items of the heap page are analyzed.
>   */
> -void
> -heapam_scan_analyze_next_block(TableScanDesc scan, BlockNumber blockno,
> -							   BufferAccessStrategy bstrategy)
> +bool
> +heapam_scan_analyze_next_block(TableScanDesc scan, ReadStream *stream)
>  {
>  	HeapScanDesc hscan = (HeapScanDesc) scan;
>  
> @@ -1076,11 +1076,15 @@ heapam_scan_analyze_next_block(TableScanDesc scan, BlockNumber blockno,
>  	 * doing much work per tuple, the extra lock traffic is probably better
>  	 * avoided.

Personally I think heapam_scan_analyze_next_block() should be inlined.
It only has a few lines. I would find it clearer inline. At the least,
there is no reason for it (or heapam_scan_analyze_next_tuple()) to take
a TableScanDesc instead of a HeapScanDesc.

>  	 */
> -	hscan->rs_cblock = blockno;
> -	hscan->rs_cindex = FirstOffsetNumber;
> -	hscan->rs_cbuf = ReadBufferExtended(scan->rs_rd, MAIN_FORKNUM,
> -										blockno, RBM_NORMAL, bstrategy);
> +	hscan->rs_cbuf = read_stream_next_buffer(stream, NULL);
> +	if (hscan->rs_cbuf == InvalidBuffer)
> +		return false;
> +
>  	LockBuffer(hscan->rs_cbuf, BUFFER_LOCK_SHARE);
> +
> +	hscan->rs_cblock = BufferGetBlockNumber(hscan->rs_cbuf);
> +	hscan->rs_cindex = FirstOffsetNumber;
> +	return true;
>  }

>  /*
> diff --git a/src/backend/commands/analyze.c b/src/backend/commands/analyze.c
> index 2fb39f3ede1..764520d5aa2 100644
> --- a/src/backend/commands/analyze.c
> +++ b/src/backend/commands/analyze.c
> @@ -1102,6 +1102,20 @@ examine_attribute(Relation onerel, int attnum, Node *index_expr)
>  	return stats;
>  }
>  
> +/*
> + * Prefetch callback function to get next block number while using
> + * BlockSampling algorithm
> + */
> +static BlockNumber
> +block_sampling_streaming_read_next(ReadStream *stream,
> +								   void *user_data,
> +								   void *per_buffer_data)
> +{
> +	BlockSamplerData *bs = user_data;
> +
> +	return BlockSampler_HasMore(bs) ? BlockSampler_Next(bs) : InvalidBlockNumber;

I don't see the point of BlockSampler_HasMore() anymore. I removed it in
the attached and made BlockSampler_Next() return InvalidBlockNumber
under the same conditions. Is there a reason not to do this? There
aren't other callers. If the BlockSampler_Next() wasn't part of an API,
we could just make it the streaming read callback, but that might be
weird as it is now.

That and my other ideas in attached. Let me know what you think.

- Melanie

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-04-04T11:03:30Z

Hi,

On Wed, 3 Apr 2024 at 23:44, Melanie Plageman <melanieplageman@gmail.com> wrote:
>
>
> I've reviewed the patches inline below and attached a patch that has
> some of my ideas on top of your patch.

Thank you!

>
> > From 8d396a42186325f920d5a05e7092d8e1b66f3cdf Mon Sep 17 00:00:00 2001
> > From: Nazir Bilal Yavuz <byavuz81@gmail.com>
> > Date: Wed, 3 Apr 2024 15:14:15 +0300
> > Subject: [PATCH v6] Use streaming read API in ANALYZE
> >
> > ANALYZE command gets random tuples using BlockSampler algorithm. Use
> > streaming reads to get these tuples by using BlockSampler algorithm in
> > streaming read API prefetch logic.
> > ---
> >  src/include/access/heapam.h              |  6 +-
> >  src/backend/access/heap/heapam_handler.c | 22 +++---
> >  src/backend/commands/analyze.c           | 85 ++++++++----------------
> >  3 files changed, 42 insertions(+), 71 deletions(-)
> >
> > diff --git a/src/include/access/heapam.h b/src/include/access/heapam.h
> > index a307fb5f245..633caee9d95 100644
> > --- a/src/include/access/heapam.h
> > +++ b/src/include/access/heapam.h
> > @@ -25,6 +25,7 @@
> >  #include "storage/bufpage.h"
> >  #include "storage/dsm.h"
> >  #include "storage/lockdefs.h"
> > +#include "storage/read_stream.h"
> >  #include "storage/shm_toc.h"
> >  #include "utils/relcache.h"
> >  #include "utils/snapshot.h"
> > @@ -388,9 +389,8 @@ extern bool HeapTupleIsSurelyDead(HeapTuple htup,
> >                                                                 struct GlobalVisState *vistest);
> >
> >  /* in heap/heapam_handler.c*/
> > -extern void heapam_scan_analyze_next_block(TableScanDesc scan,
> > -                                                                                BlockNumber blockno,
> > -                                                                                BufferAccessStrategy bstrategy);
> > +extern bool heapam_scan_analyze_next_block(TableScanDesc scan,
> > +                                                                                ReadStream *stream);
> >  extern bool heapam_scan_analyze_next_tuple(TableScanDesc scan,
> >                                                                                  TransactionId OldestXmin,
> >                                                                                  double *liverows, double *deadrows,
> > diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
> > index 0952d4a98eb..d83fbbe6af3 100644
> > --- a/src/backend/access/heap/heapam_handler.c
> > +++ b/src/backend/access/heap/heapam_handler.c
> > @@ -1054,16 +1054,16 @@ heapam_relation_copy_for_cluster(Relation OldHeap, Relation NewHeap,
> >  }
> >
> >  /*
> > - * Prepare to analyze block `blockno` of `scan`.  The scan has been started
> > - * with SO_TYPE_ANALYZE option.
> > + * Prepare to analyze block returned from streaming object.  If the block returned
> > + * from streaming object is valid, true is returned; otherwise false is returned.
> > + * The scan has been started with SO_TYPE_ANALYZE option.
> >   *
> >   * This routine holds a buffer pin and lock on the heap page.  They are held
> >   * until heapam_scan_analyze_next_tuple() returns false.  That is until all the
> >   * items of the heap page are analyzed.
> >   */
> > -void
> > -heapam_scan_analyze_next_block(TableScanDesc scan, BlockNumber blockno,
> > -                                                        BufferAccessStrategy bstrategy)
> > +bool
> > +heapam_scan_analyze_next_block(TableScanDesc scan, ReadStream *stream)
> >  {
> >       HeapScanDesc hscan = (HeapScanDesc) scan;
> >
> > @@ -1076,11 +1076,15 @@ heapam_scan_analyze_next_block(TableScanDesc scan, BlockNumber blockno,
> >        * doing much work per tuple, the extra lock traffic is probably better
> >        * avoided.
>
> Personally I think heapam_scan_analyze_next_block() should be inlined.
> It only has a few lines. I would find it clearer inline. At the least,
> there is no reason for it (or heapam_scan_analyze_next_tuple()) to take
> a TableScanDesc instead of a HeapScanDesc.

I agree.

>
> >        */
> > -     hscan->rs_cblock = blockno;
> > -     hscan->rs_cindex = FirstOffsetNumber;
> > -     hscan->rs_cbuf = ReadBufferExtended(scan->rs_rd, MAIN_FORKNUM,
> > -                                                                             blockno, RBM_NORMAL, bstrategy);
> > +     hscan->rs_cbuf = read_stream_next_buffer(stream, NULL);
> > +     if (hscan->rs_cbuf == InvalidBuffer)
> > +             return false;
> > +
> >       LockBuffer(hscan->rs_cbuf, BUFFER_LOCK_SHARE);
> > +
> > +     hscan->rs_cblock = BufferGetBlockNumber(hscan->rs_cbuf);
> > +     hscan->rs_cindex = FirstOffsetNumber;
> > +     return true;
> >  }
>
> >  /*
> > diff --git a/src/backend/commands/analyze.c b/src/backend/commands/analyze.c
> > index 2fb39f3ede1..764520d5aa2 100644
> > --- a/src/backend/commands/analyze.c
> > +++ b/src/backend/commands/analyze.c
> > @@ -1102,6 +1102,20 @@ examine_attribute(Relation onerel, int attnum, Node *index_expr)
> >       return stats;
> >  }
> >
> > +/*
> > + * Prefetch callback function to get next block number while using
> > + * BlockSampling algorithm
> > + */
> > +static BlockNumber
> > +block_sampling_streaming_read_next(ReadStream *stream,
> > +                                                                void *user_data,
> > +                                                                void *per_buffer_data)
> > +{
> > +     BlockSamplerData *bs = user_data;
> > +
> > +     return BlockSampler_HasMore(bs) ? BlockSampler_Next(bs) : InvalidBlockNumber;
>
> I don't see the point of BlockSampler_HasMore() anymore. I removed it in
> the attached and made BlockSampler_Next() return InvalidBlockNumber
> under the same conditions. Is there a reason not to do this? There
> aren't other callers. If the BlockSampler_Next() wasn't part of an API,
> we could just make it the streaming read callback, but that might be
> weird as it is now.

I agree. There is no reason to have BlockSampler_HasMore() after
streaming read API changes.

> That and my other ideas in attached. Let me know what you think.

I agree with your changes but I am not sure if others agree with all
the changes you have proposed. So, I didn't merge 0001 and your ideas
yet, instead I wrote a commit message, added some comments, changed ->
'if (bs->t >= bs->N || bs->m >= bs->n)' to 'if (K <= 0 || k <= 0)' and
attached it as 0002.

