I have talked in the past about a possible implementation of
savepoints/nested transactions. I would like to more formally outline
my ideas below.
We have talked about using WAL for such a purpose, but that requires WAL
files to remain for the life of a transaction, which seems unacceptable.
Other database systems do that, and it is a pain for administrators. I
realized we could do some sort of WAL compaction, but that seems quite
complex too.
Basically, under my plan, WAL would be unchanged. WAL's function is
crash recovery, and it would retain that. There would also be no
on-disk changes. I would use the command counter in certain cases to
identify savepoints.
My idea is to keep savepoint undo information in a private area per
backend, either in memory or on disk. We can either save the
relid/tids of modified rows, or if there are too many, discard the
saved ones and just remember the modified relids. On rollback to save
point, either clear up the modified relid/tids, or sequential scan
through the relid and clear up all the tuples that have our transaction
id and have command counters that are part of the undo savepoint.
It seems marking undo savepoint rows with a fixed aborted transaction id
would be the easiest solution.
Of course, we only remember modified rows when we are in savepoints, and
only undo them when we rollback to a savepoint. Transaction processing
remains the same.
There is no reason for other backend to be able to see savepoint undo
information, and keeping it private greatly simplifies the
implementation.
-- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610)
853-3000+ If your life is a hard drive, | 830 Blythe Avenue + Christ can be your backup. | Drexel Hill,
Pennsylvania19026
