Inconsistent error message wording for REINDEX CONCURRENTLY

On Thu, May 02, 2019 at 10:06:42AM -0400, Tom Lane wrote:

In view of the REINDEX-on-pg_class kerfuffle that we're currently
sorting through, I was very glad to see that the concurrent reindex
code doesn't even try:

regression=# reindex index concurrently pg_class_oid_index;
psql: ERROR: concurrent reindex is not supported for catalog relations
regression=# reindex table concurrently pg_class;
psql: ERROR: concurrent index creation on system catalog tables is not supported

It'd be nice though if those error messages gave the impression of having
been written on the same planet.

We could do a larger brush-up of error messages in this area, as these
are full sentences which is not a style allowed, no? The second error
message can be used as well by both CREATE INDEX CONCURRENTLY and
REINDEX CONCURRENTLY, but not the first one, so the first one needs to
be more generic than the second one. How about the following changes
for at least these two?
"cannot use REINDEX CONCURRENTLY on system catalogs"
"cannot create index on system catalog concurrently"

Then we have some other messages in index.c which could be cleaned
up.. For example at the beginning of index_constraint_create(), there
are two them, but there is much more which could be improved. Do you
think this is worth having a look and fixing?
--
Michael

Michael Paquier <michael@paquier.xyz> writes:

On Thu, May 02, 2019 at 10:06:42AM -0400, Tom Lane wrote:

regression=# reindex index concurrently pg_class_oid_index;
psql: ERROR: concurrent reindex is not supported for catalog relations
regression=# reindex table concurrently pg_class;
psql: ERROR: concurrent index creation on system catalog tables is not supported

It'd be nice though if those error messages gave the impression of having
been written on the same planet.

We could do a larger brush-up of error messages in this area, as these
are full sentences which is not a style allowed, no?

I wouldn't object to either one in isolation, it's the inconsistency
that irks me.

How about the following changes
for at least these two?
"cannot use REINDEX CONCURRENTLY on system catalogs"
"cannot create index on system catalog concurrently"

I'd suggest something like "cannot reindex a system catalog concurrently"
for both cases. The "cannot create index" wording doesn't seem to me to
be very relevant, because if you try that you'll get

regression=# create index on pg_class(relchecks);
psql: ERROR: permission denied: "pg_class" is a system catalog

Then we have some other messages in index.c which could be cleaned
up.. For example at the beginning of index_constraint_create(), there
are two them, but there is much more which could be improved. Do you
think this is worth having a look and fixing?

I'm not excited about rewording longstanding errors. These two are
new though (aren't they?)

regards, tom lane

On Sat, May 04, 2019 at 11:00:11AM -0400, Tom Lane wrote:

I'm not excited about rewording longstanding errors. These two are
new though (aren't they?)

The message you are referring to in index_create() has been introduced
as of e093dcdd with the introduction of CREATE INDEX CONCURRENTLY, and
it can be perfectly hit without REINDEX:
=# show allow_system_table_mods;
allow_system_table_mods
-------------------------
on
(1 row)
=# create index CONCURRENTLY popo on pg_class (relname);
ERROR: 0A000: concurrent index creation on system catalog tables is
not supported
LOCATION: index_create, index.c:830

So I don't agree with switching the existing error message in
index_create(). What we could do instead is to add a REINDEX-specific
error in ReindexRelationConcurrently() as done for index relkinds,
using your proposed wording.

What do you think about something like the attached then? HEAD does
not check after system indexes with REINDEX INDEX CONCURRENTLY, and I
have moved all the catalog-related tests to reindex_catalog.sql.
--
Michael

Michael Paquier <michael@paquier.xyz> writes:

The message you are referring to in index_create() has been introduced
as of e093dcdd with the introduction of CREATE INDEX CONCURRENTLY, and
it can be perfectly hit without REINDEX:
=# show allow_system_table_mods;
allow_system_table_mods
-------------------------
on
(1 row)

Oh, yeah, if you do that you can get to it.

What do you think about something like the attached then? HEAD does
not check after system indexes with REINDEX INDEX CONCURRENTLY, and I
have moved all the catalog-related tests to reindex_catalog.sql.