--
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Melanie Plageman <melanieplageman@gmail.com> — 2024-04-07T19:57:08Z

On Thu, Apr 04, 2024 at 02:03:30PM +0300, Nazir Bilal Yavuz wrote:
> 
> On Wed, 3 Apr 2024 at 23:44, Melanie Plageman <melanieplageman@gmail.com> wrote:
> >
> > I don't see the point of BlockSampler_HasMore() anymore. I removed it in
> > the attached and made BlockSampler_Next() return InvalidBlockNumber
> > under the same conditions. Is there a reason not to do this? There
> > aren't other callers. If the BlockSampler_Next() wasn't part of an API,
> > we could just make it the streaming read callback, but that might be
> > weird as it is now.
> 
> I agree. There is no reason to have BlockSampler_HasMore() after
> streaming read API changes.
> 
> > That and my other ideas in attached. Let me know what you think.
> 
> I agree with your changes but I am not sure if others agree with all
> the changes you have proposed. So, I didn't merge 0001 and your ideas
> yet, instead I wrote a commit message, added some comments, changed ->
> 'if (bs->t >= bs->N || bs->m >= bs->n)' to 'if (K <= 0 || k <= 0)' and
> attached it as 0002.

I couldn't quite let go of those changes to acquire_sample_rows(), so
attached v9 0001 implements them as a preliminary patch before your
analyze streaming read user. I inlined heapam_scan_analyze_next_block()
entirely and made heapam_scan_analyze_next_tuple() a static function in
commands/analyze.c (and tweaked the name).

I made a few tweaks to your patch since it is on top of those changes
instead of preceding them. Then 0003 is removing BlockSampler_HasMore()
since it doesn't make sense to remove it before the streaming read user
was added.

- Melanie

Re: Use streaming read API in ANALYZE

Melanie Plageman <melanieplageman@gmail.com> — 2024-04-07T20:59:26Z

On Sun, Apr 7, 2024 at 3:57 PM Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> On Thu, Apr 04, 2024 at 02:03:30PM +0300, Nazir Bilal Yavuz wrote:
> >
> > On Wed, 3 Apr 2024 at 23:44, Melanie Plageman <melanieplageman@gmail.com> wrote:
> > >
> > > I don't see the point of BlockSampler_HasMore() anymore. I removed it in
> > > the attached and made BlockSampler_Next() return InvalidBlockNumber
> > > under the same conditions. Is there a reason not to do this? There
> > > aren't other callers. If the BlockSampler_Next() wasn't part of an API,
> > > we could just make it the streaming read callback, but that might be
> > > weird as it is now.
> >
> > I agree. There is no reason to have BlockSampler_HasMore() after
> > streaming read API changes.
> >
> > > That and my other ideas in attached. Let me know what you think.
> >
> > I agree with your changes but I am not sure if others agree with all
> > the changes you have proposed. So, I didn't merge 0001 and your ideas
> > yet, instead I wrote a commit message, added some comments, changed ->
> > 'if (bs->t >= bs->N || bs->m >= bs->n)' to 'if (K <= 0 || k <= 0)' and
> > attached it as 0002.
>
> I couldn't quite let go of those changes to acquire_sample_rows(), so
> attached v9 0001 implements them as a preliminary patch before your
> analyze streaming read user. I inlined heapam_scan_analyze_next_block()
> entirely and made heapam_scan_analyze_next_tuple() a static function in
> commands/analyze.c (and tweaked the name).
>
> I made a few tweaks to your patch since it is on top of those changes
> instead of preceding them. Then 0003 is removing BlockSampler_HasMore()
> since it doesn't make sense to remove it before the streaming read user
> was added.

I realized there were a few outdated comments. Fixed in attached v10.

- Melanie

Re: Use streaming read API in ANALYZE

Andres Freund <andres@anarazel.de> — 2024-04-07T22:00:00Z

Hi,

On 2024-04-07 16:59:26 -0400, Melanie Plageman wrote:
> From 1dc2343661f3edb3b1bc4307afb0e956397eb76c Mon Sep 17 00:00:00 2001
> From: Melanie Plageman <melanieplageman@gmail.com>
> Date: Sun, 7 Apr 2024 14:55:22 -0400
> Subject: [PATCH v10 1/3] Make heapam_scan_analyze_next_[tuple|block] static.
> 
> 27bc1772fc81 removed the table AM callbacks scan_analyze_next_block and
> scan_analzye_next_tuple -- leaving their heap AM implementations only
> called by acquire_sample_rows().

Ugh, I don't think 27bc1772fc81 makes much sense. But that's unrelated to this
thread.  I did raise that separately
https://www.postgresql.org/message-id/20240407214001.jgpg5q3yv33ve6y3%40awork3.anarazel.de

Unless I seriously missed something, I see no alternative to reverting that
commit.


> @@ -1206,11 +1357,13 @@ acquire_sample_rows(Relation onerel, int elevel,
>  				break;
>  
>  			prefetch_block = BlockSampler_Next(&prefetch_bs);
> -			PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_block);
> +			PrefetchBuffer(scan->rs_base.rs_rd, MAIN_FORKNUM, prefetch_block);
>  		}
>  	}
>  #endif
>  
> +	scan->rs_cbuf = InvalidBuffer;
> +
>  	/* Outer loop over blocks to sample */
>  	while (BlockSampler_HasMore(&bs))
>  	{

I don't think it's good to move a lot of code *and* change how it is
structured in the same commit. Makes it much harder to actually see changes /
makes git blame harder to use / etc.



> From 90d115c2401567be65bcf64393a6d3b39286779e Mon Sep 17 00:00:00 2001
> From: Melanie Plageman <melanieplageman@gmail.com>
> Date: Sun, 7 Apr 2024 15:28:32 -0400
> Subject: [PATCH v10 2/3] Use streaming read API in ANALYZE
>
> The ANALYZE command prefetches and reads sample blocks chosen by a
> BlockSampler algorithm. Instead of calling Prefetch|ReadBuffer() for
> each block, ANALYZE now uses the streaming API introduced in b5a9b18cd0.
>
> Author: Nazir Bilal Yavuz
> Reviewed-by: Melanie Plageman
> Discussion: https://postgr.es/m/flat/CAN55FZ0UhXqk9v3y-zW_fp4-WCp43V8y0A72xPmLkOM%2B6M%2BmJg%40mail.gmail.com
> ---
>  src/backend/commands/analyze.c | 89 ++++++++++------------------------
>  1 file changed, 26 insertions(+), 63 deletions(-)

That's a very nice demonstration of how this makes good prefetching easier...




> From 862b7ac81cdafcda7b525e02721da14e46265509 Mon Sep 17 00:00:00 2001
> From: Melanie Plageman <melanieplageman@gmail.com>
> Date: Sun, 7 Apr 2024 15:38:41 -0400
> Subject: [PATCH v10 3/3] Obsolete BlockSampler_HasMore()
> 
> A previous commit stopped using BlockSampler_HasMore() for flow control
> in acquire_sample_rows(). There seems little use now for
> BlockSampler_HasMore(). It should be sufficient to return
> InvalidBlockNumber from BlockSampler_Next() when BlockSample_HasMore()
> would have returned false. Remove BlockSampler_HasMore().
> 
> Author: Melanie Plageman, Nazir Bilal Yavuz
> Discussion: https://postgr.es/m/flat/CAN55FZ0UhXqk9v3y-zW_fp4-WCp43V8y0A72xPmLkOM%2B6M%2BmJg%40mail.gmail.com

The justification here seems somewhat odd. Sure, the previous commit stopped
using BlockSampler_HasMore in acquire_sample_rows - but only because it was
moved to block_sampling_streaming_read_next()?

Greetings,

Andres Freund

Re: Use streaming read API in ANALYZE

Melanie Plageman <melanieplageman@gmail.com> — 2024-04-07T22:26:31Z

On Sun, Apr 07, 2024 at 03:00:00PM -0700, Andres Freund wrote:
> Hi,
> 
> On 2024-04-07 16:59:26 -0400, Melanie Plageman wrote:
> > From 1dc2343661f3edb3b1bc4307afb0e956397eb76c Mon Sep 17 00:00:00 2001
> > From: Melanie Plageman <melanieplageman@gmail.com>
> > Date: Sun, 7 Apr 2024 14:55:22 -0400
> > Subject: [PATCH v10 1/3] Make heapam_scan_analyze_next_[tuple|block] static.
> > 
> > 27bc1772fc81 removed the table AM callbacks scan_analyze_next_block and
> > scan_analzye_next_tuple -- leaving their heap AM implementations only
> > called by acquire_sample_rows().
> 
> Ugh, I don't think 27bc1772fc81 makes much sense. But that's unrelated to this
> thread.  I did raise that separately
> https://www.postgresql.org/message-id/20240407214001.jgpg5q3yv33ve6y3%40awork3.anarazel.de
> 
> Unless I seriously missed something, I see no alternative to reverting that
> commit.

Noted. I'll give up on this refactor then. Lots of churn for no gain.
Attached v11 is just Bilal's v8 patch rebased to apply cleanly and with
a few tweaks (I changed one of the loop conditions. All other changes
are to comments and commit message).

> > @@ -1206,11 +1357,13 @@ acquire_sample_rows(Relation onerel, int elevel,
> >  				break;
> >  
> >  			prefetch_block = BlockSampler_Next(&prefetch_bs);
> > -			PrefetchBuffer(scan->rs_rd, MAIN_FORKNUM, prefetch_block);
> > +			PrefetchBuffer(scan->rs_base.rs_rd, MAIN_FORKNUM, prefetch_block);
> >  		}
> >  	}
> >  #endif
> >  
> > +	scan->rs_cbuf = InvalidBuffer;
> > +
> >  	/* Outer loop over blocks to sample */
> >  	while (BlockSampler_HasMore(&bs))
> >  	{
> 
> I don't think it's good to move a lot of code *and* change how it is
> structured in the same commit. Makes it much harder to actually see changes /
> makes git blame harder to use / etc.

Yep.

