Re: Disallow UPDATE/DELETE on table with unpublished generated column as REPLICA IDENTITY

On Fri, 15 Nov 2024 at 16:45, Shlok Kyal <shlok(dot)kyal(dot)oss(at)gmail(dot)com> wrote:
>
> Thanks for providing the comments
>
> On Fri, 15 Nov 2024 at 10:59, vignesh C <vignesh21(at)gmail(dot)com> wrote:
> >
> > On Thu, 14 Nov 2024 at 15:51, Shlok Kyal <shlok(dot)kyal(dot)oss(at)gmail(dot)com> wrote:
> > >
> > > Thanks for providing the comments.
> > >
> > > On Thu, 14 Nov 2024 at 12:22, vignesh C <vignesh21(at)gmail(dot)com> wrote:
> > > >
> > > > On Wed, 13 Nov 2024 at 11:15, Shlok Kyal <shlok(dot)kyal(dot)oss(at)gmail(dot)com> wrote:
> > > > >
> > > > > Thanks for providing the comments.
> > > > >
> > > > > On Tue, 12 Nov 2024 at 12:52, Zhijie Hou (Fujitsu)
> > > > > <houzj(dot)fnst(at)fujitsu(dot)com> wrote:
> > > > > >
> > > > > > On Friday, November 8, 2024 7:06 PM Shlok Kyal <shlok(dot)kyal(dot)oss(at)gmail(dot)com> wrote:
> > > > > > >
> > > > > > > Hi Amit,
> > > > > > >
> > > > > > > On Thu, 7 Nov 2024 at 11:37, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > > > > > > >
> > > > > > > > On Tue, Nov 5, 2024 at 12:53 PM Shlok Kyal <shlok(dot)kyal(dot)oss(at)gmail(dot)com>
> > > > > > > wrote:
> > > > > > > > >
> > > > > > > > > To avoid the issue, we can disallow UPDATE/DELETE on table with
> > > > > > > > > unpublished generated column as REPLICA IDENTITY. I have attached a
> > > > > > > > > patch for the same.
> > > > > > > > >
> > > > > > > >
> > > > > > > > +CREATE PUBLICATION pub_gencol FOR TABLE testpub_gencol; UPDATE
> > > > > > > > +testpub_gencol SET a = 100 WHERE a = 1;
> > > > > > > > +ERROR: cannot update table "testpub_gencol"
> > > > > > > > +DETAIL: Column list used by the publication does not cover the
> > > > > > > > replica identity.
> > > > > > > >
> > > > > > > > This is not a correct ERROR message as the publication doesn't have
> > > > > > > > any column list associated with it. You have added the code to detect
> > > > > > > > this in the column list code path which I think is not required. BTW,
> > > > > > > > you also need to consider the latest commit 7054186c4e for this. I
> > > > > > > > guess you need to keep another flag in PublicationDesc to detect this
> > > > > > > > and then give an appropriate ERROR.
> > > > > > >
> > > > > > > I have addressed the comments and provided an updated patch. Also, I am
> > > > > > > currently working to fix this issue in back branches.
> > > > > >
> > > > > > Thanks for the patch. I am reviewing it and have some initial comments:
> > > > > >
> > > > > >
> > > > > > 1.
> > > > > > + char attgenerated = get_attgenerated(relid, attnum);
> > > > > > +
> > > > > >
> > > > > > I think it's unnecessary to initialize attgenerated here because the value will
> > > > > > be overwritten if pubviaroot is true anyway. Also, the get_attgenerated()
> > > > > > is not cheap.
> > > > > >
> > > > > Fixed
> > > > >
> > > > > > 2.
> > > > > >
> > > > > > I think the patch missed to check the case when table is marked REPLICA
> > > > > > IDENTITY FULL, and generated column is not published:
> > > > > >
> > > > > > CREATE TABLE testpub_gencol (a INT, b INT GENERATED ALWAYS AS (a + 1) STORED NOT NULL);
> > > > > > ALTER TABLE testpub_gencol REPLICA IDENTITY FULL;
> > > > > > CREATE PUBLICATION pub_gencol FOR TABLE testpub_gencol;
> > > > > > UPDATE testpub_gencol SET a = 2;
> > > > > >
> > > > > > I expected the UPDATE to fail in above case, but it can still pass after applying the patch.
> > > > > >
> > > > > Fixed
> > > > >
> > > > > > 3.
> > > > > >
> > > > > > + * If the publication is FOR ALL TABLES we can skip the validation.
> > > > > > + */
> > > > > >
> > > > > > This comment seems not clear to me, could you elaborate a bit more on this ?
> > > > > >
> > > > > I missed to handle the case FOR ALL TABLES. Have removed the comment.
> > > > >
> > > > > > 4.
> > > > > >
> > > > > > Also, I think the patch does not handle the FOR ALL TABLE case correctly:
> > > > > >
> > > > > > CREATE TABLE testpub_gencol (a INT, b INT GENERATED ALWAYS AS (a + 1) STORED NOT NULL);
> > > > > > CREATE UNIQUE INDEX testpub_gencol_idx ON testpub_gencol (b);
> > > > > > ALTER TABLE testpub_gencol REPLICA IDENTITY USING index testpub_gencol_idx;
> > > > > > CREATE PUBLICATION pub_gencol FOR ALL TABLEs;
> > > > > > UPDATE testpub_gencol SET a = 2;
> > > > > >
> > > > > > I expected the UPDATE to fail in above case as well.
> > > > > >
> > > > > Fixed
> > > > >
> > > > > > 5.
> > > > > >
> > > > > > + else if (cmd == CMD_UPDATE && !pubdesc.replident_has_valid_gen_cols)
> > > > > > + ereport(ERROR,
> > > > > > + (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
> > > > > > + errmsg("cannot update table \"%s\"",
> > > > > > + RelationGetRelationName(rel)),
> > > > > > + errdetail("REPLICA IDENTITY consists of an unpublished generated column.")));
> > > > > >
> > > > > > I think it would be better to use lower case "replica identity" to consistent
> > > > > > with other existing messages.
> > > > > >
> > > > > Fixed
> > > > >
> > > > > I have attached the updated patch here.
> > > >
> > > > Few comments:
> > > > 1) In the first check relation->rd_rel->relispartition also is checked
> > > > whereas in the below it is not checked, shouldn't the same check be
> > > > there below to avoid few of the function calls which are not required:
> > > > + if (pubviaroot && relation->rd_rel->relispartition)
> > > > + {
> > > > + publish_as_relid =
> > > > GetTopMostAncestorInPublication(pubid, ancestors, NULL);
> > > > +
> > > > + if (!OidIsValid(publish_as_relid))
> > > > + publish_as_relid = relid;
