On Sat, Apr 2, 2016 at 11:41 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
"Shulgin, Oleksandr" <oleksandr.shulgin@zalando.de> writes: > On Mon, Mar 14, 2016 at 7:55 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: >> Yeah, I don't much like that either. But I don't think we can avoid >> some refactoring there; as designed, conversion of an error message into >> user-visible form is too tightly tied to receipt of the message.
> True. Attached is a v2 which addresses all of the points raised earlier I > believe.
I took a closer look at what's going on here and realized that actually it's not that hard to decouple the message-building routine from the PGconn state, because mostly it works with fields it's extracting out of the PGresult anyway. The only piece of information that's lacking is conn->last_query. I propose therefore that instead of doing it like this, we copy last_query into error PGresults. This is strictly less added storage requirement than storing the whole verbose message would be, and it saves time compared to the v2 patch in the typical case where the application never does ask for an alternately-formatted error message. Plus we can actually support requests for any variant format, not only VERBOSE.
Attached is a libpq-portion-only version of a patch doing it this way. I've not yet looked at the psql part of your patch.
Comments?
Ah, neat, that's even better. :-)
What about regression tests? My assumption was that we won't be able to add them with the usual expected file approach, but that we also don't need it that hard. Everyone's in favor?