> > From 90d115c2401567be65bcf64393a6d3b39286779e Mon Sep 17 00:00:00 2001
> > From: Melanie Plageman <melanieplageman@gmail.com>
> > Date: Sun, 7 Apr 2024 15:28:32 -0400
> > Subject: [PATCH v10 2/3] Use streaming read API in ANALYZE
> >
> > The ANALYZE command prefetches and reads sample blocks chosen by a
> > BlockSampler algorithm. Instead of calling Prefetch|ReadBuffer() for
> > each block, ANALYZE now uses the streaming API introduced in b5a9b18cd0.
> >
> > Author: Nazir Bilal Yavuz
> > Reviewed-by: Melanie Plageman
> > Discussion: https://postgr.es/m/flat/CAN55FZ0UhXqk9v3y-zW_fp4-WCp43V8y0A72xPmLkOM%2B6M%2BmJg%40mail.gmail.com
> > ---
> >  src/backend/commands/analyze.c | 89 ++++++++++------------------------
> >  1 file changed, 26 insertions(+), 63 deletions(-)
> 
> That's a very nice demonstration of how this makes good prefetching easier...

Agreed. Yay streaming read API and Bilal!

> > From 862b7ac81cdafcda7b525e02721da14e46265509 Mon Sep 17 00:00:00 2001
> > From: Melanie Plageman <melanieplageman@gmail.com>
> > Date: Sun, 7 Apr 2024 15:38:41 -0400
> > Subject: [PATCH v10 3/3] Obsolete BlockSampler_HasMore()
> > 
> > A previous commit stopped using BlockSampler_HasMore() for flow control
> > in acquire_sample_rows(). There seems little use now for
> > BlockSampler_HasMore(). It should be sufficient to return
> > InvalidBlockNumber from BlockSampler_Next() when BlockSample_HasMore()
> > would have returned false. Remove BlockSampler_HasMore().
> > 
> > Author: Melanie Plageman, Nazir Bilal Yavuz
> > Discussion: https://postgr.es/m/flat/CAN55FZ0UhXqk9v3y-zW_fp4-WCp43V8y0A72xPmLkOM%2B6M%2BmJg%40mail.gmail.com
> 
> The justification here seems somewhat odd. Sure, the previous commit stopped
> using BlockSampler_HasMore in acquire_sample_rows - but only because it was
> moved to block_sampling_streaming_read_next()?

It didn't stop using it. It stopped being useful. The reason it existed,
as far as I can tell, was to use it as the while() loop condition in
acquire_sample_rows(). I think it makes much more sense for
BlockSampler_Next() to return InvalidBlockNumber when there are no more
blocks -- not to assert you don't call it when there aren't any more
blocks.

I didn't want to change BlockSampler_Next() in the same commit as the
streaming read user and we can't remove BlockSampler_HasMore() without
changing BlockSampler_Next().

- Melanie

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-04-08T01:20:21Z

On Mon, Apr 8, 2024 at 10:26 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:
> On Sun, Apr 07, 2024 at 03:00:00PM -0700, Andres Freund wrote:
> > >  src/backend/commands/analyze.c | 89 ++++++++++------------------------
> > >  1 file changed, 26 insertions(+), 63 deletions(-)
> >
> > That's a very nice demonstration of how this makes good prefetching easier...
>
> Agreed. Yay streaming read API and Bilal!

+1

I found a few comments to tweak, just a couple of places that hadn't
got the memo after we renamed "read stream", and an obsolete mention
of pinning buffers.  I adjusted those directly.

I ran some tests on a random basic Linux/ARM cloud box with a 7.6GB
table, and I got:

                                      cold    hot
master:                             9025ms  199ms
patched, io_combine_limit=1:        9025ms  191ms
patched, io_combine_limit=default:  8729ms  191ms

Despite being random, occasionally some I/Os must get merged, allowing
slightly better random throughput when accessing disk blocks through a
3000 IOPS drinking straw.  Looking at strace, I see 29144 pread* calls
instead of 30071, which fits that theory.  Let's see... if you roll a
fair 973452-sided dice 30071 times, how many times do you expect to
roll consecutive numbers?  Each time you roll there is a 1/973452
chance that you get the last number + 1, and we have 30071 tries
giving 30071/973452 = ~3%.  9025ms minus 3% is 8754ms.  Seems about
right.

I am not sure why the hot number is faster exactly.  (Anecdotally, I
did notice that in the cases that beat master semi-unexpectedly like
this, my software memory prefetch patch doesn't help or hurt, while in
some cases and on some CPUs there is little difference, and then that
patch seems to get a speed-up like this, which might be a clue.
*Shrug*, investigation needed.)

Pushed.  Thanks Bilal and reviewers!

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-04-08T03:46:42Z

On Mon, Apr 8, 2024 at 10:26 AM Melanie Plageman
<melanieplageman@gmail.com> wrote:
> On Sun, Apr 07, 2024 at 03:00:00PM -0700, Andres Freund wrote:
> > On 2024-04-07 16:59:26 -0400, Melanie Plageman wrote:
> > > From 862b7ac81cdafcda7b525e02721da14e46265509 Mon Sep 17 00:00:00 2001
> > > From: Melanie Plageman <melanieplageman@gmail.com>
> > > Date: Sun, 7 Apr 2024 15:38:41 -0400
> > > Subject: [PATCH v10 3/3] Obsolete BlockSampler_HasMore()
> > >
> > > A previous commit stopped using BlockSampler_HasMore() for flow control
> > > in acquire_sample_rows(). There seems little use now for
> > > BlockSampler_HasMore(). It should be sufficient to return
> > > InvalidBlockNumber from BlockSampler_Next() when BlockSample_HasMore()
> > > would have returned false. Remove BlockSampler_HasMore().
> > >
> > > Author: Melanie Plageman, Nazir Bilal Yavuz
> > > Discussion: https://postgr.es/m/flat/CAN55FZ0UhXqk9v3y-zW_fp4-WCp43V8y0A72xPmLkOM%2B6M%2BmJg%40mail.gmail.com
> >
> > The justification here seems somewhat odd. Sure, the previous commit stopped
> > using BlockSampler_HasMore in acquire_sample_rows - but only because it was
> > moved to block_sampling_streaming_read_next()?
>
> It didn't stop using it. It stopped being useful. The reason it existed,
> as far as I can tell, was to use it as the while() loop condition in
> acquire_sample_rows(). I think it makes much more sense for
> BlockSampler_Next() to return InvalidBlockNumber when there are no more
> blocks -- not to assert you don't call it when there aren't any more
> blocks.
>
> I didn't want to change BlockSampler_Next() in the same commit as the
> streaming read user and we can't remove BlockSampler_HasMore() without
> changing BlockSampler_Next().

I agree that the code looks useless if one condition implies the
other, but isn't it good to keep that cross-check, perhaps
reformulated as an assertion?  I didn't look too hard at the maths, I
just saw the words "It is not obvious that this code matches Knuth's
Algorithm S ..." and realised I'm not sure I have time to develop a
good opinion about this today.  So I'll leave the 0002 change out for
now, as it's a tidy-up that can easily be applied in the next cycle.

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-04-16T15:10:04Z

Hi,

On Wed, 3 Apr 2024 at 22:25, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>
> Hi,
>
> Thank you for looking into this!
>
> On Wed, 3 Apr 2024 at 20:17, Heikki Linnakangas <hlinnaka@iki.fi> wrote:
> >
> > On 03/04/2024 13:31, Nazir Bilal Yavuz wrote:
> > > Streaming API has been committed but the committed version has a minor
> > > change, the read_stream_begin_relation function takes Relation instead
> > > of BufferManagerRelation now. So, here is a v5 which addresses this
> > > change.
> >
> > I'm getting a repeatable segfault / assertion failure with this:
> >
> > postgres=# CREATE TABLE tengiga (i int, filler text) with (fillfactor=10);
> > CREATE TABLE
> > postgres=# insert into tengiga select g, repeat('x', 900) from
> > generate_series(1, 1400000) g;
> > INSERT 0 1400000
> > postgres=# set default_statistics_target = 10; ANALYZE tengiga;
> > SET
> > ANALYZE
> > postgres=# set default_statistics_target = 100; ANALYZE tengiga;
> > SET
> > ANALYZE
> > postgres=# set default_statistics_target =1000; ANALYZE tengiga;
> > SET
> > server closed the connection unexpectedly
> >         This probably means the server terminated abnormally
> >         before or while processing the request.
> >
> > TRAP: failed Assert("BufferIsValid(hscan->rs_cbuf)"), File:
> > "heapam_handler.c", Line: 1079, PID: 262232
> > postgres: heikki postgres [local]
> > ANALYZE(ExceptionalCondition+0xa8)[0x56488a0de9d8]
> > postgres: heikki postgres [local]
> > ANALYZE(heapam_scan_analyze_next_block+0x63)[0x5648899ece34]
> > postgres: heikki postgres [local] ANALYZE(+0x2d3f34)[0x564889b6af34]
> > postgres: heikki postgres [local] ANALYZE(+0x2d2a3a)[0x564889b69a3a]
> > postgres: heikki postgres [local] ANALYZE(analyze_rel+0x33e)[0x564889b68fa9]
> > postgres: heikki postgres [local] ANALYZE(vacuum+0x4b3)[0x564889c2dcc0]
> > postgres: heikki postgres [local] ANALYZE(ExecVacuum+0xd6f)[0x564889c2d7fe]
> > postgres: heikki postgres [local]
> > ANALYZE(standard_ProcessUtility+0x901)[0x564889f0b8b9]
> > postgres: heikki postgres [local]
> > ANALYZE(ProcessUtility+0x136)[0x564889f0afb1]
> > postgres: heikki postgres [local] ANALYZE(+0x6728c8)[0x564889f098c8]
> > postgres: heikki postgres [local] ANALYZE(+0x672b3b)[0x564889f09b3b]
> > postgres: heikki postgres [local] ANALYZE(PortalRun+0x320)[0x564889f09015]
> > postgres: heikki postgres [local] ANALYZE(+0x66b2c6)[0x564889f022c6]
> > postgres: heikki postgres [local]
> > ANALYZE(PostgresMain+0x80c)[0x564889f06fd7]
> > postgres: heikki postgres [local] ANALYZE(+0x667876)[0x564889efe876]
> > postgres: heikki postgres [local]
> > ANALYZE(postmaster_child_launch+0xe6)[0x564889e1f4b3]
> > postgres: heikki postgres [local] ANALYZE(+0x58e68e)[0x564889e2568e]
> > postgres: heikki postgres [local] ANALYZE(+0x58b7f0)[0x564889e227f0]
> > postgres: heikki postgres [local]
> > ANALYZE(PostmasterMain+0x152b)[0x564889e2214d]
> > postgres: heikki postgres [local] ANALYZE(+0x4444b4)[0x564889cdb4b4]
> > /lib/x86_64-linux-gnu/libc.so.6(+0x2724a)[0x7f7d83b6724a]
> > /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x85)[0x7f7d83b67305]
> > postgres: heikki postgres [local] ANALYZE(_start+0x21)[0x564889971a61]
> > 2024-04-03 20:15:49.157 EEST [262101] LOG:  server process (PID 262232)
> > was terminated by signal 6: Aborted
>
> I realized the same error while working on Jakub's benchmarking results.
>
> Cause: I was using the nblocks variable to check how many blocks will
> be returned from the streaming API. But I realized that sometimes the
> number returned from BlockSampler_Init() is not equal to the number of
> blocks that BlockSampler_Next() will return as BlockSampling algorithm
> decides how many blocks to return on the fly by using some random
> seeds.

