Re: PATCH: logical_work_mem and logical streaming of largein-progress transactions - Mailing list pgsql-hackers

On Thu, Jan 9, 2020 at 12:09 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
>
> On Thu, Jan 9, 2020 at 10:30 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:
> >
> > On Thu, Jan 9, 2020 at 9:35 AM Amit Kapila <amit.kapila16@gmail.com> wrote:
> > >
> > > On Wed, Jan 8, 2020 at 1:12 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:
> > > >
> > > > I have observed one more design issue.
> > > >
> > >
> > > Good observation.
> > >
> > > >  The problem is that when we
> > > > get a toasted chunks we remember the changes in the memory(hash table)
> > > > but don't stream until we get the actual change on the main table.
> > > > Now, the problem is that we might get the change of the toasted table
> > > > and the main table in different streams.  So basically, in a stream,
> > > > if we have only got the toasted tuples then even after
> > > > ReorderBufferStreamTXN the memory usage will not be reduced.
> > > >
> > >
> > > I think we can't split such changes in a different stream (unless we
> > > design an entirely new solution to send partial changes of toast
> > > data), so we need to send them together. We can keep a flag like
> > > data_complete in ReorderBufferTxn and mark it complete only when we
> > > are able to assemble the entire tuple.  Now, whenever, we try to
> > > stream the changes once we reach the memory threshold, we can check
> > > whether the data_complete flag is true, if so, then only send the
> > > changes, otherwise, we can pick the next largest transaction.  I think
> > > we can retry it for few times and if we get the incomplete data for
> > > multiple transactions, then we can decide to spill the transaction or
> > > maybe we can directly spill the first largest transaction which has
> > > incomplete data.
> > >
> > Yeah, we might do something on this line.  Basically, we need to mark
> > the top-transaction as data-incomplete if any of its subtransaction is
> > having data-incomplete (it will always be the latest sub-transaction
> > of the top transaction).  Also, for streaming, we are checking the
> > largest top transaction whereas for spilling we just need the larget
> > (sub) transaction.   So we also need to decide while picking the
> > largest top transaction for streaming, if we get a few transactions
> > with in-complete data then how we will go for the spill.  Do we spill
> > all the sub-transactions under this top transaction or we will again
> > find the larget (sub) transaction for spilling.
> >
>
> I think it is better to do later as that will lead to the spill of
> only required (minimum changes to get the memory below threshold)
> changes.
I think instead of doing this can't we just spill the changes which
are in toast_hash.  Basically, at the end of the stream, we have some
toast tuple which we could not stream because we did not have the
insert for the main table then we can spill only those changes which
are in tuple hash.  And, in the subsequent stream whenever we get the
insert for the main table at that time we can restore those changes
and stream.  We can also maintain some flag saying data is not
complete (with some change LSN number) and after that LSN we can spill
any toast change to disk until we get the change for the main table,
by this way we can avoid building tuple hash until we get the change
for the main table.

-- 
Regards,
Dilip Kumar
EnterpriseDB: http://www.enterprisedb.com