Обсуждение: Confused comment about drop replica identity index

Hi hackers,

When I doing development based by PG, I found the following comment have a
little problem in file src/include/catalog/pg_class.h.

/*
 * an explicitly chosen candidate key's columns are used as replica identity.
 * Note this will still be set if the index has been dropped; in that case it
 * has the same meaning as 'd'.
 */
#define          REPLICA_IDENTITY_INDEX    'i'

The last sentence makes me a little confused :
[......in that case it as the same meaning as 'd'.]

Now, pg-doc didn't have a clear style to describe this.


But if I drop relation's replica identity index like the comment, the action
is not as same as default.

For example:
Execute the following SQL:
create table tbl (col1 int  primary key, col2 int not null);
create unique INDEX ON tbl(col2);
alter table tbl replica identity using INDEX tbl_col2_idx;
drop index tbl_col2_idx;
create publication pub for table tbl;
delete from tbl;

Actual result:
ERROR:  cannot delete from table "tbl" because it does not have a replica identity and publishes deletes
HINT:  To enable deleting from the table, set REPLICA IDENTITY using ALTER TABLE.

Expected result in comment:
DELETE 0


I found that in the function CheckCmdReplicaIdentity, the operation described
in the comment is not considered,
When relation's replica identity index is found to be InvalidOid, an error is
reported.

Are the comment here not accurate enough?
Or we need to adjust the code according to the comments?


Regards,
Wang wei

On Tue, Dec 14, 2021 at 6:08 PM wangw.fnst@fujitsu.com
<wangw.fnst@fujitsu.com> wrote:
>
> Hi hackers,
>
> When I doing development based by PG, I found the following comment have a
> little problem in file src/include/catalog/pg_class.h.
>
> /*
>  * an explicitly chosen candidate key's columns are used as replica identity.
>  * Note this will still be set if the index has been dropped; in that case it
>  * has the same meaning as 'd'.
>  */
> #define           REPLICA_IDENTITY_INDEX        'i'
>
> The last sentence makes me a little confused :
> [......in that case it as the same meaning as 'd'.]
>
> Now, pg-doc didn't have a clear style to describe this.
>
>
> But if I drop relation's replica identity index like the comment, the action
> is not as same as default.
>
> For example:
> Execute the following SQL:
> create table tbl (col1 int  primary key, col2 int not null);
> create unique INDEX ON tbl(col2);
> alter table tbl replica identity using INDEX tbl_col2_idx;
> drop index tbl_col2_idx;
> create publication pub for table tbl;
> delete from tbl;
>
> Actual result:
> ERROR:  cannot delete from table "tbl" because it does not have a replica identity and publishes deletes
> HINT:  To enable deleting from the table, set REPLICA IDENTITY using ALTER TABLE.

I think I see where's the confusion. The table has a primary key and
so when the replica identity index is dropped, per the comment in
code, you expect that primary key will be used as replica identity
since that's what 'd' or default means.

>
> Expected result in comment:
> DELETE 0
>
>
> I found that in the function CheckCmdReplicaIdentity, the operation described
> in the comment is not considered,
> When relation's replica identity index is found to be InvalidOid, an error is
> reported.

This code in RelationGetIndexList() is not according to that comment.

   if (replident == REPLICA_IDENTITY_DEFAULT && OidIsValid(pkeyIndex))
        relation->rd_replidindex = pkeyIndex;
    else if (replident == REPLICA_IDENTITY_INDEX && OidIsValid(candidateIndex))
        relation->rd_replidindex = candidateIndex;
    else
        relation->rd_replidindex = InvalidOid;

>
> Are the comment here not accurate enough?
> Or we need to adjust the code according to the comments?
>

Comment in code is one thing, but I think PG documentation is not
covering the use case you tried. What happens when a replica identity
index is dropped has not been covered either in ALTER TABLE
https://www.postgresql.org/docs/13/sql-altertable.html or DROP INDEX
https://www.postgresql.org/docs/14/sql-dropindex.html documentation.


-- 
Best Wishes,
Ashutosh Bapat

On Tue, Dec 14, 2021 at 07:10:49PM +0530, Ashutosh Bapat wrote:
> This code in RelationGetIndexList() is not according to that comment.
>
>    if (replident == REPLICA_IDENTITY_DEFAULT && OidIsValid(pkeyIndex))
>         relation->rd_replidindex = pkeyIndex;
>     else if (replident == REPLICA_IDENTITY_INDEX && OidIsValid(candidateIndex))
>         relation->rd_replidindex = candidateIndex;
>     else
>         relation->rd_replidindex = InvalidOid;

Yeah, the comment is wrong.  If the index of a REPLICA_IDENTITY_INDEX
is dropped, I recall that the behavior is the same as
REPLICA_IDENTITY_NOTHING.

> Comment in code is one thing, but I think PG documentation is not
> covering the use case you tried. What happens when a replica identity
> index is dropped has not been covered either in ALTER TABLE
> https://www.postgresql.org/docs/13/sql-altertable.html or DROP INDEX
> https://www.postgresql.org/docs/14/sql-dropindex.html documentation.