I wanted to re-check this problem and I realized that I was wrong. I
tried using nblocks again and this time there was no failure. I looked
at block sampling logic and I am pretty sure that BlockSampler_Init()
function correctly returns the number of blocks that
BlockSampler_Next() will return. It seems 158f581923 fixed this issue
as well.

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-04-29T15:41:09Z

Hi,

On Mon, 8 Apr 2024 at 04:21, Thomas Munro <thomas.munro@gmail.com> wrote:
>
> Pushed.  Thanks Bilal and reviewers!

I wanted to discuss what will happen to this patch now that
27bc1772fc8 is reverted. I am continuing this thread but I can create
another thread if you prefer so.

After the revert of 27bc1772fc8, acquire_sample_rows() became table-AM
agnostic again. So, read stream changes might have to be pushed down
now but there are a couple of roadblocks like Melanie mentioned [1]
before.

Quote from Melanie [1]:

On Thu, 11 Apr 2024 at 19:19, Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> I am working on pushing streaming ANALYZE into heap AM code, and I ran
> into a few roadblocks.
>
> If we want ANALYZE to make the ReadStream object in heap_beginscan()
> (like the read stream implementation of heap sequential and TID range
> scans do), I don't see any way around changing the scan_begin table AM
> callback to take a BufferAccessStrategy at the least (and perhaps also
> the BlockSamplerData).
>
> read_stream_begin_relation() doesn't just save the
> BufferAccessStrategy in the ReadStream, it uses it to set various
> other things in the ReadStream object. callback_private_data (which in
> ANALYZE's case is the BlockSamplerData) is simply saved in the
> ReadStream, so it could be set later, but that doesn't sound very
> clean to me.
>
> As such, it seems like a cleaner alternative would be to add a table
> AM callback for creating a read stream object that takes the
> parameters of read_stream_begin_relation(). But, perhaps it is a bit
> late for such additions.

If we do not want to add a new table AM callback like Melanie
mentioned, it is pretty much required to pass BufferAccessStrategy and
BlockSamplerData to the initscan().

> It also opens us up to the question of whether or not sequential scan
> should use such a callback instead of making the read stream object in
> heap_beginscan().
>
> I am happy to write a patch that does any of the above. But, I want to
> raise these questions, because perhaps I am simply missing an obvious
> alternative solution.

I wonder the same, I could not think of any alternative solution to
this problem.

Another quote from Melanie [2] in the same thread:

On Thu, 11 Apr 2024 at 20:48, Melanie Plageman
<melanieplageman@gmail.com> wrote:
>
> I will also say that, had this been 6 months ago, I would probably
> suggest we restructure ANALYZE's table AM interface to accommodate
> read stream setup and to address a few other things I find odd about
> the current code. For example, I think creating a scan descriptor for
> the analyze scan in acquire_sample_rows() is quite odd. It seems like
> it would be better done in the relation_analyze callback. The
> relation_analyze callback saves some state like the callbacks for
> acquire_sample_rows() and the Buffer Access Strategy. But at least in
> the heap implementation, it just saves them in static variables in
> analyze.c. It seems like it would be better to save them in a useful
> data structure that could be accessed later. We have access to pretty
> much everything we need at that point (in the relation_analyze
> callback). I also think heap's implementation of
> table_beginscan_analyze() doesn't need most of
> heap_beginscan()/initscan(), so doing this instead of something
> ANALYZE specific seems more confusing than helpful.

If we want to implement ANALYZE specific counterparts of
heap_beginscan()/initscan(); we may think of passing
BufferAccessStrategy and BlockSamplerData to them.

Also, there is an ongoing(?) discussion about a few problems /
improvements about the acquire_sample_rows() mentioned at the end of
the 'Table AM Interface Enhancements' thread [3]. Should we wait for
these discussions to be resolved or can we resume working on this
patch?

Any kind of feedback would be appreciated.

[1] https://www.postgresql.org/message-id/CAAKRu_ZxU6hucckrT1SOJxKfyN7q-K4KU1y62GhDwLBZWG%2BROg%40mail.gmail.com
[2] https://www.postgresql.org/message-id/CAAKRu_YkphAPNbBR2jcLqnxGhDEWTKhYfLFY%3D0R_oG5LHBH7Gw%40mail.gmail.com
[3] https://www.postgresql.org/message-id/flat/CAPpHfdurb9ycV8udYqM%3Do0sPS66PJ4RCBM1g-bBpvzUfogY0EA%40mail.gmail.com

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Nazir Bilal Yavuz <byavuz81@gmail.com> — 2024-05-15T18:18:12Z

Hi,

On Mon, 29 Apr 2024 at 18:41, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>
> Hi,
>
> On Mon, 8 Apr 2024 at 04:21, Thomas Munro <thomas.munro@gmail.com> wrote:
> >
> > Pushed.  Thanks Bilal and reviewers!
>
> I wanted to discuss what will happen to this patch now that
> 27bc1772fc8 is reverted. I am continuing this thread but I can create
> another thread if you prefer so.

041b96802ef is discussed in the 'Table AM Interface Enhancements'
thread [1]. The main problems discussed about this commit is that the
read stream API is not pushed to the heap-specific code and, because
of that, the other AM implementations need to use read streams. To
push read stream API to the heap-specific code, it is pretty much
required to pass BufferAccessStrategy and BlockSamplerData to the
initscan().

I am sharing the alternative version of this patch. The first patch
just reverts 041b96802ef and the second patch is the alternative
version.

In this alternative version, the read stream API is not pushed to the
heap-specific code, but it is controlled by the heap-specific code.
The SO_USE_READ_STREAMS_IN_ANALYZE flag is introduced and set in the
heap-specific code if the scan type is 'ANALYZE'. This flag is used to
decide whether streaming API in ANALYZE will be used or not. If this
flag is set, this means heap AMs and read stream API will be used. If
it is not set, this means heap AMs will not be used and code falls
back to the version before read streams.

Pros of the alternative version:

* The existing AM implementations other than heap AM can continue to
use their AMs without any change.
* AM implementations other than heap do not need to use read streams.
* Upstream code uses the read stream API and benefits from that.

Cons of the alternative version:

* 6 if cases are added to the acquire_sample_rows() function and 3 of
them are in the while loop.
* Because of these changes, the code looks messy.

Any kind of feedback would be appreciated.

[1] https://www.postgresql.org/message-id/flat/CAPpHfdurb9ycV8udYqM%3Do0sPS66PJ4RCBM1g-bBpvzUfogY0EA%40mail.gmail.com

-- 
Regards,
Nazir Bilal Yavuz
Microsoft

Re: Use streaming read API in ANALYZE

Melanie Plageman <melanieplageman@gmail.com> — 2024-05-20T20:46:35Z

On Wed, May 15, 2024 at 2:18 PM Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
>
> On Mon, 29 Apr 2024 at 18:41, Nazir Bilal Yavuz <byavuz81@gmail.com> wrote:
> >
> > On Mon, 8 Apr 2024 at 04:21, Thomas Munro <thomas.munro@gmail.com> wrote:
> > I wanted to discuss what will happen to this patch now that
> > 27bc1772fc8 is reverted. I am continuing this thread but I can create
> > another thread if you prefer so.
>
> 041b96802ef is discussed in the 'Table AM Interface Enhancements'
> thread [1]. The main problems discussed about this commit is that the
> read stream API is not pushed to the heap-specific code and, because
> of that, the other AM implementations need to use read streams. To
> push read stream API to the heap-specific code, it is pretty much
> required to pass BufferAccessStrategy and BlockSamplerData to the
> initscan().
>
> I am sharing the alternative version of this patch. The first patch
> just reverts 041b96802ef and the second patch is the alternative
> version.
>
> In this alternative version, the read stream API is not pushed to the
> heap-specific code, but it is controlled by the heap-specific code.
> The SO_USE_READ_STREAMS_IN_ANALYZE flag is introduced and set in the
> heap-specific code if the scan type is 'ANALYZE'. This flag is used to
> decide whether streaming API in ANALYZE will be used or not. If this
> flag is set, this means heap AMs and read stream API will be used. If
> it is not set, this means heap AMs will not be used and code falls
> back to the version before read streams.

Personally, I think the alternative version here is the best option
other than leaving what is in master. However, I would vote for
keeping what is in master because 1) where we are in the release
timeline and 2) the acquire_sample_rows() code, before streaming read,
was totally block-based anyway.

If we kept what was in master, do we need to document for table AMs
how to use read_stream_next_buffer() or can we assume they will look
at the heap AM implementation?

