Bruce Momjian <pgman@candle.pha.pa.us> writes:
> I should add that I am not prepared to overhaul the pg_clog file format
> as part of adding subtransactions for 7.4. I can do the tid/sequential scan
> method for abort, or the single-lock method described.
If you think that changing the pg_clog file format would be harder than
either of those other ideas, I think you're very badly mistaken.
pg_clog is touched only by one rather simple module.
I think the other methods will be completely unacceptable from a
performance point of view. They could maybe work if subtransactions
were a seldom-used feature; but the people who want to use 'em are
mostly talking about a subtransaction for *every* command. If you
design your implementation on the assumption that subtransactions are
infrequent, it will be unusably slow.
regards, tom lane