CLUSTER on partitioned index

@cfbot: rebased

On Tue, Oct 27, 2020 at 07:33:12PM -0500, Justin Pryzby wrote:

I'm attaching a counter-proposal to your catalog change, which preserves
indisclustered on children of clustered, partitioned indexes, and invalidates
indisclustered when attaching unclustered indexes.

..and now propagates CLUSTER ON to child indexes.

I left this as separate patches to show what I mean and what's new while we
discuss it.

--
Justin

On Wed, Nov 04, 2020 at 08:23:56PM -0600, Justin Pryzby wrote:

On Tue, Oct 27, 2020 at 07:33:12PM -0500, Justin Pryzby wrote:

I'm attaching a counter-proposal to your catalog change, which preserves
indisclustered on children of clustered, partitioned indexes, and invalidates
indisclustered when attaching unclustered indexes.

..and now propagates CLUSTER ON to child indexes.

I left this as separate patches to show what I mean and what's new while we
discuss it.

This fixes some omissions in the previous patch and error in its test cases.

CLUSTER ON recurses to children, since I think a clustered parent index means
that all its child indexes are clustered. "SET WITHOUT CLUSTER" doesn't have
to recurse to children, but I did it like that for consistency and it avoids
the need to special case InvalidOid.

--
Justin

On Sun, Nov 15, 2020 at 07:53:35PM -0600, Justin Pryzby wrote:

On Wed, Nov 04, 2020 at 08:23:56PM -0600, Justin Pryzby wrote:

On Tue, Oct 27, 2020 at 07:33:12PM -0500, Justin Pryzby wrote:

I'm attaching a counter-proposal to your catalog change, which preserves
indisclustered on children of clustered, partitioned indexes, and invalidates
indisclustered when attaching unclustered indexes.

..and now propagates CLUSTER ON to child indexes.

I left this as separate patches to show what I mean and what's new while we
discuss it.

This fixes some omissions in the previous patch and error in its test cases.

CLUSTER ON recurses to children, since I think a clustered parent index means
that all its child indexes are clustered. "SET WITHOUT CLUSTER" doesn't have
to recurse to children, but I did it like that for consistency and it avoids
the need to special case InvalidOid.

The previous patch failed pg_upgrade when restoring a clustered, parent index,
since it's marked INVALID until indexes have been built on all child tables, so
CLUSTER ON was rejected on invalid index.

So I think CLUSTER ON needs to be a separate pg_dump object, to allow attaching
the child index (thereby making the parent "valid") to happen before SET
CLUSTER on the parent index.

--
Justin

On Sat, Nov 28, 2020 at 08:03:02PM -0600, Justin Pryzby wrote:

On Sun, Nov 15, 2020 at 07:53:35PM -0600, Justin Pryzby wrote:

On Wed, Nov 04, 2020 at 08:23:56PM -0600, Justin Pryzby wrote:

On Tue, Oct 27, 2020 at 07:33:12PM -0500, Justin Pryzby wrote:

I'm attaching a counter-proposal to your catalog change, which preserves
indisclustered on children of clustered, partitioned indexes, and invalidates
indisclustered when attaching unclustered indexes.

..and now propagates CLUSTER ON to child indexes.

I left this as separate patches to show what I mean and what's new while we
discuss it.

This fixes some omissions in the previous patch and error in its test cases.

CLUSTER ON recurses to children, since I think a clustered parent index means
that all its child indexes are clustered. "SET WITHOUT CLUSTER" doesn't have
to recurse to children, but I did it like that for consistency and it avoids
the need to special case InvalidOid.

The previous patch failed pg_upgrade when restoring a clustered, parent index,
since it's marked INVALID until indexes have been built on all child tables, so
CLUSTER ON was rejected on invalid index.

So I think CLUSTER ON needs to be a separate pg_dump object, to allow attaching
the child index (thereby making the parent "valid") to happen before SET
CLUSTER on the parent index.

Rebased on b5913f612 and now a3dc92600.

This patch is intertwined with the tablespace patch: not only will it get
rebase conflict, but will also need to test the functionality of
CLUSTER (TABLESPACE a) partitioned_table;

--
Justin

On Mon, Jan 18, 2021 at 12:34:59PM -0600, Justin Pryzby wrote:

On Sat, Nov 28, 2020 at 08:03:02PM -0600, Justin Pryzby wrote:

On Sun, Nov 15, 2020 at 07:53:35PM -0600, Justin Pryzby wrote:

On Wed, Nov 04, 2020 at 08:23:56PM -0600, Justin Pryzby wrote:

On Tue, Oct 27, 2020 at 07:33:12PM -0500, Justin Pryzby wrote:

I'm attaching a counter-proposal to your catalog change, which preserves
indisclustered on children of clustered, partitioned indexes, and invalidates
indisclustered when attaching unclustered indexes.

..and now propagates CLUSTER ON to child indexes.

I left this as separate patches to show what I mean and what's new while we
discuss it.

This fixes some omissions in the previous patch and error in its test cases.

CLUSTER ON recurses to children, since I think a clustered parent index means
that all its child indexes are clustered. "SET WITHOUT CLUSTER" doesn't have
to recurse to children, but I did it like that for consistency and it avoids
the need to special case InvalidOid.

The previous patch failed pg_upgrade when restoring a clustered, parent index,
since it's marked INVALID until indexes have been built on all child tables, so
CLUSTER ON was rejected on invalid index.

So I think CLUSTER ON needs to be a separate pg_dump object, to allow attaching
the child index (thereby making the parent "valid") to happen before SET
CLUSTER on the parent index.

Rebased on b5913f612 and now a3dc92600.

This resolves ORDER BY test failure with COLLATE "C".

--
Justin

Hi,
For v7-0002-Implement-CLUSTER-of-partitioned-table.patch:

+        * We have to build the list in a different memory context so it
will
+        * survive the cross-transaction processing
+        */
+       old_context = MemoryContextSwitchTo(cluster_context);

cluster_context is not modified within the loop. Can the memory context
switching code be moved outside the loop ?

