On 2021-Jul-30, Amit Kapila wrote:
> I was thinking of using toast pointer but that won't work because it
> can be different on the subscriber-side. I don't see any better ideas
> to fix this issue. This problem seems to be from the time Logical
> Replication has been introduced, so adding others (who are generally
> involved in this area) to see what they think about this bug? I think
> people might not be using toasted columns for Replica Identity due to
> which this problem has been reported yet but I feel this is quite a
> fundamental issue and we should do something about this.
In the evening before going offline a week ago I was looking at this and
my conclusion was that this was a legitimate problem: the original
implementation is faulty in that the full detoasted value is required to
be transmitted in order for downstream to be able to read the value.
I am not sure if at the level of logical decoding it is a problem
theoretically, but at least for logical replication it is clearly a
practical problem.
Reading Dilip's last posted patch that day, I had some reservations
about the API of ExtractReplicaIdentity. The new argument makes for a
very strange to explain behavior "return the key values if they are
unchanged, *or* if they are toasted" ... ??? I tried to make sense of
that, and tried to find a concept that would make sense to the whole,
but couldn't find any obvious angle in the short time I looked at it.
I haven't looked at it again.
--
Álvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/
"El Maquinismo fue proscrito so pena de cosquilleo hasta la muerte"
(Ijon Tichy en Viajes, Stanislaw Lem)