Re: lcr v5 - introduction of InvalidCommandId

On Wed, Sep 4, 2013 at 12:07 PM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:
> On 2013-09-03 11:40:57 -0400, Robert Haas wrote:
>> > 0002 wal_decoding: Introduce InvalidCommandId and declare that to be the new maximum for CommandCounterIncrement
>>
>> I'm still unconvinced we want this.
>
> Ok, so the reason for the existance of this patch is that currently
> there is no way to represent a "unset" CommandId. This is a problem for
> the following patches because we need to log the cmin, cmax of catalog
> rows and obviously there can be rows where cmax is unset.

For heap tuples, we solve this problem by using flag bits. Why not
adopt the same approach?

> The reason I chose to change the definition of CommandIds is that the
> other ondisk types we use like TransactionIds, XLogRecPtrs and such have
> an "invalid" type, CommandIds don't. Changing their definition to have 0
> - analogous to the previous examples - as their invalid value is not a
> problem because CommandIds from pg_upgraded clusters may never be used
> for anything. Going from 2^32 to 2^32-1 possible CommandIds doesn't seem
> like a problem to me. Imo the CommandIds should have been defined that
> way from the start.
>
> Makes some sense?

I don't have a problem with this if other people think it's a good
idea. But I think it needs a few +1s and not too many -1s first, and
so far (AFAIK) no one else has weighed in with an opinion.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company