Cheers

On Sat, Feb 6, 2021 at 6:46 AM Justin Pryzby <pryzby@telsasoft.com> wrote:

On Sat, Feb 06, 2021 at 08:45:49AM -0600, Justin Pryzby wrote:

On Mon, Jan 18, 2021 at 12:34:59PM -0600, Justin Pryzby wrote:

On Sat, Nov 28, 2020 at 08:03:02PM -0600, Justin Pryzby wrote:

On Sun, Nov 15, 2020 at 07:53:35PM -0600, Justin Pryzby wrote:

On Wed, Nov 04, 2020 at 08:23:56PM -0600, Justin Pryzby wrote:

On Tue, Oct 27, 2020 at 07:33:12PM -0500, Justin Pryzby wrote:

I'm attaching a counter-proposal to your catalog change, which preserves
indisclustered on children of clustered, partitioned indexes, and invalidates
indisclustered when attaching unclustered indexes.

..and now propagates CLUSTER ON to child indexes.

I left this as separate patches to show what I mean and what's new while we
discuss it.

This fixes some omissions in the previous patch and error in its test cases.

CLUSTER ON recurses to children, since I think a clustered parent index means
that all its child indexes are clustered. "SET WITHOUT CLUSTER" doesn't have
to recurse to children, but I did it like that for consistency and it avoids
the need to special case InvalidOid.

The previous patch failed pg_upgrade when restoring a clustered, parent index,
since it's marked INVALID until indexes have been built on all child tables, so
CLUSTER ON was rejected on invalid index.

So I think CLUSTER ON needs to be a separate pg_dump object, to allow attaching
the child index (thereby making the parent "valid") to happen before SET
CLUSTER on the parent index.

Rebased on b5913f612 and now a3dc92600.

This resolves ORDER BY test failure with COLLATE "C".

It occured to me that progress reporting should expose this.

I did this in the style of pg_stat_progress_create_index, adding columns
partitions_total and partitions_done showing the overall progress. The progress
of individual partitions is also visible in {blocks,tuples}_{done,total}.
This seems odd, but that's how the index view behaves.

--
Justin

On Wed, Feb 10, 2021 at 02:04:58PM -0600, Justin Pryzby wrote:

On Sat, Feb 06, 2021 at 08:45:49AM -0600, Justin Pryzby wrote:

On Mon, Jan 18, 2021 at 12:34:59PM -0600, Justin Pryzby wrote:

On Sat, Nov 28, 2020 at 08:03:02PM -0600, Justin Pryzby wrote:

On Sun, Nov 15, 2020 at 07:53:35PM -0600, Justin Pryzby wrote:

On Wed, Nov 04, 2020 at 08:23:56PM -0600, Justin Pryzby wrote:

On Tue, Oct 27, 2020 at 07:33:12PM -0500, Justin Pryzby wrote:

I'm attaching a counter-proposal to your catalog change, which preserves
indisclustered on children of clustered, partitioned indexes, and invalidates
indisclustered when attaching unclustered indexes.

..and now propagates CLUSTER ON to child indexes.

I left this as separate patches to show what I mean and what's new while we
discuss it.

This fixes some omissions in the previous patch and error in its test cases.

CLUSTER ON recurses to children, since I think a clustered parent index means
that all its child indexes are clustered. "SET WITHOUT CLUSTER" doesn't have
to recurse to children, but I did it like that for consistency and it avoids
the need to special case InvalidOid.

The previous patch failed pg_upgrade when restoring a clustered, parent index,
since it's marked INVALID until indexes have been built on all child tables, so
CLUSTER ON was rejected on invalid index.

So I think CLUSTER ON needs to be a separate pg_dump object, to allow attaching
the child index (thereby making the parent "valid") to happen before SET
CLUSTER on the parent index.

Rebased on b5913f612 and now a3dc92600.

This resolves ORDER BY test failure with COLLATE "C".

It occured to me that progress reporting should expose this.

I did this in the style of pg_stat_progress_create_index, adding columns
partitions_total and partitions_done showing the overall progress. The progress
of individual partitions is also visible in {blocks,tuples}_{done,total}.
This seems odd, but that's how the index view behaves.

Rebased on 8a8f4d8ede288c2a29105f4708e22ce7f3526149.

This also resolves an issue in the last patch which would've broken progress
reporting of vacuum full.

And take the suggestion to move memory context switching outside the loop.

--
Justin

@cfbot: rebased

Hi,
For v10-0002-Implement-CLUSTER-of-partitioned-table.patch :

or that an partitioned index was previously set clustered.

'an partitioned index' -> a partitioned index

+ * Return a List of tables and associated index, where each index is a

associated index -> associated indices

For cluster():
-       rel = table_open(tableOid, NoLock);
+       rel = table_open(tableOid, ShareUpdateExclusiveLock);

Considering the comment preceding cluster() (forced to acquire exclusive
locks on all the tables), maybe add a comment explaining why it is safe to
take ShareUpdateExclusiveLock.

+cluster_multiple_rels(List *rvs, int options)

I think the multiple in the method name is not needed since the relation is
in plural.

Cheers

On Fri, Apr 2, 2021 at 1:03 PM Justin Pryzby <pryzby@telsasoft.com> wrote:

I have to wonder if there really *is* a use case for CLUSTER in the
first place on regular tables, let alone on partitioned tables, which
are likely to be large and thus take a lot of time. What justifies
spending so much time on this implementation? My impression is that
CLUSTER is pretty much a fringe command nowadays, because of the access
exclusive lock required.

Does anybody actually use it?

--
Álvaro Herrera 39°49'30"S 73°17'W — https://www.EnterpriseDB.com/
"Ninguna manada de bestias tiene una voz tan horrible como la humana" (Orual)

On Tue, Jul 20, 2021 at 08:27:02PM -0400, Alvaro Herrera wrote:

I have to wonder if there really *is* a use case for CLUSTER in the
first place on regular tables, let alone on partitioned tables, which
are likely to be large and thus take a lot of time. What justifies
spending so much time on this implementation? My impression is that
CLUSTER is pretty much a fringe command nowadays, because of the access
exclusive lock required.

Does anybody actually use it?