OK as far as the wording goes, but now that I look at the specific tests
that are being applied, they seem rather loony, as well as inconsistent
between the two cases. IsSystemRelation *sounds* like the right thing,
but it's not, because it forbids user-relation toast tables which seems
like a restriction we need not make. I think IsCatalogRelation is the
test we actually want there. In the other place, checking
IsSystemNamespace isn't even approximately the correct way to proceed,
since it fails to reject reindexing system catalogs' toast tables.
We should be doing the equivalent of IsCatalogRelation there too.
(It's a bit of a pain that catalog.c doesn't offer a function that
makes that determination from just an OID. Should we add one?
There might be other callers for it.)

I concur that we shouldn't need a separate check for relisshared,
since all shared rels should be system catalogs.

I"m not sure I'd move these error-case tests to reindex_catalog.sql ---
bear in mind that later today, that test is either going away entirely
or at least not getting run by default anymore.

regards, tom lane

I wrote:

Michael Paquier <michael@paquier.xyz> writes:

What do you think about something like the attached then? HEAD does
not check after system indexes with REINDEX INDEX CONCURRENTLY, and I
have moved all the catalog-related tests to reindex_catalog.sql.

OK as far as the wording goes, but now that I look at the specific tests
that are being applied, they seem rather loony, as well as inconsistent
between the two cases. IsSystemRelation *sounds* like the right thing,
but it's not, because it forbids user-relation toast tables which seems
like a restriction we need not make. I think IsCatalogRelation is the
test we actually want there. In the other place, checking
IsSystemNamespace isn't even approximately the correct way to proceed,
since it fails to reject reindexing system catalogs' toast tables.
We should be doing the equivalent of IsCatalogRelation there too.
(It's a bit of a pain that catalog.c doesn't offer a function that
makes that determination from just an OID. Should we add one?
There might be other callers for it.)

After looking around a bit, I propose that we invent
"IsCatalogRelationOid(Oid reloid)" (not wedded to that name), which
is a wrapper around IsCatalogClass() that does the needful syscache
lookup for you. Aside from this use-case, it could be used in
sepgsql/dml.c, which I see is also using
IsSystemNamespace(get_rel_namespace(oid)) for the wrong thing.

I'm also thinking that it'd be a good idea to rename IsSystemNamespace
to IsCatalogNamespace. The existing naming is totally confusing given
that it doesn't square with the distinction between IsSystemRelation
and IsCatalogRelation (ie that the former includes user toast tables).
There are only five external callers of it, and per this discussion
at least two of them are wrong anyway.

I was thinking about also proposing that we rename IsSystemRelation
to IsCatalogOrToastRelation (likewise for IsSystemClass), which would
be clearer as to its semantics. However, after looking through the
code, it seems that 90% of the callers are using those functions to
decide whether to apply !allow_system_table_mods restrictions, and
indeed it's likely that some of the other 10% are wrong and should be
testing IsCatalogRelation/Class instead. So unless we want to rename
that GUC, I think the existing names of these functions are fine but
we should adjust their comments to explain that this is the primary
if not sole use-case. Another idea is to make IsSystemRelation/Class
be macros for IsCatalogOrToastRelation/Class, with the intention that
we use the former names specifically for allow_system_table_mods tests
and the latter names for anything else that happens to really want
those semantics.

There's some other cleanup I want to do in catalog.c --- many of the
comments are desperately in need of copy-editing, to start with ---
but I don't think any of the rest of it would be controversial.

regards, tom lane

On Sun, May 05, 2019 at 05:45:53PM -0400, Tom Lane wrote:

In the other place, checking IsSystemNamespace isn't even
approximately the correct way to proceed, since it fails to reject
reindexing system catalogs' toast tables.

Good point. I overlooked that part. It is easy enough to have a test
which fails for a catalog index, a catalog table, a toast table on a
system catalog and a toast index on a system catalog. However I don't
see a way to test directly that a toast relation or index on a
non-catalog relation works as we cannot run REINDEX CONCURRENTLY
within a function, and it is not possible to save the toast relation
name as a psql variable. Perhaps somebody has a trick?x

After looking around a bit, I propose that we invent
"IsCatalogRelationOid(Oid reloid)" (not wedded to that name), which
is a wrapper around IsCatalogClass() that does the needful syscache
lookup for you. Aside from this use-case, it could be used in
sepgsql/dml.c, which I see is also using
IsSystemNamespace(get_rel_namespace(oid)) for the wrong thing.