- Melanie

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-08-22T07:31:06Z

On Mon, May 20, 2024 at 10:46 PM Melanie Plageman <melanieplageman@gmail.com>
wrote:

> On Wed, May 15, 2024 at 2:18 PM Nazir Bilal Yavuz <byavuz81@gmail.com>
> wrote:
> >
> > On Mon, 29 Apr 2024 at 18:41, Nazir Bilal Yavuz <byavuz81@gmail.com>
> wrote:
> > >
> > > On Mon, 8 Apr 2024 at 04:21, Thomas Munro <thomas.munro@gmail.com>
> wrote:
> > > I wanted to discuss what will happen to this patch now that
> > > 27bc1772fc8 is reverted. I am continuing this thread but I can create
> > > another thread if you prefer so.
> >
> > 041b96802ef is discussed in the 'Table AM Interface Enhancements'
> > thread [1]. The main problems discussed about this commit is that the
> > read stream API is not pushed to the heap-specific code and, because
> > of that, the other AM implementations need to use read streams. To
> > push read stream API to the heap-specific code, it is pretty much
> > required to pass BufferAccessStrategy and BlockSamplerData to the
> > initscan().
> >
> > I am sharing the alternative version of this patch. The first patch
> > just reverts 041b96802ef and the second patch is the alternative
> > version.
> >
> > In this alternative version, the read stream API is not pushed to the
> > heap-specific code, but it is controlled by the heap-specific code.
> > The SO_USE_READ_STREAMS_IN_ANALYZE flag is introduced and set in the
> > heap-specific code if the scan type is 'ANALYZE'. This flag is used to
> > decide whether streaming API in ANALYZE will be used or not. If this
> > flag is set, this means heap AMs and read stream API will be used. If
> > it is not set, this means heap AMs will not be used and code falls
> > back to the version before read streams.
>
> Personally, I think the alternative version here is the best option
> other than leaving what is in master. However, I would vote for
> keeping what is in master because 1) where we are in the release
> timeline and 2) the acquire_sample_rows() code, before streaming read,
> was totally block-based anyway.
>
> If we kept what was in master, do we need to document for table AMs
> how to use read_stream_next_buffer() or can we assume they will look
> at the heap AM implementation?
>

Hi all,

I ran into this with the PG17 beta3 and for our use-case we need to set up
another stream (using a different relation and/or fork, but using the same
strategy) in addition to the one that is passed in to the
scan_analyze_next_block(), so to be able to do that it is necessary to have
the block sampler and the strategy from the original stream. Given that
this makes it very difficult (see below) to set up a different ReadStream
inside the TAM unless you have the BlockSampler and the BufferReadStrategy,
and the old interface did not have this problem, I would consider this a
regression.

This would be possible to solve in a few different ways:

   1. The alternate version proposed by Nazir allows you to decide which
   interface to use.
   2. Reverting the patch entirely would also solve the problem.
   3. Passing down the block sampler and the strategy to scan_begin() and
   move the ReadStream setup in analyze.c into initscan() in heapam.c, but
   this requires adding new parameters to this function.
   4. Having accessors that allow you to get the block sampler and strategy
   from the ReadStream object.

The proposed solution 1 above would still not solve the problem of allowing
a user to set up a different or extra ReadStream if they want to use the
new ReadStream interface. Reverting the ReadStream patch entirely would
also deal with the regression, but I find the ReadStream interface very
elegant since it moves the block sampling into a separate abstraction and
would like to use it, but right now there are some limitations if you want
to use it fully. The third solution above would allow that, but it requires
a change in the signature of scan_begin(), which might not be the best at
this stage of development. Proposal 4 would allow you to construct a new
stream based on the old one and might be a simple alternative solution as
well with less changes to the current code.

It is possible to capture the information in ProcessUtility() and
re-compute all the parameters, but that is quite a lot of work to get
right, especially considering that these computations are all over the
place and part of different functions at different stages (For example,
variable ring_size, needed to set up the buffer access strategy is computed
in ExecVacuum(); variable targrows, used to set up the buffer sampler, is
computed inside acquire_sample_rows(), which in turn requires to decide
what attributes to analyze, which is computed in do_analyze_rel().)

It would be great if this could be fixed before the PG17 release now that
27bc1772fc8 was reverted.
--
Best wishes,
Mats Kindahl, Timescale

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-08-24T03:33:19Z

On Thu, Aug 22, 2024 at 7:31 PM Mats Kindahl <mats@timescale.com> wrote:
> The alternate version proposed by Nazir allows you to decide which interface to use.
> Reverting the patch entirely would also solve the problem.
> Passing down the block sampler and the strategy to scan_begin() and move the ReadStream setup in analyze.c into initscan() in heapam.c, but this requires adding new parameters to this function.
> Having accessors that allow you to get the block sampler and strategy from the ReadStream object.

I'm a bit confused about how it can make sense to use the same
BlockSampler with a side relation/fork.  Could you point me at the
code?

> It would be great if this could be fixed before the PG17 release now that 27bc1772fc8 was reverted.

Ack.  Thinking...

Random thought: is there a wiki page or something where we can find
out about all the table AM projects?  For the successor to
27bc1772fc8, I hope they'll be following along.

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-08-29T13:15:40Z

On Sat, Aug 24, 2024 at 5:31 AM Thomas Munro <thomas.munro@gmail.com> wrote:

> On Thu, Aug 22, 2024 at 7:31 PM Mats Kindahl <mats@timescale.com> wrote:
> > The alternate version proposed by Nazir allows you to deide which
> interface to use.
> > Reverting the patch entirely would also solve the problem.
>

After digging through the code a little more I discovered that
there actually is another one: move the ReadStream struct into
read_stream.h.

> > Passing down the block sampler and the strategy to scan_begin() and move
> the ReadStream setup in analyze.c into initscan() in heapam.c, but this
> requires adding new parameters to this function.
> > Having accessors that allow you to get the block sampler and strategy
> from the ReadStream object.
>
> I'm a bit confused about how it can make sense to use the same
> BlockSampler with a side relation/fork.  Could you point me at the
> code?
>

Sorry, that was a bit unclear. Intention was not to re-use the block
sampler but to set a new one up with parameters from the original block
sampler, which would require access to it. (The strategy is less of a
problem since only one is used.)

To elaborate on the situation:

For the TAM in question we have two different storage areas, both are
heaps. Both relations use the same attributes "publicly" (they are
internally different, but we transform them to look the same). One of the
relations is the "default" one and is stored in rd_rel. In order to run
ANALYZE, we need to sample blocks from both relations, in slightly
different ways.

With the old interface, we faked the number of blocks in relation_size()
callback and claimed that there were N + M blocks. When then being asked
about a block by block number, we could easily pick the correct relation
and just forward the call.

With the new ReadStream API, a read-stream is (automatically) set up on the
"default" relation, but we can set up a separate read-stream inside the TAM
for the other relation. However, the difficulty is in setting it up
correctly:

We cannot use the "fake number of block"-trick since the read stream does
not only compute the block number, but actually tries to read the buffer in
the relation provided when setting up the read stream, so a block number
outside the range of this relation will not be found since it is in a
different relation.

If we could create our own read stream with both relations, that could be
solved and we could just implement the same logic, but direct it to the
correct relations depending on where we want to read the block. Unless I am
mistaken, there is already support for this since there is an array of
in-progress I/O and it would be trivial to extend this with more
relations+forks, if you have access to the structure definition. The
ReadStream struct is, however, an opaque struct so it's hard to hack around
with it. Just making the struct declaration public would potentially solve
a lot of problems here. (See attached patch, which is close to the minimum
of what is needed to allow extension writers to tweak the contents.)

Since both relations are using the same attributes with the same
"analyzability", having that information would be useful to compute the
targrows for setting up the additional stream, but it is computed in
do_analyze_rel() and not further propagated, so it needs to be re-computed
if we want to set up a separate read-stream.

> > It would be great if this could be fixed before the PG17 release now
> that 27bc1772fc8 was reverted.
>
> Ack.  Thinking...
>

Right now I think that just making the ReadStream struct available in the
header file is the best approach. It is a safe and low-risk fix (so
something that can be added to a beta) and will allow extension writers to
hack to their hearts' contents. In addition to that, being able to select
what interface to use would also help.

> Random thought: is there a wiki page or something where we can find
> out about all the table AM projects?  For the successor to
> 27bc1772fc8, I hope they'll be following along.
>

At this point, unfortunately not, we are quite early in this. Once I have
something, I'll share.
-- 
Best wishes,
Mats Kindahl, Timescale

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-09-04T10:40:47Z

Thanks for the explanation.  I think we should revert it.  IMHO it was
a nice clean example of a streaming transformation, but unfortunately
it transformed an API that nobody liked in the first place, and broke
some weird and wonderful workarounds.  Let's try again in 18.

Re: Use streaming read API in ANALYZE

Robert Haas <robertmhaas@gmail.com> — 2024-09-04T15:36:21Z

On Wed, Sep 4, 2024 at 6:38 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> Thanks for the explanation.  I think we should revert it.  IMHO it was
> a nice clean example of a streaming transformation, but unfortunately
> it transformed an API that nobody liked in the first place, and broke
> some weird and wonderful workarounds.  Let's try again in 18.

The problem I have with this is that we just released RC1. I suppose
if we have to make this change it's better to do it sooner than later,
but are we sure we want to whack this around this close to final
release?

-- 
Robert Haas
EDB: http://www.enterprisedb.com

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-09-04T23:34:21Z

On Thu, Sep 5, 2024 at 3:36 AM Robert Haas <robertmhaas@gmail.com> wrote:
> On Wed, Sep 4, 2024 at 6:38 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> > Thanks for the explanation.  I think we should revert it.  IMHO it was
> > a nice clean example of a streaming transformation, but unfortunately
> > it transformed an API that nobody liked in the first place, and broke
> > some weird and wonderful workarounds.  Let's try again in 18.
>
> The problem I have with this is that we just released RC1. I suppose
> if we have to make this change it's better to do it sooner than later,
> but are we sure we want to whack this around this close to final
> release?