Yeah, I am not getting really excited about doing anything here
either. I thought for some time about the interactions with
indisclustered and partitioned tables, but anything I could come up
with felt clunky.
--
Michael

On Tue, 2021-07-20 at 20:27 -0400, Alvaro Herrera wrote:

I have to wonder if there really *is* a use case for CLUSTER in the
first place on regular tables, let alone on partitioned tables, which
are likely to be large and thus take a lot of time. What justifies
spending so much time on this implementation? My impression is that
CLUSTER is pretty much a fringe command nowadays, because of the access
exclusive lock required.

Does anybody actually use it?

I see is used in the field occasionally, as it can really drastically
improve performance. But I admit is is not frequently used.

In a data warehouse, which is updated only occasionally, running
CLUSTER after an update can make a lot of sense.

I personally think that it is enough to be able to cluster the table
partiton by partition.

Yours,
Laurenz Albe

On Tue, Jul 20, 2021 at 08:27:02PM -0400, Alvaro Herrera wrote:

I have to wonder if there really *is* a use case for CLUSTER in the
first place on regular tables, let alone on partitioned tables, which
are likely to be large and thus take a lot of time.

The cluster now is done one partition at a time, so it might take a long time,
but doesn't lock the whole partition heirarchy. Same as VACUUM (since v10) and
(since v14) REINDEX.

The patch series would be simpler if partitioned indexes weren't allowed to be
marked CLUSTERED ON. Then, "USING <index>" would be required, which is a step
forward from not supporting cluster on partitioned index at all. As attached.
It's arguably true that the follow-up patches supporting indisclustered on
partitioned indexes aren't worth the trouble.

For sure CLUSTER is useful, see eg.
https://github.com/bucardo/check_postgres/issues/29

It's sometimes important that the table is clustered to allow index scan to
work well (or be chosen at all).

If a table is scanned by an index, and isn't well-clustered, then a larger
fraction (multiple) of the table will be read than what's optimal. That
requires more IO, and more cache space.

A year ago, I partitioned one of our previously-unpartitioned tables, and ended
up clustering the partitions on their partition key (and indexed column) using
\gexec. This was preferable to doing INSERT .. SELECT .. ORDER BY, which
would've made the initial process slower - maybe unjustifiably slower for some
customers. Cluster (using \gexec) was something I was able to do afterward,
for completeness, since I expect the partitions to be mostly-clustered
automatically, so it was bothering me that the existing data was unordered, and
that it might behave differently in the future.

What justifies spending so much time on this implementation?

Actually, I don't use partitioned indexes at all here, so this is not for us..
I worked on this after Adger asked about CIC on partitioned tables (for which I
have a patch in the queue). Isn't it worth supporting that (or should we
include an example about how to use format() with %I and \gexec) ?

VACUUM [FULL] has recursed into partitions since v10 (f0e44751d).
REINDEX supports partitioned tables in v14 (a6642b3ae).
Partitioned indexes exist since v11 (as you well know), so it's somewhat odd
that CLUSTER isn't supported, and seems increasingly weird as decreasing number
of DDL commands are not supported. Supporting DDL on partitioned tables
supports the idea that the physical partitions can be seen as an implementation
detail by the DBA, which I understand was the intent since v10.

You're right that I wouldn't plan to *routinely* re-cluster a partitioned
table. Rather, I'd cluster only its "recent" *partitions*, and leave the old
ones alone. Or cluster the partitions, a single time, once they're no longer
recent. I don't think the feature is marginal just because I don't use it
routinely.

My impression is that CLUSTER is pretty much a fringe command nowadays,
because of the access exclusive lock required.

A step forward would be to integrate something like pg_repack/reorg/squeeze.
I used pg_repack --index until v12 got REINDEX CONCURRENTLY. The goal there
was to improve index scans on some large, append-only partitions where the
planner gave an index scan, but performance was poor (now, we use BRIN so it
works well without reindex). I tested that this would still be an issue by
creating a non-brin index for a single day's table (even with v13 deduplication
and v12 TID tiebreak).

As I see it, support for partitioned cluster is orthogonal to an
online/concurrent cluster, which is a job for another patch.

--
Justin

On Sun, 12 Sept 2021 at 22:10, Justin Pryzby <pryzby@telsasoft.com> wrote:

On Tue, Jul 20, 2021 at 08:27:02PM -0400, Alvaro Herrera wrote:

I have to wonder if there really *is* a use case for CLUSTER in the
first place on regular tables, let alone on partitioned tables, which
are likely to be large and thus take a lot of time.

The cluster now is done one partition at a time, so it might take a long time,
but doesn't lock the whole partition heirarchy. Same as VACUUM (since v10) and
(since v14) REINDEX.

Note: The following review is based on the assumption that this v11
revision was meant to contain only one patch. I put this up as a note,
because it seemed quite limited when compared to earlier versions of
the patchset.

I noticed that you store the result of find_all_inheritors(...,
NoLock) in get_tables_to_cluster_partitioned, without taking care of
potential concurrent partition hierarchy changes that the comment on
find_all_inheritors warns against or documenting why it is safe, which
sounds dangerous in case someone wants to adapt the code. One problem
I can think of is that only storing reloid and indoid is not
necessarily safe, as they might be reused by drop+create table running
parallel to the CLUSTER command.

Apart from that, I think it would be useful (though not strictly
necessary for this patch) if you could adapt the current CLUSTER
progress reporting to report the progress for the whole partition
hierarchy, instead of a new progress report for each relation, as was
the case in earlier versions of the patchset.

The v11 patch seems quite incomplete when compared to previous
discussions, or at the very least is very limited (no ALTER TABLE
clustering commands for partitioned tables, but `CLUSTER ptable USING
pindex` is supported). If v11 is the new proposed direction for ptable
clustering, could you also document these limitations in the
cluster.sgml and alter_table.sgml docs?

[ v11-0001-Implement-CLUSTER-of-partitioned-table.patch ]

diff --git a/src/test/regress/expected/cluster.out b/src/test/regress/expected/cluster.out
...
+ALTER TABLE clstrpart SET WITHOUT CLUSTER;
+ERROR:  cannot mark index clustered in partitioned table

This error message does not seem to match my expectation as a user: I
am not trying to mark an index as clustered, and for a normal table
"SET WITHOUT CLUSTER" does not fail for unclustered tables. I think
that behaviour of normal unclustered tables should be shared here as
well. At the very least, the error message should be changed.

