Re: [GENERAL] Using an SMP machine to make multiple indices on the same

> -----Original Message-----
> From: Hiroshi Inoue
>
> Tom Lane wrote:
> >
> > Martin Weinberg <weinberg@osprey.astro.umass.edu> writes:
> > > Yes, I understand locking the table, but empirically, two index
> > > creations will not run simultaneously on the same table.
> >
> > Hmm, on trying it you are right.  The second index creation blocks here:
> >
> > #6  0x1718e0 in XactLockTableWait (xid=17334) at lmgr.c:344
> > #7  0x9e530 in heap_mark4update (relation=0xc1be62f8, tuple=0x7b03b7f0,
> >     buffer=0x7b03b828) at heapam.c:1686
> > #8  0xcb410 in LockClassinfoForUpdate (relid=387785, rtup=0x7b03b7f0,
> >     buffer=0x7b03b828, confirmCommitted=0 '\000') at index.c:1131
> > #9  0xcb534 in IndexesAreActive (relid=387785,
> confirmCommitted=1 '\001')
> >     at index.c:1176
> > #10 0xf0f04 in DefineIndex (heapRelationName=0x400aab20 "tenk1",
> >     indexRelationName=0x400aab00 "anotherj",
> accessMethodName=0x59f48 "btree",
> >     attributeList=0x400aab80, unique=0, primary=0, predicate=0x0,
> >     rangetable=0x0) at indexcmds.c:133
> > #11 0x17e118 in ProcessUtility (parsetree=0x400aaba0, dest=Remote)
> >     at utility.c:905
> >
> > Essentially it's trying to do a SELECT FOR UPDATE on the pg_class tuple
> > of the relation before it starts building the index.
> >
> > I have opined before that LockClassinfoForUpdate is a mistake that
> > shouldn't exist at all, since acquiring the proper lock on the relation
> > ought to be sufficient.
>
> As I've already mentioned many times I never agree with you.
>
> > I see no need for locking the pg_class tuple,
> > and certainly none for doing so at the beginning of the operation rather
> > than the end.
> >
> > Hiroshi, I think you defended it last time; any comments?
>
> Hmm the excluive row level lock by FOR UPDATE is too strong
> in this case. OK I would change IndexesAreActive() to not
> acquire a lock on the pg_class tuple for user tables because
> reindex doesn't need to handle relhasindex for user tables
> since 7.1.

In the end, I changed DefineIndex() to not call IndexesAreActive().

regards,
Hiroshi Inoue