Not sure about the DROP INDEX page, but I'd be fine with mentioning
that in the ALTER TABLE page in the paragraph related to REPLICA
IDENTITY.  While on it, I would be tempted to switch this stuff to use
a list of <variablelist> for all the option values.  That would be
much easier to read.

[ ... thinks a bit ... ]

FWIW, this brings back some memories, as of this thread:
https://www.postgresql.org/message-id/20200522035028.GO2355@paquier.xyz

See also commit fe7fd4e from August 2020, where some tests have been
added.  I recall seeing this incorrect comment from last year's
thread and it may have been mentioned in one of the surrounding
threads..  Maybe I just let it go back then.  I don't know.
--
Michael

On Tue, Dec 15, 2021 at 11:25AM, Michael Paquier wrote:
> Yeah, the comment is wrong.  If the index of a REPLICA_IDENTITY_INDEX is
> dropped, I recall that the behavior is the same as REPLICA_IDENTITY_NOTHING.

Thank you for your response.
I agreed that the comment is wrong.


> Not sure about the DROP INDEX page, but I'd be fine with mentioning that in the
> ALTER TABLE page in the paragraph related to REPLICA IDENTITY.  While on it, I
> would be tempted to switch this stuff to use a list of <variablelist> for all the option
> values.  That would be much easier to read.

Yeah, if we can add some details to pg-doc and code comments, I think it will
be more friendly to PG users and developers.

Regards,
Wang wei

On Wed, Dec 15, 2021 at 09:18:26AM +0000, wangw.fnst@fujitsu.com wrote:
> Yeah, if we can add some details to pg-doc and code comments, I think it will
> be more friendly to PG users and developers.

Would you like to write a patch to address all that?
Thanks,
--
Michael

On Tue, Dec 16, 2021 at 06:40AM, Michael Paquier wrote:
> Would you like to write a patch to address all that?

OK, I will push it soon.


Regards,
Wang wei

On 2021-Dec-15, Michael Paquier wrote:

> On Tue, Dec 14, 2021 at 07:10:49PM +0530, Ashutosh Bapat wrote:
> > This code in RelationGetIndexList() is not according to that comment.
> > 
> >    if (replident == REPLICA_IDENTITY_DEFAULT && OidIsValid(pkeyIndex))
> >         relation->rd_replidindex = pkeyIndex;
> >     else if (replident == REPLICA_IDENTITY_INDEX && OidIsValid(candidateIndex))
> >         relation->rd_replidindex = candidateIndex;
> >     else
> >         relation->rd_replidindex = InvalidOid;
> 
> Yeah, the comment is wrong.  If the index of a REPLICA_IDENTITY_INDEX
> is dropped, I recall that the behavior is the same as
> REPLICA_IDENTITY_NOTHING.

Hmm, so if a table has REPLICA IDENTITY INDEX and there is a publication
with an explicit column list, then we need to forbid the DROP INDEX for
that index.

I wonder why don't we just forbid DROP INDEX of an index that's been
defined as replica identity.  It seems quite silly an operation to
allow.

-- 
Álvaro Herrera           39°49'30"S 73°17'W  —  https://www.EnterpriseDB.com/
"Ed is the standard text editor."
      http://groups.google.com/group/alt.religion.emacs/msg/8d94ddab6a9b0ad3

On Thu, Dec 16, 2021 at 03:08:46PM -0300, Alvaro Herrera wrote:
> Hmm, so if a table has REPLICA IDENTITY INDEX and there is a publication
> with an explicit column list, then we need to forbid the DROP INDEX for
> that index.

Hmm.  I have not followed this thread very closely.

> I wonder why don't we just forbid DROP INDEX of an index that's been
> defined as replica identity.  It seems quite silly an operation to
> allow.

The commit logs talk about b23b0f55 here for this code, to ease the
handling of relcache entries for rd_replidindex.  07cacba is the
origin of the logic (see RelationGetIndexList).  Andres?

I don't think that this is really an argument against putting more
restrictions as anything that deals with an index drop, including the
internal ones related to constraints, would need to go through
index_drop(), and new features may want more restrictions in place as
you say.

Now, I don't see a strong argument in changing this behavior either
(aka I have not looked at what this implies for the new publication
types), and we still need to do something for the comment/docs in
existing branches, anyway.  So I would still fix this gap as a first
step, then deal with the rest on HEAD as necessary.
--
Michael

On Thu, Dec 16, 2021 at 03:08:46PM -0300, Alvaro Herrera wrote:

> Hmm, so if a table has REPLICA IDENTITY INDEX and there is a publication

> with an explicit column list, then we need to forbid the DROP INDEX for

> that index.

Hmm.  I have not followed this thread very closely.

> I wonder why don't we just forbid DROP INDEX of an index that's been

> defined as replica identity.  It seems quite silly an operation to

> allow.