ALTER TABLE clstrpart CLUSTER ON clstrpart_idx;
ERROR: cannot mark index clustered in partitioned table

A "HINT: use the CLUSTER command to cluster partitioned tables" (or
equivalent) should be added if we decide to keep the clustering APIs
of ALTER TABLE disabled for partitioned tables, as CLUSTER is now
implemented for partitioned tables.

-DROP TABLE clstrpart;

I believe that this cleanup should not be fully removed, but moved to
before '-- Test CLUSTER with external tuplesorting', as the table is
not used after that line.

Kind regards,

Matthias van de Meent

On Thu, Sep 23, 2021 at 08:18:41PM +0200, Matthias van de Meent wrote:

On Sun, 12 Sept 2021 at 22:10, Justin Pryzby <pryzby@telsasoft.com> wrote:

On Tue, Jul 20, 2021 at 08:27:02PM -0400, Alvaro Herrera wrote:

I have to wonder if there really *is* a use case for CLUSTER in the
first place on regular tables, let alone on partitioned tables, which
are likely to be large and thus take a lot of time.

The cluster now is done one partition at a time, so it might take a long time,
but doesn't lock the whole partition heirarchy. Same as VACUUM (since v10) and
(since v14) REINDEX.

Note: The following review is based on the assumption that this v11
revision was meant to contain only one patch. I put this up as a note,
because it seemed quite limited when compared to earlier versions of
the patchset.

Alvaro's critique was that the patchset was too complicated for what was
claimed to be a marginal feature. My response was to rearrange the patchset to
its minimal form, supporting CLUSTER without marking the index as clustered.

I noticed that you store the result of find_all_inheritors(...,
NoLock) in get_tables_to_cluster_partitioned, without taking care of
potential concurrent partition hierarchy changes that the comment on
find_all_inheritors warns against or documenting why it is safe, which
sounds dangerous in case someone wants to adapt the code. One problem
I can think of is that only storing reloid and indoid is not
necessarily safe, as they might be reused by drop+create table running
parallel to the CLUSTER command.

The parallel code in vacuum is expand_vacuum_rel(), which is where the
corresponding things happens for vacuum full. This patch is to make cluster()
do all the same stuff before calling cluster_rel().

What VACUUM tries to do is to avoid erroring if a partition is dropped while
cluster is running. cluster_rel() does the same thing by calling
cluster_multiple_rels() ,which uses CLUOPT_RECHECK.

If the OIDs wrapped around, I think existing vacuum could accidentally process
a new table with the same OID as a dropped partition. I think cluster would
*normally* catch that case and error in check_index_is_clusterable():
| Check that index is in fact an index on the given relation

Arguably VACUUM FULL could call cluster() (not cluster_rel()) and pass the
partitioned table rather than first expanding it. But non-full vacuum needs to
expand partitioned tables anyway.

Apart from that, I think it would be useful (though not strictly
necessary for this patch) if you could adapt the current CLUSTER
progress reporting to report the progress for the whole partition
hierarchy, instead of a new progress report for each relation, as was
the case in earlier versions of the patchset.

Yea, but this is already true for VACUUM FULL (which uses CLUSTER and supports
partitioned tables since v10) and REINDEX.
See also /messages/by-id/20210216064214.GI28165@telsasoft.com

My goal is to present a minimal patch and avoid any nonessential complexity.

The v11 patch seems quite incomplete when compared to previous
discussions, or at the very least is very limited (no ALTER TABLE
clustering commands for partitioned tables, but `CLUSTER ptable USING
pindex` is supported). If v11 is the new proposed direction for ptable
clustering, could you also document these limitations in the
cluster.sgml and alter_table.sgml docs?

You said it's less complete, but it's is due to deliberate reduction in scope.
cluster.sgml says:
+    Clustering a partitioned table clusters each of its partitions using the
+    partition of the specified partitioned index (which must be specified).

The ALTER restriction hasn't changed, so I didn't touch the documentation.

I am still curious myself to know if this is the direction the patch should
move.

[ v11-0001-Implement-CLUSTER-of-partitioned-table.patch ]

diff --git a/src/test/regress/expected/cluster.out b/src/test/regress/expected/cluster.out
...
+ALTER TABLE clstrpart SET WITHOUT CLUSTER;
+ERROR:  cannot mark index clustered in partitioned table

This error message does not seem to match my expectation as a user: I
am not trying to mark an index as clustered, and for a normal table
"SET WITHOUT CLUSTER" does not fail for unclustered tables. I think
that behaviour of normal unclustered tables should be shared here as
well. At the very least, the error message should be changed.

This is the pre-existing behavior.

-DROP TABLE clstrpart;

I believe that this cleanup should not be fully removed, but moved to
before '-- Test CLUSTER with external tuplesorting', as the table is
not used after that line.

You're right - this was from when the patchset handled CLUSTER ON.
Leaving the index allows testing in pg_dump - a large part of the complexity of
the elided patches is to handle restoring a partitioned index, without
violating the rule that partitions of an clustered index must also be
clustered. I adjusted this in my local branch.

Thanks for looking. I'm going to see about updating comments based on
corresponding parts of vacuum and on this message itself.

--
Justin

On Thu, Sep 23, 2021 at 06:56:26PM -0500, Justin Pryzby wrote:

On Thu, Sep 23, 2021 at 08:18:41PM +0200, Matthias van de Meent wrote:

Note: The following review is based on the assumption that this v11
revision was meant to contain only one patch. I put this up as a note,
because it seemed quite limited when compared to earlier versions of
the patchset.

Alvaro's critique was that the patchset was too complicated for what was
claimed to be a marginal feature. My response was to rearrange the patchset to
its minimal form, supporting CLUSTER without marking the index as clustered.

My goal is to present a minimal patch and avoid any nonessential complexity.

FWIW, my opinion on the matter is similar to Alvaro's, and an extra
read of the patch gives me the same impression. Let's see if others
have an opinion on the matter.

Thanks for looking. I'm going to see about updating comments based on
corresponding parts of vacuum and on this message itself.