> > > > + }
> > > > +
> > > >
> > > > + if (pubviaroot)
> > > > + {
> > > > + /* attribute name in the child table */
> > > > + char *colname =
> > > > get_attname(relid, attnum, false);
> > > > +
> > > > + /*
> > > > + * Determine the attnum for the
> > > > attribute name in parent (we
> > > > + * are using the column list defined
> > > > on the parent).
> > > > + */
> > > > + attnum = get_attnum(publish_as_relid, colname);
> > > > + attgenerated =
> > > > get_attgenerated(publish_as_relid, attnum);
> > > > + }
> > > > + else
> > > > + attgenerated = get_attgenerated(relid, attnum);
> > > I have updated the if condititon
> > >
> > > > 2) I think we could use check_and_fetch_column_list to see that it is
> > > > not a column list publication instead of below code:
> > > > + if (!puballtables)
> > > > + {
> > > > + tuple = SearchSysCache2(PUBLICATIONRELMAP,
> > > > +
> > > > ObjectIdGetDatum(publish_as_relid),
> > > > +
> > > > ObjectIdGetDatum(pubid));
> > > > +
> > > > + if (!HeapTupleIsValid(tuple))
> > > > + return false;
> > > > +
> > > > + (void) SysCacheGetAttr(PUBLICATIONRELMAP, tuple,
> > > > +
> > > > Anum_pg_publication_rel_prattrs,
> > > > + &isnull);
> > > > +
> > > > + ReleaseSysCache(tuple);
> > > > + }
> > > > +
> > > > + if(puballtables || isnull)
> > >
> > > Yes we can use it. I have updated the patch.
> > >
> > > > 3) Since there is only a single statement, remove the enclosing parenthisis:
> > > > + if (!pubform->pubgencols &&
> > > > + (pubform->pubupdate || pubform->pubdelete) &&
> > > > + replident_has_unpublished_gen_col(pubid,
> > > > relation, ancestors,
> > > > +
> > > > pubform->pubviaroot, pubform->puballtables))
> > > > + {
> > > > + pubdesc->replident_has_valid_gen_cols = false;
> > > > + }
> > > >
> > > Fixed
> > >
> > > > 4) Pgindent should be run there are few issues:
> > > > 4.a)
> > > > +extern bool replident_has_unpublished_gen_col(Oid pubid, Relation relation,
> > > > +
> > > > List *ancestors, bool pubviaroot, bool
> > > > puballtables);
> > > > 4.b)
> > > > + }
> > > > +
> > > > + if(puballtables || isnull)
> > > > + {
> > > > + int x;
> > > > + Bitmapset *idattrs = NULL;
> > > > 4.c)
> > > > + * generated column we should error out.
> > > > + */
> > > > + if(relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL &&
> > > > + relation->rd_att->constr &&
> > > > relation->rd_att->constr->has_generated_stored)
> > > > + result = true;
> > > > 4.d)
> > > > + while ((x = bms_next_member(idattrs, x)) >= 0)
> > > > + {
> > > > + AttrNumber attnum = (x +
> > > > FirstLowInvalidHeapAttributeNumber);
> > > > + char attgenerated;
> > > >
> > > Fixed
> > >
> > > > 5) You could do this in a single line comment:
> > > > + /*
> > > > + * Check if any REPLICA IDENTITY column is an generated column.
> > > > + */
> > > > + while ((x = bms_next_member(idattrs, x)) >= 0)
> > > >
> > >
> > > Fixed
> > >
> > > > 6) I felt one of update or delete is enough in this case as the code
> > > > path is same:
> > > > +UPDATE testpub_gencol SET a = 100 WHERE a = 1;
> > > > +DELETE FROM testpub_gencol WHERE a = 100;
> > > > +
> > > > +-- error - generated column "b" is not published and REPLICA IDENTITY
> > > > is set FULL
> > > > +ALTER TABLE testpub_gencol REPLICA IDENTITY FULL;
> > > > +UPDATE testpub_gencol SET a = 100 WHERE a = 1;
> > > > +DELETE FROM testpub_gencol WHERE a = 100;
> > > > +DROP PUBLICATION pub_gencol;
> > > > +
> > > > +-- ok - generated column "b" is published and is part of REPLICA IDENTITY
> > > > +CREATE PUBLICATION pub_gencol FOR TABLE testpub_gencol with
> > > > (publish_generated_columns = true);
> > > > +UPDATE testpub_gencol SET a = 100 WHERE a = 1;
> > > > +DELETE FROM testpub_gencol WHERE a = 100;
> > >
> > > Removed the 'DELETE' case.
> > >
> > > I have addressed the comments and updated the patch.
> >
> > Few comments:
> > 1) Current patch will not handle this scenario where subset of columns
> > are specified in the replica identity index:
> > CREATE TABLE t1 (a INT not null, a1 int not null, a2 int not null, b
> > INT GENERATED ALWAYS AS (a + 1) STORED NOT NULL);
> > create unique index idx1_t1 on t1(a, a1);
> >
> > -- Replica identity will have subset of table columns
> > alter table t1 replica identity using index idx1_t1 ;
> > insert into t1 values(1,1,1);
> > create publication pub1 for table t1;
> >
> > postgres=# update t1 set a = 2;
> > UPDATE 1
> >
> > I felt we should throw an error in this case too.
> >
> I feel the above behaviour is expected. I think we can specify a
> subset of columns in the replica identity index as per documentation
> [1]. Thoughts?
>
> > 2) Instead of checking if replica identity has a generated column, can
> > we check if the columns that will be published and the columns in the
> > replica identity matches:
> > + if (pubviaroot && relation->rd_rel->relispartition)
> > + {
> > + /* attribute name in the child table */
> > + char *colname =
> > get_attname(relid, attnum, false);
> > +
> > + /*
> > + * Determine the attnum for the
> > attribute name in parent (we
> > + * are using the column list defined
> > on the parent).
> > + */
> > + attnum = get_attnum(publish_as_relid, colname);
> > + attgenerated =
> > get_attgenerated(publish_as_relid, attnum);
> > + }
> > + else
> > + attgenerated = get_attgenerated(relid, attnum);
> >
> Fixed
>
> > 3) publish_as_relid will be set accordingly based on pubviaroot, so it
> > need not be initialized:
> > + Oid relid = RelationGetRelid(relation);
> > + Oid publish_as_relid = RelationGetRelid(relation);
> > + bool result = false;
> >
> Fixed
>
> I have addressed the comments and attached the updated patch.

