On 22.11.21 01:47, Tomas Vondra wrote:
So I think just decoding the sequence tuples is a better solution - for large transactions (consuming many values from the sequence) it may be more expensive (i.e. send more records to replica). But I doubt that matters too much - it's likely negligible compared to other data for large transactions.
I agree that the original approach is better. It was worth trying out this alternative, but it seems quite complicated. I note that a lot of additional code had to be added around several areas of the code, whereas the original patch really just touched the logical decoding code, as it should. This doesn't prevent anyone from attempting to optimize things somehow in the future, but for now let's move forward with the simple approach.