It doesn't feel right to just discard the patch at this stage, and it
needs an update, so I have moved it to the next CF for now, waiting on
author. If this does not really move on, my suggestion is to discard
the patch at the end of next CF, aka 2022-01.
--
Michael

On Fri, Dec 03, 2021 at 10:16:24AM +0900, Michael Paquier wrote:

On Thu, Sep 23, 2021 at 06:56:26PM -0500, Justin Pryzby wrote:

On Thu, Sep 23, 2021 at 08:18:41PM +0200, Matthias van de Meent wrote:

Note: The following review is based on the assumption that this v11
revision was meant to contain only one patch. I put this up as a note,
because it seemed quite limited when compared to earlier versions of
the patchset.

Alvaro's critique was that the patchset was too complicated for what was
claimed to be a marginal feature. My response was to rearrange the patchset to
its minimal form, supporting CLUSTER without marking the index as clustered.

My goal is to present a minimal patch and avoid any nonessential complexity.

FWIW, my opinion on the matter is similar to Alvaro's, and an extra
read of the patch gives me the same impression. Let's see if others
have an opinion on the matter.

You and Alvaro thought the patch was too complicated for its value, so I
reduced the scope to its essential form.

CLUSTER was claimed to be of marginal utility, since the table is locked.
But locking one partition at a time would be less disruptive than locking an
equivalent non-partitioned table.

There's only about a dozen, other remaining restrictions/limitations on
partitioned tables, (AMs, triggers, identity, generated, exclusion, CIC,
FREEZE).

Since it's supported to VACUUM (including VACUUM FULL) and REINDEX a
partitioned table, I'm still suprised there's much hesitation to support
CLUSTER (which is used by vacuum full).

Thanks for looking. I'm going to see about updating comments based on
corresponding parts of vacuum and on this message itself.

It doesn't feel right to just discard the patch at this stage, and it
needs an update, so I have moved it to the next CF for now, waiting on
author. If this does not really move on, my suggestion is to discard
the patch at the end of next CF, aka 2022-01.

This includes minor updates based on Mathias review (commit message and test
case).

--
Justin

On Tue, Jul 20, 2021 at 08:27:02PM -0400, Alvaro Herrera wrote:

I have to wonder if there really *is* a use case for CLUSTER in the
first place on regular tables, let alone on partitioned tables, which
are likely to be large and thus take a lot of time. What justifies
spending so much time on this implementation? My impression is that
CLUSTER is pretty much a fringe command nowadays, because of the access
exclusive lock required.

Does anybody actually use it?

I hope that Alvaro will comment on the simplified patches. If multiple people
think the patch isn't worth it, feel free to close it. But I don't see how
complexity could be the reason.

--
Justin

On 2022-Feb-23, Justin Pryzby wrote:

I hope that Alvaro will comment on the simplified patches. If multiple people
think the patch isn't worth it, feel free to close it. But I don't see how
complexity could be the reason.

I gave your patch a look and it seems a reasonable thing to do. Maybe
not terribly useful in most cases, but there may be some cases for which
it is. I found some part of it a bit repetitive, so I moved things
around a bit. What do think about this?

--
Álvaro Herrera 48°01'N 7°57'E — https://www.EnterpriseDB.com/
"After a quick R of TFM, all I can say is HOLY CR** THAT IS COOL! PostgreSQL was
amazing when I first started using it at 7.2, and I'm continually astounded by
learning new features and techniques made available by the continuing work of
the development team."
Berend Tober, http://archives.postgresql.org/pgsql-hackers/2007-08/msg01009.php

I realized after posting that we used to allow clustering toast tables,
but after my changes we no longer do. (Justin's version had a
RELKIND_HAS_STORAGE test here instead, which seemed a little too lax.) I
don't know why we allowed it and I don't know of anyone who has ever
used that feature and we don't have any test coverage for it, but I
don't have any reason to explicitly disallow it either. So I propose to
continue to allow it:

From 05ba6124422fb7c2fd19575e905e444ba3eef1e5 Mon Sep 17 00:00:00 2001

From: Alvaro Herrera <alvherre@alvh.no-ip.org>
Date: Thu, 31 Mar 2022 12:49:57 +0200
Subject: [PATCH] allow to cluster toast tables

---
src/backend/commands/cluster.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/backend/commands/cluster.c b/src/backend/commands/cluster.c
index 8417cbdb67..b391d7c434 100644
--- a/src/backend/commands/cluster.c
+++ b/src/backend/commands/cluster.c
@@ -451,7 +451,8 @@ cluster_rel(Oid tableOid, Oid indexOid, ClusterParams *params)
 	}

 	Assert(OldHeap->rd_rel->relkind == RELKIND_RELATION ||
-		   OldHeap->rd_rel->relkind == RELKIND_MATVIEW);
+		   OldHeap->rd_rel->relkind == RELKIND_MATVIEW ||
+		   OldHeap->rd_rel->relkind == RELKIND_TOASTVALUE);

/*
* All predicate locks on the tuples or pages are about to be made
--
2.30.2

--
Álvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/
"Java is clearly an example of money oriented programming" (A. Stepanov)

On Thu, Mar 31, 2022 at 6:54 AM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:

I realized after posting that we used to allow clustering toast tables,
but after my changes we no longer do. (Justin's version had a
RELKIND_HAS_STORAGE test here instead, which seemed a little too lax.) I
don't know why we allowed it and I don't know of anyone who has ever
used that feature and we don't have any test coverage for it, but I
don't have any reason to explicitly disallow it either. So I propose to
continue to allow it:

I think that's probably a good decision. It's certainly useful to have
a way to force a rewrite of a TOAST table, although a lot of people
who would benefit from that operation probably don't know that they
need it, or don't know that they need just that, and end up rewriting
both the main table and the TOAST table. Whether it's useful to be
able to run CLUSTER specifically rather than VACUUM FULL on the TOAST
table is less clear, but I don't think we're likely to save anything
by forbidding it. Maybe we should consider adding a test, though.

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

On Wed, Mar 30, 2022 at 10:51:43PM +0200, Alvaro Herrera wrote:

On 2022-Feb-23, Justin Pryzby wrote:

I hope that Alvaro will comment on the simplified patches. If multiple people
think the patch isn't worth it, feel free to close it. But I don't see how
complexity could be the reason.