Few comments:
1) I felt we can return from here after identifying it is replica
identity full instead of processing further:
+ /*
+ * REPLICA IDENTITY can be FULL only if there is no
column list for
+ * publication. If REPLICA IDENTITY is set as FULL and
relation has a
+ * generated column we should error out.
+ */
+ if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL &&
+ relation->rd_att->constr &&
+ relation->rd_att->constr->has_generated_stored)
+ result = true;

2) columns bms also should be freed here:
/* replica identity column, not covered by the
column list */
+ if (!bms_is_member(attnum, columns))
+ {
+ result = true;
+ break;
+ }
+ }
+
+ bms_free(idattrs);

3) Error detail message should begin with upper case:
3.a)
@@ -809,6 +809,12 @@ CheckCmdReplicaIdentity(Relation rel, CmdType cmd)
errmsg("cannot update table \"%s\"",
RelationGetRelationName(rel)),
errdetail("Column list used by the
publication does not cover the replica identity.")));
+ else if (cmd == CMD_UPDATE && !pubdesc.replident_has_valid_gen_cols)
+ ereport(ERROR,
+ (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
+ errmsg("cannot update table \"%s\"",
+ RelationGetRelationName(rel)),
+ errdetail("replica identity consists
of an unpublished generated column.")));
else if (cmd == CMD_DELETE && !pubdesc.rf_valid_for_delete)
ereport(ERROR,
(errcode(ERRCODE_INVALID_COLUMN_REFERENCE),

3.b) Similarly here too:
errdetail("Column list used by the
publication does not cover the replica identity.")));
+ else if (cmd == CMD_DELETE && !pubdesc.replident_has_valid_gen_cols)
+ ereport(ERROR,
+ (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
+ errmsg("cannot delete from table \"%s\"",
+ RelationGetRelationName(rel)),
+ errdetail("replica identity consists
of an unpublished generated column.")));

Regards,
Vignesh