Re: Replication vs. float timestamps is a disaster

On Sun, Feb 19, 2017 at 9:19 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Robert Haas <robertmhaas(at)gmail(dot)com> writes:
>> On Sun, Feb 19, 2017 at 3:31 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>> Thoughts? Should we double down on trying to make this work according
>>> to the "all integer timestamps" protocol specs, or cut our losses and
>>> change the specs?
>
>> I vote for doubling down. It's bad enough that we have so many
>> internal details that depend on this setting; letting that cascade
>> into the wire protocol seems like it's just letting the chaos spread
>> farther and wider.
>
> How do you figure that it's not embedded in the wire protocol already?
> Not only the replicated data for a timestamp column, but also the
> client-visible binary I/O format, depend on this. I think having some
> parts of the protocol use a different timestamp format than other parts
> is simply weird, and as this exercise has shown, it's bug-prone as all
> get out.

You've got a point, but if we don't make the replication protocol
consistently use integers, then every client that cares about those
fields has to be able to handle either format, or at least detect that
it got a different format than it was expecting and fail cleanly.
How's that going to work?

Also, I feel like there's some difference between the data stored in a
cluster and the metadata which describes the cluster. Making the
interpretation of the metadata depend on the cluster configuration
feels circular, like using non-ASCII characters in the name of an
encoding.

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