I gave your patch a look and it seems a reasonable thing to do. Maybe
not terribly useful in most cases, but there may be some cases for which
it is. I found some part of it a bit repetitive, so I moved things
around a bit. What do think about this?

Thanks for looking at it.

The changes to finish_heap_swap() and get_tables_to_cluster() are superfluous,
right ?

I think this comment is worth preserving (it'd be okay if it lived in the
commit message).
- * Expand partitioned relations for CLUSTER (the corresponding
- * thing for VACUUM FULL happens in and around expand_vacuum_rel()

+ if (rel != NULL) In this case, maybe it should Assert() that it's
relkind=p (mostly for purposes of self-documentation).

+ partition of the specified partitioned index (which must be specified).
This is my own language, but now seems repetitive. I think the parenthetical
part should be a separate sentance: "For partitioned indexes, the index may not
be omitted.".

Otherwise looks ok.

diff --git a/doc/src/sgml/ref/cluster.sgml b/doc/src/sgml/ref/cluster.sgml
index b3463ae5c46..fbc090cd0b0 100644
--- a/doc/src/sgml/ref/cluster.sgml
+++ b/doc/src/sgml/ref/cluster.sgml
@@ -199,7 +199,8 @@ CLUSTER [VERBOSE]

    <para>
     Clustering a partitioned table clusters each of its partitions using the
-    partition of the specified partitioned index (which must be specified).
+    partition of the specified partitioned index.  When clustering a
+    partitioned table, the index may not be omitted.
    </para>

  </refsect1>
diff --git a/src/backend/commands/cluster.c b/src/backend/commands/cluster.c
index 8417cbdb67f..412147f05bc 100644
--- a/src/backend/commands/cluster.c
+++ b/src/backend/commands/cluster.c
@@ -231,6 +231,7 @@ cluster(ParseState *pstate, ClusterStmt *stmt, bool isTopLevel)
 	params.options |= CLUOPT_RECHECK;
 	if (rel != NULL)
 	{
+		Assert (rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
 		check_index_is_clusterable(rel, indexOid, true, AccessShareLock);
 		rtcs = get_tables_to_cluster_partitioned(cluster_context, indexOid);

@@ -451,6 +452,7 @@ cluster_rel(Oid tableOid, Oid indexOid, ClusterParams *params)
}

Assert(OldHeap->rd_rel->relkind == RELKIND_RELATION ||
+ OldHeap->rd_rel->relkind == RELKIND_TOASTVALUE ||
OldHeap->rd_rel->relkind == RELKIND_MATVIEW);

 	/*
diff --git a/src/test/regress/expected/cluster.out b/src/test/regress/expected/cluster.out
index 3f2758d13f6..6cf18c8d321 100644
--- a/src/test/regress/expected/cluster.out
+++ b/src/test/regress/expected/cluster.out
@@ -305,6 +305,8 @@ WHERE pg_class.oid=indexrelid
 ---------
 (0 rows)

+-- Verify that toast is clusterable
+CLUSTER pg_toast.pg_toast_826 USING pg_toast_826_index;
 -- Verify that clustering all tables does in fact cluster the right ones
 CREATE USER regress_clstr_user;
 CREATE TABLE clstr_1 (a INT PRIMARY KEY);
diff --git a/src/test/regress/sql/cluster.sql b/src/test/regress/sql/cluster.sql
index 74118993a82..ae27c35f65d 100644
--- a/src/test/regress/sql/cluster.sql
+++ b/src/test/regress/sql/cluster.sql
@@ -104,6 +104,9 @@ WHERE pg_class.oid=indexrelid
 	AND pg_class_2.relname = 'clstr_tst'
 	AND indisclustered;

+-- Verify that toast is clusterable
+CLUSTER pg_toast.pg_toast_826 USING pg_toast_826_index;
+
 -- Verify that clustering all tables does in fact cluster the right ones
 CREATE USER regress_clstr_user;
 CREATE TABLE clstr_1 (a INT PRIMARY KEY);

On Thu, Mar 31, 2022 at 12:54:36PM +0200, Alvaro Herrera wrote:

I realized after posting that we used to allow clustering toast tables,
but after my changes we no longer do. (Justin's version had a
RELKIND_HAS_STORAGE test here instead, which seemed a little too lax.) I
don't know why we allowed it and I don't know of anyone who has ever
used that feature and we don't have any test coverage for it, but I
don't have any reason to explicitly disallow it either. So I propose to
continue to allow it:

Good catch.

My daily vacuum script would've discovered that they're no longer supported, as
it tests for (among other things) c.relkind IN ('r','t'). That clusters tables
that have an indisclustered set and vacuums various others. (BTW, it's the
same script that discovered in 2019 that clustering on expressional indexes had
been broken by the heapam changes).

I think the response should be to add a test case, which could be 0001 or
00099.

Thanks, pushed.

--
Álvaro Herrera 48°01'N 7°57'E — https://www.EnterpriseDB.com/
"La fuerza no está en los medios físicos
sino que reside en una voluntad indomable" (Gandhi)

Small things here.

1. in VACUUM FULL we only process partitions that are owned by the
invoking user. We don't have this test in the new code. I'm not sure
why do we do that there; is it worth doing the same here?

2. We should silently skip a partition that's a foreign table, I
suppose.

3. We do mark the index on the partitions as indisclustered AFAICS (we
claim that the partitioned table's index is not marked, which is
accurate). So users doing unadorned CLUSTER afterwards will get the
partitions clustered too, once they cluster the partitioned table. If
they don't want this, they would have to ALTER TABLE to remove the
marking. How likely is that this will be a problem? Maybe documenting
this point is enough.

--
Álvaro Herrera Breisgau, Deutschland — https://www.EnterpriseDB.com/
Voy a acabar con todos los humanos / con los humanos yo acabaré
voy a acabar con todos (bis) / con todos los humanos acabaré ¡acabaré! (Bender)

On Sat, Apr 02, 2022 at 07:11:47PM +0200, Alvaro Herrera wrote:

Thanks, pushed.

