On 4/2/22 12:43, Amit Kapila wrote:
> On Sat, Apr 2, 2022 at 5:47 AM Tomas Vondra
> <tomas.vondra@enterprisedb.com> wrote:
>>
>> On 4/1/22 17:02, Tomas Vondra wrote:
>>
>> The only option I see is reworking the decoding so that it does not need
>> the snapshot at all. We'll need to stash the changes just like any other
>> change, apply them at end of transaction, and the main difference
>> between transactional and non-transactional case would be what happens
>> at abort. Transactional changes would be discarded, non-transactional
>> would be applied anyway.
>>
>
> I think in the above I am not following how we can make it work
> without considering *snapshot at all* because based on that we would
> have done the initial sync (copy_sequence) and if we don't follow that
> later it can lead to inconsistency. I might be missing something here.
>
Well, what I meant to say is that we can't consider the snapshot at this
phase of decoding. We'd still consider it later, at commit/abort time,
of course. I.e. it'd be fairly similar to what heap_decode/DecodeInsert
does, for example. AFAIK this does not build the snapshot anywhere.
>> The challenge is this reorders the sequence changes, so we'll need to
>> reconcile them somehow. One option would be to simply (1) apply the
>> change with the highest LSN in the transaction, and then walk all other
>> in-progress transactions and changes for that sequence with a lower LSN.
>> Not sure how complex/expensive that would be, though. Another problem is
>> not all increments are WAL-logged, of course, not sure about that.
>>
>> Another option might be revisiting the approach proposed by Hannu in
>> September [1], i.e. tracking sequences touched in a transaction, and
>> then replicating the current state at that particular moment.
>>
>
> I'll think about that approach as well.
>
Thanks!
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