Hmmm. A wrapper on top of IsCatalogClass() implies that we would need
to open the related relation directly in the new function so as it is
possible to grab its pg_class entry. We could imply that the function
takes a ShareLock all the time, but that's not going to be true most
of the time and the recent discussions around lock upgrades stress me
a bit, and I'd rather not add new race conditions or upgrade hazards.
We should have an extra argument with the lock mode, but we have
nothing in catalog.c of that kind, and that does not feel consistent
with the current interface. At the end I have made the choice to not
reinvent the world, and just get a Relation from the parent table when
looking after an index relkind so as IsCatalogRelation() is used for
the check.

What do you think about the updated patch attached? I have removed
the tests from reindex_catalog.sql, and added more coverage into
create_index.sql.
--
Michael

I wrote:

After looking around a bit, I propose that we invent
"IsCatalogRelationOid(Oid reloid)" (not wedded to that name), which
is a wrapper around IsCatalogClass() that does the needful syscache
lookup for you. Aside from this use-case, it could be used in
sepgsql/dml.c, which I see is also using
IsSystemNamespace(get_rel_namespace(oid)) for the wrong thing.

I'm also thinking that it'd be a good idea to rename IsSystemNamespace
to IsCatalogNamespace. The existing naming is totally confusing given
that it doesn't square with the distinction between IsSystemRelation
and IsCatalogRelation (ie that the former includes user toast tables).
There are only five external callers of it, and per this discussion
at least two of them are wrong anyway.

After studying the callers of these catalog.c functions for awhile,
I realized that IsCatalogRelation/Class are really fundamentally wrong,
and have been so for a long time. The reason is that while they will
return FALSE for tables in the information_schema, they will return
TRUE for toast tables attached to the information_schema tables.
(They're toast tables, and they have OIDs below FirstNormalObjectId,
so there you have it.) This is wrong on its face: if those tables don't
need to be protected as catalogs, why should their TOAST appendages
need it? Moreover, if you drop and recreate information_schema, you'll
start getting different behavior for them, which is even sillier.

I was driven to this realization by the following very confused (and
confusing) bit in ReindexMultipleTables:

/*
* Skip system tables that index_create() would reject to index
* concurrently. XXX We need the additional check for
* FirstNormalObjectId to skip information_schema tables, because
* IsCatalogClass() here does not cover information_schema, but the
* check in index_create() will error on the TOAST tables of
* information_schema tables.
*/
if (concurrent &&
(IsCatalogClass(relid, classtuple) || relid < FirstNormalObjectId))
{

That's nothing but a hack, and the reason it's necessary is that
index_create will throw error if IsCatalogRelation is true, which
it will be for information_schema TOAST tables --- but not for their
parent tables that are being examined here.

After looking around, it seems to me that the correct definition for
IsCatalogRelation is just "is the OID less than FirstBootstrapObjectId?".
Currently we could actually restrict it to "less than
FirstGenbkiObjectId", because all the catalogs, indexes, and TOAST tables
have hand-assigned OIDs --- but perhaps someday we'll let genbki.pl
assign some of those OIDs, so I prefer the weaker constraint. In any
case, this gives us a correct separation between objects that are
traceable to the bootstrap data and those that are created by plain SQL
later in initdb.

With this, the Form_pg_class argument to IsCatalogClass becomes
vestigial. I'm tempted to get rid of that function altogether in
favor of direct calls to IsCatalogRelationOid, but haven't done so
in the attached.

Comments?

regards, tom lane

On Tue, May 07, 2019 at 05:19:38PM -0400, Tom Lane wrote:

That's nothing but a hack, and the reason it's necessary is that
index_create will throw error if IsCatalogRelation is true, which
it will be for information_schema TOAST tables --- but not for their
parent tables that are being examined here.

Oh. Good catch. That's indeed crazy now that I look closer at that.

After looking around, it seems to me that the correct definition for
IsCatalogRelation is just "is the OID less than FirstBootstrapObjectId?".
Currently we could actually restrict it to "less than
FirstGenbkiObjectId", because all the catalogs, indexes, and TOAST tables
have hand-assigned OIDs --- but perhaps someday we'll let genbki.pl
assign some of those OIDs, so I prefer the weaker constraint. In any
case, this gives us a correct separation between objects that are
traceable to the bootstrap data and those that are created by plain SQL
later in initdb.

