Re: Ryu floating point output patch

Hi,

On 2019-01-11 13:33:54 -0500, Robert Haas wrote:
> On Fri, Jan 11, 2019 at 1:30 PM Andres Freund <andres@anarazel.de> wrote:
> > On 2019-01-11 12:55:54 -0500, Robert Haas wrote:
> > > About as much as I can say is that if you commit this patch, and as a
> > > result, a dump-and-restore from v11 to v12 causes a change in the bits
> > > on disk that would not have occurred without this patch, that's very
> > > bad.  What ramifications there will be if the output changes in ways
> > > that don't affect the bits that get written to the platter -- I don't
> > > know.
> >
> > Are you trying to highlight the potential dangers of this patch, or are
> > you seeing a concrete risk of that in the discussion? Both ryu's and the
> > current output (as used by pg_dump, with extra_float_digits = 3) ought
> > to be roundtrip safe.  There's possibly some chance for differences in
> > insignificant bits, but that ought to be pretty much it?
> >
> > The current discussion would probably change the data that's stored on disk
> > between ryu/non-ryu for queries like
> > CREATE TABLE AS SELECT floatcol::text::float FROM ...
> > or
> > CREATE TABLE AS SELECT floatcol::text FROM ...
> >
> > when not specifyiung extra_float_digits = 3, but the ryu case would be
> > more precise, so that ought to be good?
> 
> I'm not complaining about the patch.  I'm saying that the patch can't
> change people's existing data when they dump and restore.  If you're
> saying it doesn't, cool.  I don't mind if the answers to queries get
> more precise -- that's good, as you say -- just that we don't change
> their data.

It'd potentially change data between an non-ryu->{ryu,non-ryu} and
ryu->ryu versions, for dumps that didn't specify extra_float_digits =
3. But that'd not be pg_dump, and already broken previously, so I don't
have particularly much sympathy?

And of course it'd change the dump's text contents between ryu and
non-ryu backends even with extra_float_digits = 3, but the resulting
floats ought to be the same. It's just that ryu is better at figuring
out what the minimal text representation is than the current code.

Greetings,

Andres Freund