Fix PyPy compatibility in adbapi init

[Valodim]

Following up to #2760 (and with #3098 fixed), this PR adds the simple required fix for the database initialization with adbapi to be able to run on pypy again.

I also added pypy35 to the tox and travis configs - let's see if that works :)

Signed-Off-By: Vincent Breitmoser look@my.amazin.horse

[Valodim]

Looks like the pypy35 test went through equally well as the python35 one.

But pypy35-postgres caused a bunch of errors, mostly "too many connections" on the psql server? The pypy tests also ran significantly longer than the py3 ones, not sure what to make of that :) On the plus side though, setting up tox for pypy was straightforward and worked almost right out of the box, yay

[luminoso]

pypy tests also ran significantly longer than the py3 ones

I think this is expected due to JIT compilations. However, it should run faster after that

[richvdh]

please see CONTRIBUTING.rst - this needs basing on develop, and needs sign-off.

[Valodim]

Well. The test runs work now, but the pypy-postgres run takes a whoopin' 43 minutes for some reason. Wonder what the problem there is... anyways, I'll leave this here for now :)

[richvdh]

we're excluding python 3.6 deliberately: we get four workers on Travis, so this one just slows things down pointlessly.

[Valodim]

ah, sorry. I think this one snuck during the rebase, didn't mean to add it

[richvdh]

the pypy-postgres run takes a whoopin' 43 minutes

Well, that sounds like a show-stopper :-p

[Valodim]

Sure does. Even the regular test takes 2.5 times as long :(

There's three ways from here:

• figure out why it's slow and fix it :) heh. maybe if someone comes along who knows a lot about pypy
• keep only the pypy-sqlite test around. it's not prohibitively slow but should still catch most pypy breakage
• comment out the pypy tests in the CI pipeline, keeping pypy support in its current stepmotherly state. given that pypy's primary selling point is performance, but it's slower by a factor two in these (inadequate) tests, this is still a reasonable option

[luminoso]

@Valodim give it a try running the pypy tests with "--jit off" to check if JIT compilations are the ones causing the tests taking so long

[richvdh]

We can't really take this while it makes the CI runs take over 45 minutes. I suggest you remove the postgres job from travis.yml, and we can investigate why it is so slow elsewhere.

[luminoso]

Any news about this?

[richvdh]

seems to have stalled, so closing for now. We can reopen it if anyone wants to continue work.