With this, the Form_pg_class argument to IsCatalogClass becomes
vestigial. I'm tempted to get rid of that function altogether in
favor of direct calls to IsCatalogRelationOid, but haven't done so
in the attached.

I think that removing entirely IsCatalogClass() is just better as if
any extension uses this routine, then it could potentially simplify
its code because needing Form_pg_class means usually opening a
Relation, and this can be removed.

With IsCatalogClass() removed, the only dependency with Form_pg_class
comes from IsToastClass() which is not used at all except in
IsSystemClass(). Wouldn't it be better to remove entirely
IsToastClass() and switch IsSystemClass() to use a namespace OID
instead of Form_pg_class?

With your patch, ReindexRelationConcurrently() does not complain for
REINDEX TABLE CONCURRENTLY for a catalog table and would trigger the
error from index_create(), which is at the origin of this thread. The
check with IsSharedRelation() for REINDEX INDEX CONCURRENTLY is
useless and the error message generated for IsCatalogRelationOid()
still needs to be improved. Would you prefer to include those changes
in your patch? Or should I work on top of what you are proposing
(your patch does not include negative tests for toast index and
tables on catalogs either).
--
Michael

Michael Paquier <michael@paquier.xyz> writes:

On Tue, May 07, 2019 at 05:19:38PM -0400, Tom Lane wrote:

With this, the Form_pg_class argument to IsCatalogClass becomes
vestigial. I'm tempted to get rid of that function altogether in
favor of direct calls to IsCatalogRelationOid, but haven't done so
in the attached.

I think that removing entirely IsCatalogClass() is just better as if
any extension uses this routine, then it could potentially simplify
its code because needing Form_pg_class means usually opening a
Relation, and this can be removed.

Yeah, it's clearly easier to use without the extra argument.

With IsCatalogClass() removed, the only dependency with Form_pg_class
comes from IsToastClass() which is not used at all except in
IsSystemClass(). Wouldn't it be better to remove entirely
IsToastClass() and switch IsSystemClass() to use a namespace OID
instead of Form_pg_class?

Not sure. The way it's defined has the advantage of being more
independent of exactly what the implementation of the "is a toast table"
check is. Also, I looked around to see if any callers could really be
simplified if they only had to pass the table OID, and didn't find much;
almost all of them are looking at the pg_class tuple themselves, typically
to check the relkind too. So we'd not make any net savings in syscache
lookups by changing IsSystemClass's API. I'm kind of inclined to leave
it alone.

With your patch, ReindexRelationConcurrently() does not complain for
REINDEX TABLE CONCURRENTLY for a catalog table and would trigger the
error from index_create(), which is at the origin of this thread. The
check with IsSharedRelation() for REINDEX INDEX CONCURRENTLY is
useless and the error message generated for IsCatalogRelationOid()
still needs to be improved. Would you prefer to include those changes
in your patch? Or should I work on top of what you are proposing
(your patch does not include negative tests for toast index and
tables on catalogs either).

Yes, we still need to do your patch on top of this one (or really
either order would do). I think keeping them separate is good.

BTW, when I was looking at this I got dissatisfied about another
aspect of the wording of the relevant error messages: a lot of them
are like, for example

errmsg("cannot reindex concurrently this type of relation")));

While that matches the command syntax we're using, it's just horrid
English grammar. Better would be

errmsg("cannot reindex this type of relation concurrently")));

Can we change that while we're at it?

regards, tom lane

On Wed, May 08, 2019 at 08:31:54AM -0400, Tom Lane wrote:

Michael Paquier <michael@paquier.xyz> writes:

With IsCatalogClass() removed, the only dependency with Form_pg_class
comes from IsToastClass() which is not used at all except in
IsSystemClass(). Wouldn't it be better to remove entirely
IsToastClass() and switch IsSystemClass() to use a namespace OID
instead of Form_pg_class?

Not sure. The way it's defined has the advantage of being more
independent of exactly what the implementation of the "is a toast table"
check is. Also, I looked around to see if any callers could really be
simplified if they only had to pass the table OID, and didn't find much;
almost all of them are looking at the pg_class tuple themselves, typically
to check the relkind too. So we'd not make any net savings in syscache
lookups by changing IsSystemClass's API. I'm kind of inclined to leave
it alone.

Hmm. Okay. It would have been nice to remove completely the
dependency to Form_pg_class from this set of APIs, but I can see your
point.