The commit logs talk about b23b0f55 here for this code, to ease the

handling of relcache entries for rd_replidindex.  07cacba is the

origin of the logic (see RelationGetIndexList).  Andres?

I don't think that this is really an argument against putting more

restrictions as anything that deals with an index drop, including the

internal ones related to constraints, would need to go through

index_drop(), and new features may want more restrictions in place as

you say.

Now, I don't see a strong argument in changing this behavior either

(aka I have not looked at what this implies for the new publication

types), and we still need to do something for the comment/docs in

existing branches, anyway.  So I would still fix this gap as a first

step, then deal with the rest on HEAD as necessary.

On Tue, Dec 16, 2021 at 10:27AM, Michael Paquier wrote:
> On Tue, Dec 16, 2021 at 06:40AM, Michael Paquier wrote:
> > Would you like to write a patch to address all that?
>
> OK, I will push it soon.


Here is a patch to correct wrong comment about REPLICA_IDENTITY_INDEX, And improve the pg-doc.


Regards,
Wang wei

On Mon, Dec 20, 2021 at 03:46:13AM +0000, wangw.fnst@fujitsu.com wrote:
> Here is a patch to correct wrong comment about
> REPLICA_IDENTITY_INDEX, And improve the pg-doc.

That's mostly fine.  I have made some adjustments as per the
attached.

+         The default for non-system tables. Records the old values of the columns
+         of the primary key, if any. The default for non-system tables.
The same sentence is repeated twice.

+         Records no information about the old row.(This is the
default for system tables.)
For consistency with the rest, this could drop the parenthesis for the
second sentence.

+       <term><literal>USING INDEX index_name</literal></term>
This should use <replaceable> as markup for index_name.

Pondering more about this thread, I don't think we should change the
existing behavior in the back-branches, but I don't have any arguments
about doing such changes on HEAD to help the features being worked
on, either.  So I'd like to apply and back-patch the attached, as a
first step, to fix the inconsistency.
--
Michael

On Mon, Dec 20, 2021 at 03:46:13AM +0000, wangw.fnst@fujitsu.com wrote:

> Here is a patch to correct wrong comment about

> REPLICA_IDENTITY_INDEX, And improve the pg-doc.

That's mostly fine.  I have made some adjustments as per the

attached.

Pondering more about this thread, I don't think we should change the

existing behavior in the back-branches, but I don't have any arguments

about doing such changes on HEAD to help the features being worked

on, either.  So I'd like to apply and back-patch the attached, as a

first step, to fix the inconsistency.

On Tue, Dec 20, 2021 at 19:11PM, Michael Paquier wrote:
> That's mostly fine.  I have made some adjustments as per the attached.

Thanks for reviewing.


> +         The default for non-system tables. Records the old values of the columns
> +         of the primary key, if any. The default for non-system tables.
> The same sentence is repeated twice.
>
> +         Records no information about the old row.(This is the
> default for system tables.)
> For consistency with the rest, this could drop the parenthesis for the second
> sentence.
>
> +       <term><literal>USING INDEX index_name</literal></term>
> This should use <replaceable> as markup for index_name.

The change looks good to me.



Regards,
Wang wei

On Tues, Dec 21, 2021 8:47 AM Michael Paquier <michael@paquier.xyz> wrote:
> On Mon, Dec 20, 2021 at 11:57:32AM -0300, Euler Taveira wrote:
> > What do you think about the attached patch? It forbids the DROP INDEX.
> > We might add a detail message but I didn't in this patch.
>
> Yeah.  I'd agree about doing something like that on HEAD, and that would help
> with some of the logirep-related patch currently being worked on, as far as I
> understood.

Hi,

I think forbids DROP INDEX might not completely solve this problem. Because
user could still use other command to delete the index, for example: ALTER
TABLE DROP COLUMN. After dropping the column, the index on it will also be
dropped.

Besides, user can also ALTER REPLICA IDENTITY USING INDEX "primary key", and in
this case, when they ALTER TABLE DROP CONSTR "PRIMARY KEY", the replica
identity index will also be dropped.

Best regards,
Hou zj

On Thu, Dec 30, 2021 at 06:45:30AM +0000, houzj.fnst@fujitsu.com wrote:
> I think forbids DROP INDEX might not completely solve this problem. Because
> user could still use other command to delete the index, for example: ALTER
> TABLE DROP COLUMN. After dropping the column, the index on it will also be
> dropped.
>
> Besides, user can also ALTER REPLICA IDENTITY USING INDEX "primary key", and in
> this case, when they ALTER TABLE DROP CONSTR "PRIMARY KEY", the replica
> identity index will also be dropped.

Indexes related to any other object type, like constraints, are
dropped as part of index_drop() as per the handling of dependencies.
So, by putting a restriction there, any commands would take this code
path, and fail when trying to drop an index used as a replica
identity.  Why would that be logically a problem?  We may want errors
with more context for such cases, though, as complaining about an
object not directly known by the user when triggering a different
command, like a constraint index, could be confusing.
--
Michael