I hear you.  But I definitely don't want to (and likely can't at this
point) make any of the other proposed changes, and I also don't want
to break Timescale.  That seems to leave only one option: go back to
the v16 API for RC2, and hope that the ongoing table AM discussions
for v18 (CF #4866) will fix all the problems for the people whose TAMs
don't quack like a "heap", and the people whose TAMs do and who would
not like to duplicate the code, and the people who want streaming I/O.

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-09-05T06:45:18Z

On Thu, Sep 5, 2024 at 1:34 AM Thomas Munro <thomas.munro@gmail.com> wrote:

> On Thu, Sep 5, 2024 at 3:36 AM Robert Haas <robertmhaas@gmail.com> wrote:
> > On Wed, Sep 4, 2024 at 6:38 AM Thomas Munro <thomas.munro@gmail.com>
> wrote:
> > > Thanks for the explanation.  I think we should revert it.  IMHO it was
> > > a nice clean example of a streaming transformation, but unfortunately
> > > it transformed an API that nobody liked in the first place, and broke
> > > some weird and wonderful workarounds.  Let's try again in 18.
> >
> > The problem I have with this is that we just released RC1. I suppose
> > if we have to make this change it's better to do it sooner than later,
> > but are we sure we want to whack this around this close to final
> > release?
>
> I hear you.  But I definitely don't want to (and likely can't at this
> point) make any of the other proposed changes, and I also don't want
> to break Timescale.  That seems to leave only one option: go back to
> the v16 API for RC2, and hope that the ongoing table AM discussions
> for v18 (CF #4866) will fix all the problems for the people whose TAMs
> don't quack like a "heap", and the people whose TAMs do and who would
> not like to duplicate the code, and the people who want streaming I/O.
>

Forgive me for asking, but I am not entirely sure why the ReadStream struct
is opaque. The usual reasons are:

   - You want to provide an ABI to allow extensions to work with new major
   versions without re-compiling. Right now it is necessary to recompile
   extensions anyway, this does not seem to apply. (Because there are a lot of
   other changes that you need when switching versions because of the lack of
   a stable ABI for other parts of the code. However, it might be that the
   goal is to support it eventually, and then it would make sense to start
   making structs opaque.)
   - You want to ensure that you can make modifications *inside* a major
   version without breaking ABIs and requiring a re-compile. In this case, you
   could still follow safe practice of adding new fields last, not relying on
   the size of the struct for anything (e.g., no arrays of these structures,
   just pointers to them), etc. However, if you want to be *very* safe and
   support very drastic changes inside a major version, it needs to be opaque,
   so this could be the reason.

Is it either of these reasons, or is there another reason?

Making the ReadStream API non-opaque (that is, moving the definition to the
header file) would at least solve our problem (unless I am mistaken).
However, I am ignorant about long-term plans which might affect this, so
there might be a good reason to revert it for reasons I am not aware of.
-- 
Best wishes,
Mats Kindahl, Timescale

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-09-05T09:12:07Z

On Thu, Sep 5, 2024 at 6:45 PM Mats Kindahl <mats@timescale.com> wrote:
> Forgive me for asking, but I am not entirely sure why the ReadStream struct is opaque. The usual reasons are:
>
> You want to provide an ABI to allow extensions to work with new major versions without re-compiling. Right now it is necessary to recompile extensions anyway, this does not seem to apply. (Because there are a lot of other changes that you need when switching versions because of the lack of a stable ABI for other parts of the code. However, it might be that the goal is to support it eventually, and then it would make sense to start making structs opaque.)
> You want to ensure that you can make modifications inside a major version without breaking ABIs and requiring a re-compile. In this case, you could still follow safe practice of adding new fields last, not relying on the size of the struct for anything (e.g., no arrays of these structures, just pointers to them), etc. However, if you want to be very safe and support very drastic changes inside a major version, it needs to be opaque, so this could be the reason.
>
> Is it either of these reasons, or is there another reason?
>
> Making the ReadStream API non-opaque (that is, moving the definition to the header file) would at least solve our problem (unless I am mistaken). However, I am ignorant about long-term plans which might affect this, so there might be a good reason to revert it for reasons I am not aware of.

The second thing.  Also there are very active plans[1] to change the
internal design of ReadStream in 18, since the goal is to drive true
asynchronous I/O, and the idea of ReadStream was to create a simple
API to let many consumers start using it, so that we can drive
efficient modern system interfaces below that API, so having people
depending on how it works would not be great.

But let's talk about how that would actually look, for example if we
exposed the struct or you took a photocopy of it...  I think your idea
must be something like: if you could access struct ReadStream's
internals, you could replace stream->callback with an interceptor
callback, and if the BlockSampler had been given the fake N + M
relation size, the interceptor could overwrite
stream->ios[next_io_index].op.smgr and return x - N if the intercepted
callback returned x >= N.  (Small detail: need to check
stream->fast_path and use 0 instead or something like that, but maybe
we could change that.)  One minor problem that jumps out is that
read_stream.c could inappropriately merge blocks from the two
relations into one I/O.  Hmm, I guess you'd have to teach the
interceptor not to allow that: if switching between the two relation,
and if the block number would coincide with
stream->pending_read_blocknum + stream->pending_read_nblocks, it would
need to pick a new block instead (interfering with the block sampling
algorithm, but only very rarely).  Is this what you had in mind, or
something else?

(BTW I have a patch to teach read_stream.c about multi-smgr-relation
streams, by adding a different constructor with a different callback
that returns smgr, fork, block instead of just the block, but it
didn't make it into 17.)

[1] https://www.postgresql.org/message-id/flat/uvrtrknj4kdytuboidbhwclo4gxhswwcpgadptsjvjqcluzmah@brqs62irg4dt

Re: Use streaming read API in ANALYZE

Michael Banck <mbanck@gmx.net> — 2024-09-09T15:36:39Z

Hi,

On Thu, Sep 05, 2024 at 09:12:07PM +1200, Thomas Munro wrote:
> On Thu, Sep 5, 2024 at 6:45 PM Mats Kindahl <mats@timescale.com> wrote:
> > Making the ReadStream API non-opaque (that is, moving the definition
> > to the header file) would at least solve our problem (unless I am
> > mistaken). However, I am ignorant about long-term plans which might
> > affect this, so there might be a good reason to revert it for
> > reasons I am not aware of.
> 
> The second thing.

I am a bit confused about the status of this thread. Robert mentioned
RC1, so I guess it pertains to v17 but I don't see it on the open item
wiki list?

Does the above mean you are going to revert it for v17, Thomas? And if
so, what exactly? The ANALYZE changes on top of the streaming read API
or something else about that API that is being discussed on this thread?

I am also asking because this feature (i.e. Use streaming read API in
ANALYZE) is being mentioned in the release announcement and that was
just frozen for translations.

Michael

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-09-09T22:27:43Z

On Tue, Sep 10, 2024 at 3:36 AM Michael Banck <mbanck@gmx.net> wrote:
> I am a bit confused about the status of this thread. Robert mentioned
> RC1, so I guess it pertains to v17 but I don't see it on the open item
> wiki list?

Yes, v17.  Alight, I'll add an item.

> Does the above mean you are going to revert it for v17, Thomas? And if
> so, what exactly? The ANALYZE changes on top of the streaming read API
> or something else about that API that is being discussed on this thread?

I might have been a little pessimistic in that assessment.  Another
workaround that seems an awful lot cleaner and less invasive would be
to offer a new ReadStream API function that provides access to block
numbers and the strategy, ie the arguments of v16's
scan_analyze_next_block() function.  Mats, what do you think about
this?  (I haven't tried to preserve the prefetching behaviour, which
probably didn't actually too work for you in v16 anyway at a guess,
I'm just looking for the absolute simplest thing we can do to resolve
this API mismatch.)  TimeScale could then continue to use its v16
coding to handle the two-relations-in-a-trenchcoat problem, and we
could continue discussing how to make v18 better.

I looked briefly at another non-heap-like table AM, the Citus Columnar
TAM.  I am not familiar with that code and haven't studied it deeply
this morning, but its _next_block() currently just returns true, so I
think it will somehow need to change to counting calls and returning
false when it thinks its been called enough times (otherwise the loop
in acquire_sample_rows() won't terminate, I think?).  I suppose an
easy way to do that without generating extra I/O or having to think
hard about how to preserve the loop cound from v16 would be to use
this function.

I think there are broadly three categories of TAMs with respect to
ANALYZE block sampling: those that are very heap-like (blocks of one
SMgrRelation) and can just use the stream directly, those that are not
at all heap-like (doing something completely different to sample
tuples and ignoring the block aspect but using _next_block() to
control the loop), and then Timescale's case which is sort of
somewhere in between: almost heap-like from the point of view of this
sampling code, ie working with blocks, but fudging the meaning of
block numbers, which we didn't anticipate.  (I wonder if it fails to
sample fairly across the underlying relation boundary anyway because
their data densities must surely be quite different, but that's not
what we're here to talk about.)

. o O { We need that wiki page listing TAMs with links to the open
source ones... }

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-09-10T04:04:00Z

On Tue, Sep 10, 2024 at 10:27 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> Mats, what do you think about
> this?  (I haven't tried to preserve the prefetching behaviour, which
> probably didn't actually too work for you in v16 anyway at a guess,
> I'm just looking for the absolute simplest thing we can do to resolve
> this API mismatch.)  TimeScale could then continue to use its v16
> coding to handle the two-relations-in-a-trenchcoat problem, and we
> could continue discussing how to make v18 better.

