Re: [BUG] Logical replication failure "ERROR: could not map filenode "base/13237/442428" to relation OID" with catalog modifying txns

On Tue, Jul 12, 2022 at 10:28 AM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
>
> On Tue, Jul 12, 2022 at 9:48 AM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >
> > On Fri, Jul 8, 2022 at 8:20 PM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> > >
> > > On Fri, Jul 8, 2022 at 5:59 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > > >
> > > > On Fri, Jul 8, 2022 at 12:46 PM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> > > > >
> > > > > On Fri, Jul 8, 2022 at 3:27 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > > > > >
> > > > >
> > > > > > 1.
> > > > > > In ReorderBufferGetCatalogChangesXacts(), isn't it better to use the
> > > > > > list length of 'catchange_txns' to allocate xids array? If we can do
> > > > > > so, then we will save the need to repalloc as well.
> > > > >
> > > > > Since ReorderBufferGetcatalogChangesXacts() collects all ongoing
> > > > > catalog modifying transactions, the length of the array could be
> > > > > bigger than the one taken last time. We can start with the previous
> > > > > length but I think we cannot remove the need for repalloc.
> > > > >
> > > >
> > > > It is using the list "catchange_txns" to form xid array which
> > > > shouldn't change for the duration of
> > > > ReorderBufferGetCatalogChangesXacts(). Then the caller frees the xid
> > > > array after its use. Next time in
> > > > ReorderBufferGetCatalogChangesXacts(), the fresh allocation for xid
> > > > array happens, so not sure why repalloc would be required?
> > >
> > > Oops, I mistook catchange_txns for catchange->xcnt. You're right.
> > > Starting with the length of catchange_txns should be sufficient.
> > >
> >
> > I've attached an updated patch.
> >
> > While trying this idea, I noticed there is no API to get the length of
> > dlist, as we discussed offlist. Alternative idea was to use List
> > (T_XidList) but I'm not sure it's a great idea since deleting an xid
> > from the list is O(N), we need to implement list_delete_xid, and we
> > need to make sure allocating list node in the reorder buffer context.
> > So in the patch, I added a variable, catchange_ntxns, to keep track of
> > the length of the dlist. Please review it.
> >
>
> I'm doing benchmark tests and will share the results.
>

I've done benchmark tests to measure the overhead introduced by doing
bsearch() every time when decoding a commit record. I've simulated a
very intensified situation where we decode 1M commit records while
keeping builder->catchange.xip array but the overhead is negilible:

HEAD: 584 ms
Patched: 614 ms

I've attached the benchmark script I used. With increasing
LOG_SNAPSHOT_INTERVAL_MS to 90000, the last decoding by
pg_logicla_slot_get_changes() decodes 1M commit records while keeping
catalog modifying transactions.

Regards,

--
Masahiko Sawada
EDB: https://www.enterprisedb.com/