Re: Savepoints in transactions for speed?

On Wed, Nov 28, 2012 at 8:28 PM, Jeff Davis  wrote:
>
> The main problem with a long-running delete or update transaction is
> that the dead tuples (deleted tuples or the old version of an updated
> tuple) can't be removed until the transaction finishes. That can cause
> temporary "bloat", but 1.5M records shouldn't be noticeable.

Not really that fast if you have indices (and who doesn't have a PK or two).

I've never been able to update (update) 2M rows in one transaction in
reasonable times (read: less than several hours) without dropping
indices. Doing it in batches is way faster if you can't drop the
indices, and if you can leverage HOT updates.