Thank you for revisiting it, and thanks to Zhihong Yu for earlier review.

I'll look into your outstanding questions later this week.

--
Justin

On Sat, Apr 02, 2022 at 07:21:11PM +0200, Alvaro Herrera wrote:

Small things here.

1. in VACUUM FULL we only process partitions that are owned by the
invoking user. We don't have this test in the new code. I'm not sure
why do we do that there; is it worth doing the same here?

That dates to a556549d7 (see also cbe24a6dd8 for an earlier commit in CLUSTER
itself). The reason was to avoid blocking if an unprivileged user runs VACUUM
FULL which would try to lock things (including shared catalogs) before checking
if they have permission to vacuum them. That commit also initially checks the
owner of the partitioned table, and then re-checking owner of partitions later
on.

A similar issue exists here. But 1) catalog tables are not partitioned, and,
2) ownership of a partitioned table is checked immediately. So the problem can
only occur if a user who owns a partitioned table but doesn't own all its
partitions tries to cluster it, and it blocks behind another session. Fixing
this is probably a good idea, but seems improbable that it would avoid a DOS.

2. We should silently skip a partition that's a foreign table, I
suppose.

I think it's not needed, since the loop over index children doesn't see a child
index on the foreign table. ?

3. We do mark the index on the partitions as indisclustered AFAICS (we
claim that the partitioned table's index is not marked, which is
accurate). So users doing unadorned CLUSTER afterwards will get the
partitions clustered too, once they cluster the partitioned table. If
they don't want this, they would have to ALTER TABLE to remove the
marking. How likely is that this will be a problem? Maybe documenting
this point is enough.

It seems at least as likely that someone would *want* the partitions to be
marked clustered as that someone would want them to be unchanged.

The cluster mark accurately reflects having been clustered. It seems unlikely
that a user would want something else to be clustered later by "cluster;".
Since clustering on a partitioned table wasn't supported before, nothing weird
will happen to someone who upgrades to v15 unless they elect to use the new
feature. As this seems to be POLA, it doesn't even need to be documented. ?

On Mon, Apr 11, 2022 at 7:06 AM Justin Pryzby <pryzby@telsasoft.com> wrote:

On Sat, Apr 02, 2022 at 07:21:11PM +0200, Alvaro Herrera wrote:

Small things here.

1. in VACUUM FULL we only process partitions that are owned by the
invoking user. We don't have this test in the new code. I'm not sure
why do we do that there; is it worth doing the same here?

That dates to a556549d7 (see also cbe24a6dd8 for an earlier commit in
CLUSTER
itself). The reason was to avoid blocking if an unprivileged user runs
VACUUM
FULL which would try to lock things (including shared catalogs) before
checking
if they have permission to vacuum them. That commit also initially checks
the
owner of the partitioned table, and then re-checking owner of partitions
later
on.

A similar issue exists here. But 1) catalog tables are not partitioned,
and,
2) ownership of a partitioned table is checked immediately. So the
problem can
only occur if a user who owns a partitioned table but doesn't own all its
partitions tries to cluster it, and it blocks behind another session.
Fixing
this is probably a good idea, but seems improbable that it would avoid a
DOS.

2. We should silently skip a partition that's a foreign table, I
suppose.

I think it's not needed, since the loop over index children doesn't see a
child
index on the foreign table. ?

3. We do mark the index on the partitions as indisclustered AFAICS (we
claim that the partitioned table's index is not marked, which is
accurate). So users doing unadorned CLUSTER afterwards will get the
partitions clustered too, once they cluster the partitioned table. If
they don't want this, they would have to ALTER TABLE to remove the
marking. How likely is that this will be a problem? Maybe documenting
this point is enough.

It seems at least as likely that someone would *want* the partitions to be
marked clustered as that someone would want them to be unchanged.

The cluster mark accurately reflects having been clustered. It seems
unlikely
that a user would want something else to be clustered later by "cluster;".
Since clustering on a partitioned table wasn't supported before, nothing
weird
will happen to someone who upgrades to v15 unless they elect to use the new
feature. As this seems to be POLA, it doesn't even need to be
documented. ?

Hi,
For v13-0002-cluster-early-ownership-check-of-partitions.patch :

only for it to fails ownership check anyway

to fails -> to fail

Cheers

On Mon, Apr 11, 2022 at 09:06:09AM -0500, Justin Pryzby wrote:

On Sat, Apr 02, 2022 at 07:21:11PM +0200, Alvaro Herrera wrote:

1. in VACUUM FULL we only process partitions that are owned by the
invoking user. We don't have this test in the new code. I'm not sure
why do we do that there; is it worth doing the same here?

I think that adding a test is a good idea for such things. Perhaps we
could have an isolation test, but what Justin is proposing seems good
enough to me for this goal.

That dates to a556549d7 (see also cbe24a6dd8 for an earlier commit in CLUSTER
itself). The reason was to avoid blocking if an unprivileged user runs VACUUM
FULL which would try to lock things (including shared catalogs) before checking
if they have permission to vacuum them. That commit also initially checks the
owner of the partitioned table, and then re-checking owner of partitions later
on.

A similar issue exists here. But 1) catalog tables are not partitioned, and,
2) ownership of a partitioned table is checked immediately. So the problem can
only occur if a user who owns a partitioned table but doesn't own all its
partitions tries to cluster it, and it blocks behind another session. Fixing
this is probably a good idea, but seems improbable that it would avoid a DOS.

Catalogs are out of the picture as you say and I would not worry about
them becoming somewhat partitioned even in the far future. Are you
saying that it is possible for a user kicking a CLUSTER command on a
partitioned table who has no ownership on some of the partitions to
do some blocking table_open() calls if the permission check is not
done in get_tables_to_cluster_partitioned()? Hence, this user could
block the access to such partitions? I am not sure that we need to
add any new ownership checks here as CLUOPT_RECHECK gets added to the
parameters in cluster() before calling cluster_multiple_rels(), then
we do a mix of try_relation_open() with a skip when we are not the
owner anymore. So this logic looks sound to me. In short, you don't
need this extra check, and the test proposed in 0002 keeps the same
behavior.