Yes, we still need to do your patch on top of this one (or really
either order would do). I think keeping them separate is good.

Okay, glad to hear. That's what I wanted to do.

While that matches the command syntax we're using, it's just horrid
English grammar. Better would be

errmsg("cannot reindex this type of relation concurrently")));

Can we change that while we're at it?

No problem to do that. I'll brush up all that once you commit the
first piece you have come up with, and reuse the new API of catalog.c
you are introducing based on the table OID.
--
Michael

Michael Paquier <michael@paquier.xyz> writes:

No problem to do that. I'll brush up all that once you commit the
first piece you have come up with, and reuse the new API of catalog.c
you are introducing based on the table OID.

Pushed my stuff, have at it.

regards, tom lane

On Wed, May 08, 2019 at 11:28:35PM -0400, Tom Lane wrote:

Michael Paquier <michael@paquier.xyz> writes:

No problem to do that. I'll brush up all that once you commit the
first piece you have come up with, and reuse the new API of catalog.c
you are introducing based on the table OID.

Pushed my stuff, have at it.

Thanks. Attached is what I get to after scanning the error messages
in indexcmds.c and index.c. Perhaps you have more comments about it?
--
Michael

Michael Paquier <michael@paquier.xyz> writes:

On Wed, May 08, 2019 at 11:28:35PM -0400, Tom Lane wrote:

Pushed my stuff, have at it.

Thanks. Attached is what I get to after scanning the error messages
in indexcmds.c and index.c. Perhaps you have more comments about it?

LGTM, thanks.

regards, tom lane

On Thu, May 09, 2019 at 02:08:39PM -0400, Tom Lane wrote:

LGTM, thanks.

Thanks for double-checking, committed. I am closing the open item.
--
Michael

On 2019-May-09, Michael Paquier wrote:

On Wed, May 08, 2019 at 11:28:35PM -0400, Tom Lane wrote:

Michael Paquier <michael@paquier.xyz> writes:

No problem to do that. I'll brush up all that once you commit the
first piece you have come up with, and reuse the new API of catalog.c
you are introducing based on the table OID.

Pushed my stuff, have at it.

Thanks. Attached is what I get to after scanning the error messages
in indexcmds.c and index.c. Perhaps you have more comments about it?

I do :-) There are a couple of "is not supported" messages that are
annoyingly similar but different:
git grep --show-function 'reindex.*supported' -- *.c

src/backend/commands/indexcmds.c=ReindexMultipleTables(const char *objectName, ReindexObjectType objectKind,
src/backend/commands/indexcmds.c: errmsg("concurrent reindex of system catalogs is not supported")));
src/backend/commands/indexcmds.c: errmsg("concurrent reindex is not supported for catalog relations, skipping all")));
src/backend/commands/indexcmds.c=ReindexRelationConcurrently(Oid relationOid, int options)
src/backend/commands/indexcmds.c: errmsg("concurrent reindex is not supported for shared relations")));
src/backend/commands/indexcmds.c: errmsg("concurrent reindex is not supported for catalog relations")));

It seems strange to have some cases say "cannot do foo" and other cases
say "foo is not supported". However, I think having
ReindexMultipleTables say "cannot reindex a system catalog" would be
slightly wrong (since we're not reindexing one but many) -- so it would
have to be "cannot reindex system catalogs". And in order to avoid
having two messages that are essentially identical except in number, I
propose to change the others to use the plural too. So the one you just
committed

+				/* A system catalog cannot be reindexed concurrently */
+				if (IsCatalogRelationOid(relationOid))
+					ereport(ERROR,
+							(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+							 errmsg("cannot reindex a system catalog concurrently")));

would become "cannot reindex system catalogs concurrently", identical to
the one in ReindexMultipleTables.

--
�lvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On Tue, May 14, 2019 at 11:32:52AM -0400, Alvaro Herrera wrote:

I do :-)

And actually I am happy to see somebody raising that point. The
current log messages are quite inconsistent for a couple of years now
but I did not bother changing anything other than the new strings per
the feedback I got until, well, yesterday.

It seems strange to have some cases say "cannot do foo" and other cases
say "foo is not supported". However, I think having
ReindexMultipleTables say "cannot reindex a system catalog" would be
slightly wrong (since we're not reindexing one but many) -- so it would
have to be "cannot reindex system catalogs". And in order to avoid
having two messages that are essentially identical except in number, I
propose to change the others to use the plural too. So the one you just
committed

would become "cannot reindex system catalogs concurrently", identical to
the one in ReindexMultipleTables.