. o O { Spitballing here: if we add that tiny function I showed to get
you unstuck for v17, then later in v18, if we add a multi-relation
ReadStream constructor/callback (I have a patch somewhere, I want to
propose that as it is needed for streaming recovery), you could
construct a new ReadSteam of your own that is daisy-chained from that
one.  You could keep using your N + M block numbering scheme if you
want to, and the callback of the new stream could decode the block
numbers and redirect to the appropriate relation + real block number.
That way you'd get I/O concurrency for both relations (for now just
read-ahead advice, but see Andres's AIO v2 thread).  That'd
essentially be a more supported version of the 'access the struct
internals' idea (or at least my understanding of what you had in
mind), through daisy-chained streams.  A little weird maybe, and maybe
the redesign work will result in something completely
different/better... just a thought... }

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-09-10T17:07:18Z

On Thu, Sep 5, 2024 at 11:12 AM Thomas Munro <thomas.munro@gmail.com> wrote:

> On Thu, Sep 5, 2024 at 6:45 PM Mats Kindahl <mats@timescale.com> wrote:
> > Forgive me for asking, but I am not entirely sure why the ReadStream
> struct is opaque. The usual reasons are:
> >
> > You want to provide an ABI to allow extensions to work with new major
> versions without re-compiling. Right now it is necessary to recompile
> extensions anyway, this does not seem to apply. (Because there are a lot of
> other changes that you need when switching versions because of the lack of
> a stable ABI for other parts of the code. However, it might be that the
> goal is to support it eventually, and then it would make sense to start
> making structs opaque.)
> > You want to ensure that you can make modifications inside a major
> version without breaking ABIs and requiring a re-compile. In this case, you
> could still follow safe practice of adding new fields last, not relying on
> the size of the struct for anything (e.g., no arrays of these structures,
> just pointers to them), etc. However, if you want to be very safe and
> support very drastic changes inside a major version, it needs to be opaque,
> so this could be the reason.
> >
> > Is it either of these reasons, or is there another reason?
> >
> > Making the ReadStream API non-opaque (that is, moving the definition to
> the header file) would at least solve our problem (unless I am mistaken).
> However, I am ignorant about long-term plans which might affect this, so
> there might be a good reason to revert it for reasons I am not aware of.
>
> The second thing.  Also there are very active plans[1] to change the
> internal design of ReadStream in 18, since the goal is to drive true
> asynchronous I/O, and the idea of ReadStream was to create a simple
> API to let many consumers start using it, so that we can drive
> efficient modern system interfaces below that API, so having people
> depending on how it works would not be great.
>

That is understandable, since you usually do not want to have to re-compile
the extension for different minor versions. However, it would be a rare
case with extensions that are meddling with this, so might not turn out to
be a big problem in reality, as long as it is very clear to all involved
that this might change and that you make an effort to avoid binary
incompatibility by removing or changing types for fields.


> But let's talk about how that would actually look, for example if we
> exposed the struct or you took a photocopy of it...  I think your idea
> must be something like: if you could access struct ReadStream's
> internals, you could replace stream->callback with an interceptor
> callback, and if the BlockSampler had been given the fake N + M
> relation size, the interceptor could overwrite
> stream->ios[next_io_index].op.smgr and return x - N if the intercepted
> callback returned x >= N.  (Small detail: need to check
> stream->fast_path and use 0 instead or something like that, but maybe
> we could change that.)


Yes, this is what I had in mind, but I did not dig too deeply into the code.


> One minor problem that jumps out is that
> read_stream.c could inappropriately merge blocks from the two
> relations into one I/O.  Hmm, I guess you'd have to teach the
> interceptor not to allow that: if switching between the two relation,
> and if the block number would coincide with
> stream->pending_read_blocknum + stream->pending_read_nblocks, it would
> need to pick a new block instead (interfering with the block sampling
> algorithm, but only very rarely).  Is this what you had in mind, or
> something else?
>

Hmmm... I didn't look too closely at this. Since the block number comes
from the callback, I guess we could make sure to have a "padding" block
between the regions so that we "break" any suite of blocks, which I think
is what you mean with "teach the interceptor not to allow that", but I
would have to write a patch to make sure.


>
> (BTW I have a patch to teach read_stream.c about multi-smgr-relation
> streams, by adding a different constructor with a different callback
> that returns smgr, fork, block instead of just the block, but it
> didn't make it into 17.)
>

Without having looked at the patch, this sounds like the correct way to do
it.


>
> [1]
> https://www.postgresql.org/message-id/flat/uvrtrknj4kdytuboidbhwclo4gxhswwcpgadptsjvjqcluzmah@brqs62irg4dt
>


-- 
Best wishes,
Mats Kindahl, Timescale

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-09-13T08:01:23Z

On Tue, Sep 10, 2024 at 12:28 AM Thomas Munro <thomas.munro@gmail.com>
wrote:

> On Tue, Sep 10, 2024 at 3:36 AM Michael Banck <mbanck@gmx.net> wrote:
> > I am a bit confused about the status of this thread. Robert mentioned
> > RC1, so I guess it pertains to v17 but I don't see it on the open item
> > wiki list?
>
> Yes, v17.  Alight, I'll add an item.
>
> > Does the above mean you are going to revert it for v17, Thomas? And if
> > so, what exactly? The ANALYZE changes on top of the streaming read API
> > or something else about that API that is being discussed on this thread?
>
> I might have been a little pessimistic in that assessment.  Another
> workaround that seems an awful lot cleaner and less invasive would be
> to offer a new ReadStream API function that provides access to block
> numbers and the strategy, ie the arguments of v16's
> scan_analyze_next_block() function.  Mats, what do you think about
> this?  (I haven't tried to preserve the prefetching behaviour, which
> probably didn't actually too work for you in v16 anyway at a guess,
> I'm just looking for the absolute simplest thing we can do to resolve
> this API mismatch.)  TimeScale could then continue to use its v16
> coding to handle the two-relations-in-a-trenchcoat problem, and we
> could continue discussing how to make v18 better.
>

In the original code we could call the methods with an "adjusted" block
number, so the entire logic worked as before because we could just
recursively forward the call with modified parameters. This is a little
different with the new API.

> I looked briefly at another non-heap-like table AM, the Citus Columnar
> TAM.  I am not familiar with that code and haven't studied it deeply
> this morning, but its _next_block() currently just returns true, so I
> think it will somehow need to change to counting calls and returning
> false when it thinks its been called enough times (otherwise the loop
> in acquire_sample_rows() won't terminate, I think?).  I suppose an
> easy way to do that without generating extra I/O or having to think
> hard about how to preserve the loop cound from v16 would be to use
> this function.
>

Yes, but we are re-using the heapam so forwarding the call to it, which not
only fetches the next block it also reads the buffer. Since you could just
pass in the block number before, it just worked.

As mentioned, we intended to set up a new ReadStream for the "internal"
relation ourselves (I think this is what you mean with "daisy-chain" in the
followup to this mail), but then you need targrows, which is based on
vacattrstats, which is computed with code that is currently either inline
(the loop over the attributes in do_analyze_rel), or static (the
examine_attribute function). We can write our own code for this, it would
help to have the code that does this work callable, or be able to extract
parameters from the existing readstream to at least get a hint. This would
allow us to just get the vacuum attribute stats for an arbitrary relation
and then run the same computations as in do_analyze_rel. Being able to do
the same for the indexes is less important since this is an "internal"
relation and the "public" indexes are the ones that matter.

I attached a tentative patch for this, just doing some refactorings, and
will see if that is sufficient for the current work by trying to use it. (I
thought I would be able to verify this today, but am a little delayed so
I'm sending this anyway.)

A patch like this is a minimal refactoring so should be safe even in an RC.
I have deliberately not tried to do a more serious refactoring although I
see that there are some duplications when doing the same work with the
indexes and it would probably be possible to make a more generic function
for this.

> I think there are broadly three categories of TAMs with respect to
> ANALYZE block sampling: those that are very heap-like (blocks of one
> SMgrRelation) and can just use the stream directly, those that are not
> at all heap-like (doing something completely different to sample
> tuples and ignoring the block aspect but using _next_block() to
> control the loop), and then Timescale's case which is sort of
> somewhere in between: almost heap-like from the point of view of this
> sampling code, ie working with blocks, but fudging the meaning of
> block numbers, which we didn't anticipate.

In this case the block numbers are only from a different relation, so they
are still valid blocks, just encoded in a funny way. The block numbers
trick is just a hack, but the gist is that we want to sample an
arbitrary number of relations/forks when running analysis, not just the
"front-facing" one.

> (I wonder if it fails to
> sample fairly across the underlying relation boundary anyway because
> their data densities must surely be quite different, but that's not
> what we're here to talk about.)
>

Yes, they are, so this is kind-of-a-hack-to-get-it-roughly-correct. The
ideal scenario would be to be able to run the same analysis of that is done
in do_analyze_rel on the "hidden" relation to get an accurate targetrows.
This is what I am trying now with the attached patch.

>
> . o O { We need that wiki page listing TAMs with links to the open
> source ones... }

-- 
Best wishes,
Mats Kindahl, Timescale

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-09-13T08:10:13Z

On Tue, Sep 10, 2024 at 6:04 AM Thomas Munro <thomas.munro@gmail.com> wrote:

> On Tue, Sep 10, 2024 at 10:27 AM Thomas Munro <thomas.munro@gmail.com>
> wrote:
> > Mats, what do you think about
> > this?  (I haven't tried to preserve the prefetching behaviour, which
> > probably didn't actually too work for you in v16 anyway at a guess,
> > I'm just looking for the absolute simplest thing we can do to resolve
> > this API mismatch.)  TimeScale could then continue to use its v16
> > coding to handle the two-relations-in-a-trenchcoat problem, and we
> > could continue discussing how to make v18 better.
>
> . o O { Spitballing here: if we add that tiny function I showed to get
> you unstuck for v17, then later in v18, if we add a multi-relation
> ReadStream constructor/callback (I have a patch somewhere, I want to
> propose that as it is needed for streaming recovery), you could
> construct a new ReadSteam of your own that is daisy-chained from that
> one.  You could keep using your N + M block numbering scheme if you
> want to, and the callback of the new stream could decode the block
> numbers and redirect to the appropriate relation + real block number.
>

I think it is good to make as small changes as possible to the RC, so agree
with this approach. Looking at the patch. I think it will work, but I'll do
some experimentation with the patch.

Just asking, is there any particular reason why you do not want to *add*
new functions for opaque objects inside a major release? After all, that
was the reason they were opaque from the beginning and extending with new
functions would not break any existing code, not even from the ABI
perspective.