2. We should silently skip a partition that's a foreign table, I
suppose.

I think it's not needed, since the loop over index children doesn't see a child
index on the foreign table?

Hmm. That may be a sign to add an assertion, at least, or something
based on RELKIND_HAS_STORAGE().

I was wondering what 0001 was doing here as that's a separate issue,
but it looked fine so I have applied it.

+       /* Use a permanent memory context for the result list */
+       old_context = MemoryContextSwitchTo(cluster_context);
+
        rtc = (RelToCluster *) palloc(sizeof(RelToCluster));

Independently of the extra ownership check, the memory context
manipulation has to be fixed and the code shoudl switch to
RelToCluster only when saving an item.

+CREATE ROLE ptnowner;
Roles that are created in the regression tests need to be prefixed
with "regress_", or some buildfarm members will complain. FWIW, I
enforce -DENFORCE_REGRESSION_TEST_NAME_RESTRICTIONS in all my dev
builds.

I have added an open item for now, but the whole looks
straight-forward to me.
--
Michael

On Wed, Apr 13, 2022 at 03:50:15PM +0900, Michael Paquier wrote:

That dates to a556549d7 (see also cbe24a6dd8 for an earlier commit in CLUSTER
itself). The reason was to avoid blocking if an unprivileged user runs VACUUM
FULL which would try to lock things (including shared catalogs) before checking
if they have permission to vacuum them. That commit also initially checks the
owner of the partitioned table, and then re-checking owner of partitions later
on.

A similar issue exists here. But 1) catalog tables are not partitioned, and,
2) ownership of a partitioned table is checked immediately. So the problem can
only occur if a user who owns a partitioned table but doesn't own all its
partitions tries to cluster it, and it blocks behind another session. Fixing
this is probably a good idea, but seems improbable that it would avoid a DOS.

Catalogs are out of the picture as you say and I would not worry about
them becoming somewhat partitioned even in the far future. Are you
saying that it is possible for a user kicking a CLUSTER command on a
partitioned table who has no ownership on some of the partitions to
do some blocking table_open() calls if the permission check is not
done in get_tables_to_cluster_partitioned()? Hence, this user could
block the access to such partitions? I am not sure that we need to
add any new ownership checks here as CLUOPT_RECHECK gets added to the
parameters in cluster() before calling cluster_multiple_rels(), then
we do a mix of try_relation_open() with a skip when we are not the
owner anymore. So this logic looks sound to me. In short, you don't
need this extra check, and the test proposed in 0002 keeps the same
behavior.

Are you sure? The ownership re-check in cluster_rel() occurs after acquiring
locks.

s1:
postgres=# CREATE TABLE p(i int) PARTITION BY LIST (i);
postgres=# CREATE TABLE p1 PARTITION OF p FOR VALUES IN (1);
postgres=# CREATE TABLE p2 PARTITION OF p FOR VALUES IN (2);
postgres=# CREATE INDEX ON p (i);
postgres=# CREATE ROLE po WITH LOGIN;
postgres=# ALTER TABLE p OWNER TO po;
postgres=# begin; SELECT FROM p1;

s2:
postgres=> SET client_min_messages =debug;
postgres=> CLUSTER VERBOSE p USING p_i_idx ;
LOG: process 26058 still waiting for AccessExclusiveLock on relation 39577 of database 5 after 1000.105 ms
postgres=> SELECT 39577::regclass;
regclass | p1

On Wed, Apr 13, 2022 at 05:52:14AM -0500, Justin Pryzby wrote:

Are you sure? The ownership re-check in cluster_rel() occurs after acquiring
locks.

Yep, you are right. However, the SQL test does not check for this
blocking scenario. In short, removing the new ACL check in
get_tables_to_cluster_partitioned() makes the test behave the same
way. Could you implement an isolation check to make sure that the
difference is visible? The SQL check looks useful in itself, either
way.
--
Michael

Thanks for the patch -- I have pushed it now, with some wording changes
and renaming the role to regress_* to avoid buildfarm's ire.

Michaël in addition proposes an isolation test. I'm not sure; is it
worth the additional test run time? It doesn't seem a critical issue.
But if anybody feels like contributing one, step right ahead.

--
Álvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/

On Thu, Apr 14, 2022 at 10:37:06PM +0200, Alvaro Herrera wrote:

Thanks for the patch -- I have pushed it now, with some wording changes
and renaming the role to regress_* to avoid buildfarm's ire.

Cool, thanks.

Michaël in addition proposes an isolation test. I'm not sure; is it
worth the additional test run time? It doesn't seem a critical issue.
But if anybody feels like contributing one, step right ahead.

Well, I am a bit annoyed that we don't actually check that a CLUSTER
command does not block when doing a CLUSTER on a partitioned table
while a lock is held on one of its partitions. So, attached is a
proposal of patch to improve the test coverage in this area. While on
it, I have added a test with a normal table. You can see the
difference once you remove the ACL check added recently in
get_tables_to_cluster_partitioned(). What do you think?
--
Michael

On Sat, Apr 16, 2022 at 08:58:50PM +0900, Michael Paquier wrote:

Well, I am a bit annoyed that we don't actually check that a CLUSTER
command does not block when doing a CLUSTER on a partitioned table
while a lock is held on one of its partitions. So, attached is a
proposal of patch to improve the test coverage in this area. While on
it, I have added a test with a normal table. You can see the
difference once you remove the ACL check added recently in
get_tables_to_cluster_partitioned(). What do you think?

This was the last reason why this was listed as an open item, so,
hearing nothing, I have applied this patch to add those extra tests,
and switched the item as fixed.
--
Michael

On 2022-Apr-26, Michael Paquier wrote:

On Sat, Apr 16, 2022 at 08:58:50PM +0900, Michael Paquier wrote:

Well, I am a bit annoyed that we don't actually check that a CLUSTER
command does not block when doing a CLUSTER on a partitioned table
while a lock is held on one of its partitions. So, attached is a
proposal of patch to improve the test coverage in this area.

This was the last reason why this was listed as an open item, so,
hearing nothing, I have applied this patch to add those extra tests,
and switched the item as fixed.

Thank you!

--
Álvaro Herrera 48°01'N 7°57'E — https://www.EnterpriseDB.com/