Re: Fast index build vs. PITR

On Tue, 2004-06-01 at 03:13, Alvaro Herrera wrote:
> A completely different idea would be to log a "logical index creation",
> so that during normal recovery those entries are saved somewhere; after
> the rest of WAL recovery is done, the system is taken into a more normal
> post-recovery pre-usable state, on which those indexes are recreated
> from user data. This would be cheapest in WAL traffic, but probably
> it'll also require more code and new hooks in the startup mechanism.
> Also, it'd require examining later WAL entries that refer to the index
> and act accordingly (e.g. ignore the entry if it modifies the index, and
> forget the creation if it's a DROP INDEX command.)
>

There will be many ways to optimise recovery once we have PITR
working...

The current code does a straight replay of all changes. We can imagine
lots of different multi-pass or lookahead strategies for replaying xlog
records, but please lets wait awhile...

> Not that I like neither of those ideas really ... issuing normal WAL
> index creation traffic if PITR is active is certainly the easiest way.

I agree, certainly for now.

Best Regards, Simon Riggs