Christoph Berg <christoph.berg@credativ.de> writes:
> A customer recently upgraded their jdbc driver from 8.4 to 9.2. This
> enabled the binary wire protocol (effectively between 9.1 and 9.2).
> They reported that single precision values inserted into a
> numeric(10,2) column were suddenly rounded wrongly, i.e. 10000.18 was
> inserted as 10000.20, while that worked before. Of course we told them
> that single is the wrong data type for this, but still, this is a
> regression.
I'm not sure that it's fair to characterize that as a regression.
If anything, it's more sensible than what happened before.
> But if extra_float_digits > 0 is set, I'd expect not only the float4
> output to be affected by it, but also casts to other datatypes,
This proposal scares me. extra_float_digits is strictly a matter of
I/O representation, it does not affect any internal calculations.
Moreover, since one of the fundamental attributes of type numeric
is that it's supposed to give platform-independent results, I don't
like the idea that you're likely to get platform-dependent results
of conversions from float4/float8.
I think your customer got bit by his own bad coding practice, and
that should be the end of it.
regards, tom lane