Обсуждение: ALTER TABLE: warn when actions do not recurse to partitions

For now, I’ve limited the change to REPLICA IDENTITY to see whether there are objections to this approach. If there are none, I plan to extend the same warning behavior to the other sub-commands listed above. After that, users can reasonably assume that an ALTER TABLE partitioned_table ... action will recurse to child partitions unless a warning explicitly tells them otherwise.

Now, the message is like:

```

evantest=# alter table sensor_data replica identity full;

NOTICE:  REPLICA IDENTITY does not apply to partitions (1 affected)

ALTER TABLE

Hi Chao

On 13/01/2026 05:02, Chao Li wrote:
> 
> PSA v3:
> 
> * Rephrased the notice message as David's suggestion.
> * Removed partition count from notice message.
> * If a partitioned table doesn't have any partition, then suppress the
> message.

I've been playing with this patch, and it seems to work as expected -
I'm surprised it didn't break any existing tests :). Do you plan to
extend this patch to other subcommands mentioned in your initial post,
such as SET STATISTICS?

Thanks for the patch

Best, Jim

== tests ==

CREATE TABLE m (a int NOT NULL, b int) PARTITION BY RANGE (a);
CREATE TABLE m_p1 PARTITION OF m FOR VALUES FROM (1) TO (10);
CREATE TABLE m_p2 PARTITION OF m FOR VALUES FROM (10) TO (20);

CREATE UNIQUE INDEX m_idx ON m(a);
CREATE UNIQUE INDEX m_p1_idx ON m_p1(a);
CREATE UNIQUE INDEX m_p2_idx ON m_p2(a);
CREATE TABLE
CREATE TABLE
CREATE TABLE
CREATE INDEX
CREATE INDEX
CREATE INDEX

-- issue a NOTICE (m has partitions)
ALTER TABLE m REPLICA IDENTITY NOTHING;
ALTER TABLE m REPLICA IDENTITY FULL;
ALTER TABLE m REPLICA IDENTITY DEFAULT;
ALTER TABLE m REPLICA IDENTITY USING INDEX m_idx;
NOTICE:  present partitions not affected
HINT:  partitions may be modified individually using separate commands
ALTER TABLE
NOTICE:  present partitions not affected
HINT:  partitions may be modified individually using separate commands
ALTER TABLE
NOTICE:  present partitions not affected
HINT:  partitions may be modified individually using separate commands
ALTER TABLE
NOTICE:  present partitions not affected
HINT:  partitions may be modified individually using separate commands
ALTER TABLE

-- does not issue a NOTICE (with ONLY: no recursion into partitions)
ALTER TABLE ONLY m REPLICA IDENTITY NOTHING;
ALTER TABLE ONLY m REPLICA IDENTITY FULL;
ALTER TABLE ONLY m REPLICA IDENTITY DEFAULT;
ALTER TABLE ONLY m REPLICA IDENTITY USING INDEX m_idx;
ALTER TABLE
ALTER TABLE
ALTER TABLE
ALTER TABLE

-- does not issue a NOTICE (p1 has no partitions)
ALTER TABLE m_p1 REPLICA IDENTITY NOTHING;
ALTER TABLE m_p1 REPLICA IDENTITY FULL;
ALTER TABLE m_p1 REPLICA IDENTITY DEFAULT;
ALTER TABLE m_p1 REPLICA IDENTITY USING INDEX m_p1_idx;
ALTER TABLE
ALTER TABLE
ALTER TABLE
ALTER TABLE

-- does not issue a NOTICE (m no longer has partitions)
DROP TABLE m_p1, m_p2;
DROP TABLE

ALTER TABLE m REPLICA IDENTITY NOTHING;
ALTER TABLE m REPLICA IDENTITY FULL;
ALTER TABLE m REPLICA IDENTITY DEFAULT;
ALTER TABLE m REPLICA IDENTITY USING INDEX m_idx;
ALTER TABLE
ALTER TABLE
ALTER TABLE
ALTER TABLE

> On Jan 13, 2026, at 19:16, Jim Jones <jim.jones@uni-muenster.de> wrote:
>
> Hi Chao
>
> On 13/01/2026 05:02, Chao Li wrote:
>>
>> PSA v3:
>>
>> * Rephrased the notice message as David's suggestion.
>> * Removed partition count from notice message.
>> * If a partitioned table doesn't have any partition, then suppress the
>> message.
>
> I've been playing with this patch, and it seems to work as expected -
> I'm surprised it didn't break any existing tests :). Do you plan to
> extend this patch to other subcommands mentioned in your initial post,
> such as SET STATISTICS?
>
> Thanks for the patch
>
> Best, Jim
>

Hi Jim,

Thanks for your testing. Yes, I plan to add the notice to other sub-commands as needed. I only updated REPLICA IDENTITY
firstto call for feedback. As you see, David has suggested the great wording for the notice message. Once the change on
REPLICAIDENTITY is reviewed, it’s easy to extend to other sub-commands. 

Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/

Hi Chao

On 22/01/2026 06:45, Chao Li wrote:
> evantest=# alter table p_test replica identity full, alter column
> username set (n_distinct = 0.1);
> NOTICE:  ALTER action REPLICA IDENTITY on relation "p_test" does not
> affect present partitions
> HINT:  partitions may be modified individually, or specify ONLY to
> suppress this message
> NOTICE:  ALTER action ALTER COLUMN ... SET on relation "p_test" does not
> affect present partitions
> HINT:  partitions may be modified individually, or specify ONLY to
> suppress this message
> ALTER TABLE


One could argue that encapsulating all conditions in
EmitPartitionNoRecurseNotice(), meaning it is called all the time, is
slightly inefficient, but the impact is really negligible in this case -
and it is how it is done in similar functions in tablecmds.c :) The code
LGTM.

One small thing:

errhint is supposed to be capitalised - see Error Message Style Guide[1]

"Detail and hint messages: Use complete sentences, and end each with a
period. Capitalize the first word of sentences. Put two spaces after the
period if another sentence follows (for English text; might be
inappropriate in other languages)."

ereport(NOTICE,
    errmsg("ALTER action %s on relation \"%s\" does not affect present
partitions",
           action_str,
           RelationGetRelationName(rel)),
    errhint("partitions may be modified individually, or specify ONLY to
suppress this message"));

What about this?

HINT: To update partitions, apply the command to each one individually,
or specify ONLY to suppress this message.

I'll test the newly covered subcomands tomorrow.

Best, Jim

1 - https://www.postgresql.org/docs/current/error-style-guide.html

> On Jan 23, 2026, at 03:27, Jim Jones <jim.jones@uni-muenster.de> wrote:
>
> Hi Chao
>
> On 22/01/2026 06:45, Chao Li wrote:
>> evantest=# alter table p_test replica identity full, alter column
>> username set (n_distinct = 0.1);
>> NOTICE:  ALTER action REPLICA IDENTITY on relation "p_test" does not
>> affect present partitions
>> HINT:  partitions may be modified individually, or specify ONLY to
>> suppress this message
>> NOTICE:  ALTER action ALTER COLUMN ... SET on relation "p_test" does not
>> affect present partitions
>> HINT:  partitions may be modified individually, or specify ONLY to
>> suppress this message
>> ALTER TABLE
>
>
> One could argue that encapsulating all conditions in
> EmitPartitionNoRecurseNotice(), meaning it is called all the time, is
> slightly inefficient, but the impact is really negligible in this case -
> and it is how it is done in similar functions in tablecmds.c :) The code
> LGTM.

Hi Jim, thanks a lot for the review.

>
> One small thing:
>
> errhint is supposed to be capitalised - see Error Message Style Guide[1]

Thanks for the info, I wasn’t aware of that. When I wrote the code, I searched “errhint” over the source tree, and
didn’tfind a standard to follow. 

>
> "Detail and hint messages: Use complete sentences, and end each with a
> period. Capitalize the first word of sentences. Put two spaces after the
> period if another sentence follows (for English text; might be
> inappropriate in other languages)."
>
> ereport(NOTICE,
> errmsg("ALTER action %s on relation \"%s\" does not affect present
> partitions",
>   action_str,
>   RelationGetRelationName(rel)),
> errhint("partitions may be modified individually, or specify ONLY to
> suppress this message"));
>
> What about this?
>
> HINT: To update partitions, apply the command to each one individually,
> or specify ONLY to suppress this message.

Looks good. I will integrate your edit to the next version.

>
> I'll test the newly covered subcomands tomorrow.

Thanks again for testing. I will wait to see the test results and address all issues together in next version.

Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/

On 23/01/2026 01:11, Chao Li wrote:
> I will wait to see the test results and address all issues together in next version.

While testing some edge cases I found out that the NOTICE is being
emitted too early in the code path, e.g.

postgres=# ALTER TABLE m ALTER COLUMN b SET COMPRESSION pglz;
NOTICE:  ALTER action ALTER COLUMN ... SET COMPRESSION on relation "m"
does not affect present partitions
HINT:  partitions may be modified individually, or specify ONLY to
suppress this message
ERROR:  column data type integer does not support compression

I'd argue that emitting only the ERROR message in this case would be the
right approach. What about moving the EmitPartitionNoRecurseNotice()
call to ATExecCmd, right **after** the changes were successfully
executed? For instance, in the case I mentioned above, you could explore:

@@ -5446,6 +5475,8 @@ ATExecCmd(List **wqueue, AlteredTableInfo *tab,
                case AT_SetCompression: /* ALTER COLUMN SET COMPRESSION */
                        address = ATExecSetCompression(rel, cmd->name,
cmd->def,

          lockmode);
+                       /* Emit notice after validation passes */
+                       EmitPartitionNoRecurseNotice(cmd->subtype, rel,
cmd->recurse, false);
                        break;

Not sure if cmd->recurse is propagated in this code path. If not, you
might need to do it manually, e.g.

@@ -4936,6 +4937,14 @@ ATPrepCmd(List **wqueue, Relation rel,
AlterTableCmd *cmd,
         */
        cmd = copyObject(cmd);

+       if (recurse)
+               cmd->recurse = true;
+

I'm not saying it should be exactly this way, but it sounds more
reasonable to me to emit the NOTICE only if we know that the command is
going to be successfully executed (or was successfully executed).

This patch touches a lot of regression tests, but mostly to add the
keyword ONLY to the ALTER TABLE statements, to avoid the NOTICE message,
so that's ok.

Thanks!

Best, Jim