> That way you'd get I/O concurrency for both relations (for now just
> read-ahead advice, but see Andres's AIO v2 thread).  That'd
> essentially be a more supported version of the 'access the struct
> internals' idea (or at least my understanding of what you had in
> mind), through daisy-chained streams.  A little weird maybe, and maybe
> the redesign work will result in something completely
> different/better... just a thought... }
>

I'll take a look at the thread. I really think the ReadStream abstraction
is a good step in the right direction.
-- 
Best wishes,
Mats Kindahl, Timescale

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-09-14T12:14:29Z

On Fri, Sep 13, 2024 at 10:10 AM Mats Kindahl <mats@timescale.com> wrote:

> On Tue, Sep 10, 2024 at 6:04 AM Thomas Munro <thomas.munro@gmail.com>
> wrote:
>
>> On Tue, Sep 10, 2024 at 10:27 AM Thomas Munro <thomas.munro@gmail.com>
>> wrote:
>> > Mats, what do you think about
>> > this?  (I haven't tried to preserve the prefetching behaviour, which
>> > probably didn't actually too work for you in v16 anyway at a guess,
>> > I'm just looking for the absolute simplest thing we can do to resolve
>> > this API mismatch.)  TimeScale could then continue to use its v16
>> > coding to handle the two-relations-in-a-trenchcoat problem, and we
>> > could continue discussing how to make v18 better.
>>
>> . o O { Spitballing here: if we add that tiny function I showed to get
>> you unstuck for v17, then later in v18, if we add a multi-relation
>> ReadStream constructor/callback (I have a patch somewhere, I want to
>> propose that as it is needed for streaming recovery), you could
>> construct a new ReadSteam of your own that is daisy-chained from that
>> one.  You could keep using your N + M block numbering scheme if you
>> want to, and the callback of the new stream could decode the block
>> numbers and redirect to the appropriate relation + real block number.
>>
>
> I think it is good to make as small changes as possible to the RC, so
> agree with this approach. Looking at the patch. I think it will work, but
> I'll do some experimentation with the patch.
>
> Just asking, is there any particular reason why you do not want to *add*
> new functions for opaque objects inside a major release? After all, that
> was the reason they were opaque from the beginning and extending with new
> functions would not break any existing code, not even from the ABI
> perspective.
>
>
>> That way you'd get I/O concurrency for both relations (for now just
>> read-ahead advice, but see Andres's AIO v2 thread).  That'd
>> essentially be a more supported version of the 'access the struct
>> internals' idea (or at least my understanding of what you had in
>> mind), through daisy-chained streams.  A little weird maybe, and maybe
>> the redesign work will result in something completely
>> different/better... just a thought... }
>>
>
> I'll take a look at the thread. I really think the ReadStream abstraction
> is a good step in the right direction.
> --
> Best wishes,
> Mats Kindahl, Timescale
>

Hi Thomas,

I used the combination of your patch and making the computation of
vacattrstats for a relation available through the API and managed to
implement something that I think does the right thing. (I just sampled a
few different statistics to check if they seem reasonable, like most common
vals and most common freqs.) See attached patch.

I need the vacattrstats to set up the two streams for the internal
relations. I can just re-implement them in the same way as is already done,
but this seems like a small change that avoids unnecessary code
duplication.
-- 
Best wishes,
Mats Kindahl, Timescale

Re: Use streaming read API in ANALYZE

Thomas Munro <thomas.munro@gmail.com> — 2024-09-18T03:13:20Z

On Sun, Sep 15, 2024 at 12:14 AM Mats Kindahl <mats@timescale.com> wrote:
> I used the combination of your patch and making the computation of vacattrstats for a relation available through the API and managed to implement something that I think does the right thing. (I just sampled a few different statistics to check if they seem reasonable, like most common vals and most common freqs.) See attached patch.

Cool.  I went ahead and committed that small new function and will
mark the open item closed.

> I need the vacattrstats to set up the two streams for the internal relations. I can just re-implement them in the same way as is already done, but this seems like a small change that avoids unnecessary code duplication.

Unfortunately we're not in a phase where we can make non-essential
changes, we're right about to release and we're only committing fixes,
and it seems like you have a way forward (albeit with some
duplication).  We can keep talking about that for v18.

From your earlier email:
> I'll take a look at the thread. I really think the ReadStream abstraction is a good step in the right direction.

Here's something you or your colleagues might be interested in: I was
looking around for a fun extension to streamify as a demo of the
technology, and I finished up writing a quick patch to streamify
pgvector's HNSW index scan, which worked well enough to share[1] (I
think it should in principle be able to scale with the number of graph
connections, at least 16x), but then people told me that it's of
limited interest because everybody knows that HNSW indexes have to fit
in memory (I think there may also be memory prefetch streaming
opportunities, unexamined for now).  But that made me wonder what the
people with the REALLY big indexes do for hyperdimensional graph
search on a scale required to build Skynet, and that led me back to
Timescale pgvectorscale[2].  I see two obvious signs that this thing
is eminently and profitably streamifiable: (1) The stated aim is
optimising for indexes that don't fit in memory, hence "Disk" in the
name of the research project it is inspired by, (2) I see that
DIskANN[3] is aggressively using libaio (Linux) and overlapped/IOCP
(Windows).  So now I am waiting patiently for a Rustacean to show up
with patches for pgvectorscale to use ReadStream, which would already
get read-ahead advice and vectored I/O (Linux, macOS, FreeBSD soon
hopefully), and hopefully also provide a nice test case for the AIO
patch set which redirects buffer reads through io_uring (Linux,
basically the newer better libaio) or background I/O workers (other
OSes, which works surprisingly competitively).  Just BTW for
comparison with DiskANN we have also had early POC-quality patches
that drive AIO with overlapped/IOCP (Windows) which will eventually be
rebased and proposed (Windows isn't really a primary target but we
wanted to validate that the stuff we're working on has abstractions
that will map to the obvious system APIs found in the systems
PostgreSQL targets).  For completeness, I've also had it mostly
working on the POSIX AIO of FreeBSD, HP-UX and AIX (though we dropped
support for those last two so that was a bit of a dead end).

[1] https://www.postgresql.org/message-id/flat/CA%2BhUKGJ_7NKd46nx1wbyXWriuZSNzsTfm%2BrhEuvU6nxZi3-KVw%40mail.gmail.com
[2] https://github.com/timescale/pgvectorscale
[3] https://github.com/microsoft/DiskANN

Re: Use streaming read API in ANALYZE

Mats Kindahl <mats@timescale.com> — 2024-09-20T06:36:42Z

On Wed, Sep 18, 2024 at 5:13 AM Thomas Munro <thomas.munro@gmail.com> wrote:

> On Sun, Sep 15, 2024 at 12:14 AM Mats Kindahl <mats@timescale.com> wrote:
> > I used the combination of your patch and making the computation of
> vacattrstats for a relation available through the API and managed to
> implement something that I think does the right thing. (I just sampled a
> few different statistics to check if they seem reasonable, like most common
> vals and most common freqs.) See attached patch.
>
> Cool.  I went ahead and committed that small new function and will
> mark the open item closed.
>

Thank you Thomas, this will help a lot.


> > I need the vacattrstats to set up the two streams for the internal
> relations. I can just re-implement them in the same way as is already done,
> but this seems like a small change that avoids unnecessary code duplication.
>
> Unfortunately we're not in a phase where we can make non-essential
> changes, we're right about to release and we're only committing fixes,
> and it seems like you have a way forward (albeit with some
> duplication).  We can keep talking about that for v18.
>

Yes, I can work around this by re-implementing the same code that is
present in PostgreSQL.


>
> From your earlier email:
> > I'll take a look at the thread. I really think the ReadStream
> abstraction is a good step in the right direction.
>
> Here's something you or your colleagues might be interested in: I was
> looking around for a fun extension to streamify as a demo of the
> technology, and I finished up writing a quick patch to streamify
> pgvector's HNSW index scan, which worked well enough to share[1] (I
> think it should in principle be able to scale with the number of graph
> connections, at least 16x), but then people told me that it's of
> limited interest because everybody knows that HNSW indexes have to fit
> in memory (I think there may also be memory prefetch streaming
> opportunities, unexamined for now).  But that made me wonder what the
> people with the REALLY big indexes do for hyperdimensional graph
> search on a scale required to build Skynet, and that led me back to
> Timescale pgvectorscale[2].  I see two obvious signs that this thing
> is eminently and profitably streamifiable: (1) The stated aim is
> optimising for indexes that don't fit in memory, hence "Disk" in the
> name of the research project it is inspired by, (2) I see that
> DIskANN[3] is aggressively using libaio (Linux) and overlapped/IOCP
> (Windows).  So now I am waiting patiently for a Rustacean to show up
> with patches for pgvectorscale to use ReadStream, which would already
> get read-ahead advice and vectored I/O (Linux, macOS, FreeBSD soon
> hopefully), and hopefully also provide a nice test case for the AIO
> patch set which redirects buffer reads through io_uring (Linux,
> basically the newer better libaio) or background I/O workers (other
> OSes, which works surprisingly competitively).  Just BTW for
> comparison with DiskANN we have also had early POC-quality patches
> that drive AIO with overlapped/IOCP (Windows) which will eventually be
> rebased and proposed (Windows isn't really a primary target but we
> wanted to validate that the stuff we're working on has abstractions
> that will map to the obvious system APIs found in the systems
> PostgreSQL targets).  For completeness, I've also had it mostly
> working on the POSIX AIO of FreeBSD, HP-UX and AIX (though we dropped
> support for those last two so that was a bit of a dead end).




> [1]
> https://www.postgresql.org/message-id/flat/CA%2BhUKGJ_7NKd46nx1wbyXWriuZSNzsTfm%2BrhEuvU6nxZi3-KVw%40mail.gmail.com
> [2] https://github.com/timescale/pgvectorscale
> [3] https://github.com/microsoft/DiskANN
>

Thanks Thomas, this looks really interesting. I've forwarded it to the
pgvectorscale team.
-- 
Best wishes,
Mats Kindahl, Timescale