On 23.12.21 12:40, Fabien COELHO wrote:
>> In [0], it was reported that certain replication commands result in
>> infinite loops because of faulty error handling. This still happens.
>> I wrote a test for it, attached here. (I threw in a few more basic
>> tests, just to have some more coverage that was lacking, and to have a
>> file to put the new test in.)
>
> Hmmm… For some unclear reason on errors on a PGRES_COPY_* state
> PQgetResult keeps on returning an empty result. PQexec manually ignores
> it, so I did the same with a comment, but for me the real bug is somehow
> in PQgetResult behavior…
>
>> In [1], it was reported that server crashes produce duplicate error
>> messages. This also still happens. I didn't write a test for it,
>> maybe you have an idea. (Obviously, we could check whether the error
>> message is literally there twice in the output, but that doesn't seem
>> very general.) But it's easy to test manually: just have psql
>> connect, shut down the server, then run a query.
>
> This is also a feature/bug of libpq which happens to be hidden by
> PQexec: when one command crashes PQgetResult actually returns *2*
> results. First one with the FATAL message, second one when libpq figures
> out that the connection was lost with the second message appended to the
> first. PQexec just happen to silently ignore the first result. I added a
> manual reset of the error message when first shown so that it is not
> shown twice. It is unclear to me whether the reset should be somewhere
> in libpq instead. I added a voluntary crash at the end of the psql test.
I agree that these two behaviors in libpq are dubious, especially the
second one. I want to spend some time analyzing this more and see if
fixes in libpq might be appropriate.