There are also a couple of similar, much older, error messages in
index_create() for concurrent creation. Do you think that these
should be changed? I can see benefits for translators to unify things
a bit more, but these do not directly apply to REINDEX, and all
messages are a bit different depending on the context. One argument
to change them is that they don't comply with the project style as
they use full sentences.

Perhaps something like the attached for the REINDEX portion would make
the world a better place? What do you think?
--
Michael

On Wed, May 15, 2019 at 11:17:51AM +0900, Michael Paquier wrote:

Perhaps something like the attached for the REINDEX portion would make
the world a better place? What do you think?

Alvaro, do you have extra thoughts about this patch improving the
error message consistency for REINDEX CONCURRENTLY. I quite like the
suggestions you made and this makes the error strings more
project-like, so I would like to apply it.
--
Michael

On 2019-May-27, Michael Paquier wrote:

On Wed, May 15, 2019 at 11:17:51AM +0900, Michael Paquier wrote:

Perhaps something like the attached for the REINDEX portion would make
the world a better place? What do you think?

Alvaro, do you have extra thoughts about this patch improving the
error message consistency for REINDEX CONCURRENTLY. I quite like the
suggestions you made and this makes the error strings more
project-like, so I would like to apply it.

I wonder if we really want to abolish all distinction between "cannot do
X" and "Y is not supported". I take the former to mean that the
operation is impossible to do for some reason, while the latter means we
just haven't implemented it yet and it seems likely to get implemented
in a reasonable timeframe. See some excellent commentary about about
the "can not" wording at
/messages/by-id/CA+TgmoYS8jKhETyhGYTYMcbvGPwYY=qA6yYp9B47MX7MweE25w@mail.gmail.com

I notice your patch changes "catalog relations" to "system catalogs".
I think we predominantly prefer the latter, so that part of your change
seems OK. (In passing, I noticed we have a couple of places using
"system catalog tables", which is weird.)

I think reindexing system catalogs concurrently is a complex enough
undertaking that implementing it is far enough in the future that the
"cannot" wording is okay; but reindexing partitioned tables is not so
obviously out of the question. We do have "is not yet implemented" in a
couple of other places, so all things considered I'm not so sure about
changing that one to "cannot".

--
�lvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On Mon, May 27, 2019 at 12:20:58AM -0400, Alvaro Herrera wrote:

I wonder if we really want to abolish all distinction between "cannot do
X" and "Y is not supported". I take the former to mean that the
operation is impossible to do for some reason, while the latter means we
just haven't implemented it yet and it seems likely to get implemented
in a reasonable timeframe. See some excellent commentary about about
the "can not" wording at
/messages/by-id/CA+TgmoYS8jKhETyhGYTYMcbvGPwYY=qA6yYp9B47MX7MweE25w@mail.gmail.com

Incorrect URL?

I notice your patch changes "catalog relations" to "system catalogs".
I think we predominantly prefer the latter, so that part of your change
seems OK. (In passing, I noticed we have a couple of places using
"system catalog tables", which is weird.)

Good point. These are not new though, so I would prefer not touch
those parts for this patch.
src/backend/catalog/index.c: errmsg("user-defined
indexes on system catalog tables are not supported")));
src/backend/catalog/index.c: errmsg("concurrent index
creation on system catalog tables is not supported")));
src/backend/catalog/index.c: errmsg("user-defined
indexes on system catalog tables are not supported")));
src/backend/parser/parse_clause.c: errmsg("ON CONFLICT
is not supported with system catalog tables"),

I think reindexing system catalogs concurrently is a complex enough
undertaking that implementing it is far enough in the future that the
"cannot" wording is okay; but reindexing partitioned tables is not so
obviously out of the question.

I am not sure that we actually can without much complication, as
technically locks on catalogs may get released before commit if I
recall correctly.

We do have "is not yet implemented" in a
couple of other places, so all things considered I'm not so sure about
changing that one to "cannot".

Okay. I can live with this difference. Not changing the string in
ReindexRelationConcurrently() has the merit to be consistent with the
existing ones in reindex_relation() and ReindexPartitionedIndex().
Please find attached an updated version. What do you think?
--
Michael