?hel kenal p?eval, N, 2006-01-26 kell 15:07, kirjutas Christopher
Browne:
> Now that version 1.1.5 has been released, it now makes sense to look
> at what features need to get in in order to get version 1.2 to being
> "release-ready."
> 
> A number of things complained about were consciously left alone
> because, in the "stable" stream, we should (arguably) primarily be
> fixing specific engine problems being experienced that _don't_ point
> to major reimplementation of things.
> 
> In order not to invalidate RPM packaging work, for instance, we didn't
> take on some of the build environment changes that got suggested,
> notably cleaning up the usage of --prefix and other configure-time
> parameters.  It is now "fair game" to do so; we'll presumably break
> the RPMs, but that's necessary to clean things up.  Bugs #1507, 1508,
> 1509, 1510, 1497, 1473, 1475 all seem to be relevant to the build
> environment.
> 
> Other things likely on the 1.2 "track":
> 
> - The stored function that regenerates sl_listen needs revising
>   (#1485)
> 
> - EXECUTE SCRIPT will fail if an early query creates columns on
>   tables, and then later queries manipulate those columns
> 
> - Rotating periodically between sl_log_1 and sl_log_2 will resolve
>   some issues relating to XID rollover

Some but not all :(

For example if a cluster has several nodes producing data, and they are
tunneled through a single node, then the xid-wraparound will still break
btree indexes and you are still losing data.

Perhaps we should move to storing xid's as plain bigints and fix
snapshot type accordingly. (Marko Kreen has done the 8byte-xid (with
wraparound detection) and 8byte-snapshot part for one of our internal
projects).

Btw, what is the reasoning behind not storing full snapshots in sl_event
as its own snapshot type? why is it broken down into 3 fields (xxid,
xxid, text) ?

Perhaps a general DONT.DO document should also be compiled and included
with slony documenting more common ways to shoot yourself in the foot.

----------------
Hannu