On Fri, Mar 4, 2011 at 8:20 AM, Robert Haas <robertmhaas@xxxxxxxxx> wrote: > On Fri, Mar 4, 2011 at 4:21 AM, Matt Burke <mattblists@xxxxxxxxxxxxx> wrote: >> Robert Haas wrote: >>> Old row versions have to be kept around until they're no longer of >>> interest to any still-running transaction. >> >> Thanks for the explanation. >> >> Regarding the snippet above, why would the intermediate history of >> multiply-modified uncommitted rows be of interest to anything, or is the >> current behaviour simply "cheaper" overall in terms of cpu/developer time? > > Because in theory you could have a cursor open. You could open a > cursor, start to read from it, then make an update. Now the cursor > needs to see things as they were before the update. > > We might be able to do some optimization here if we had some > infrastructure to detect when a backend has no registered snapshots > with a command-ID less than the command-ID of the currently active > snapshot, but nobody's put in the effort to figure out exactly what's > involved and whether it makes sense. It's a problem I'm interested > in, but #(needed round-tuits) > #(actual round-tuits). Not just cursors, but pl/pgsql for example is also pretty aggressive about grabbing snapshots. Also,t' is a matter of speculation if the case of a single row being updated a high number of times in a single transaction merits such complicated optimizations. It bears repeating: Explicit transaction control (beyond the dblink type hacks that currently exist) in backend scripting would solve many cases where this is a problem in practice without having to muck around in the mvcc engine. Autonomous transactions are another way to do this... merlin -- Sent via pgsql-performance mailing list (pgsql-performance@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
