On Sat, Sep 21, 2019 at 10:52 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> would work. The naming convention I was putting in didn't allow for
> identical DB and user names, and it didn't seem to me that it was
> worth overriding that convention to keep this particular test
> result the same. The error output still proves that ecpg *tried*
> to connect to the database named the same as the user, which is
> the only real point of this test case as compared to its siblings.
> And why not test the error path, anyway?
Seems reasonable to me.
> > I'm asking about this because I find that "make installcheck-world"
> > sometimes fails at this point when run against my local installation
> > -- the test "test connect/test" reliably fails.
>
> Um, you're confusing me ... is it "sometimes", or is it "reliably"?
It's reliable. Doesn't have to be a Valgrind build, or even a debug build.
> It does seem like there might be something to poke at here. We never
> did figure out quite what the other reporter was seeing.
>
> Given the lack of overlap of user and database names, we can be
> pretty darn certain that the test case shouldn't accidentally
> succeed due to timing issues. I wonder if there could be some kind
> of uninitialized-variable bug inside ecpg that lets it sometimes
> connect to the wrong DB name (perhaps one it'd previously connected
> to) if the DB name is omitted from the command?
That seems possible. IIRC Valgrind has certain limitations when it
comes to detecting stack corruption.
--
Peter Geoghegan
