> Well, it wouldn't be twenty plus transactions updating the
> same record. Concurrent transactions can't modify a single
> record. On the other hand, you might hit that limit if each
> transaction modified a different record and all the records
> happened to be on the same page. But, as I said, I haven't
> tracked through all the code to see where the limit is
> enforced.
>
> What page size are you using?

I leave all configuration parameters at their defaults. Page size is
4096 B.

Just an idea - do you think that theoretically it would help to make
the record artificially longer (ie. stuffing it by another large CHAR
field with some hard to compress value) ? Longer records could mean
less records per page, and would reduce the probabitity of the error.
Awkward solution, but in short term it can help me.

Well, I will try to identify and isolate the problem and prepare a
test case for the devel list. At least now I know where to look for
it.

Thanks again
Michal