Alvaro Herrera <alvherre@dcc.uchile.cl> writes:
> On Fri, May 28, 2004 at 03:19:29PM -0400, Tom Lane wrote:
>> We'd still need a plain CommandCounterIncrement facility, which means
>> that actually a subtransaction would have to be a group of CIDs not just
>> one.
> Right, this is why I suggested runlength (the group is contiguous).
Not necessarily.
> Right. We only need to store the "borders" though. Not even that: only
> the start, because the end is what is current at AbortSubTransaction()
> time.
Nope. Think about sub-subtransactions.
A runlength encoding might be worth using, though, since you're right
that subxacts would tend to get runs of consecutive CIDs.
In theory we could also use a runlength representation for the master
bitmap of aborted CIDs, but that would be costly since you would have to
do a search, and not just a trivial index, to check the state of a
particular CID. My inclination would be to stick with a bitmap for now.
It'd be easy enough to revise the implementation later if that gut feel
proves out wrong.
regards, tom lane