COMPAT: Ignore numexpr <= v2 (at least for now)

[jtratner]

Earlier versions of numexpr don't work with pandas because of API changes. For now we should not use numexpr if version too low and then revisit whether we can easily support it later.

[ghost]

Also should raise warning when using query/eval that require numexpr. #5969.

[cpcloud]

Hm .. _check_engine should raise an ImportError if numexpr can't be imported

[cpcloud]

that's the second thing that happens when you call pd.eval. Let me try and repro #5969

[jreback]

@cpcloud let's create a sep issue for query/eval (and do a PR for 0.13.1)

[cpcloud]

you mean for #5969?

[jreback]

yep...I think #5969 got closed , but the non-installed numexpr was causing it (or can just reopen that one), and see if you can repro

[cpcloud]

ok repopening ...

[cpcloud]

Just as a note to self:

Somehow, without numexpr installed, eval plugs along until the AttributeError thrown by trying to call numexpr.evaluate is raised

[jreback]

@jtratner / @cpcloud going to be able to get to this?

[cpcloud]

i can do this weekend

[jtratner]

Yes, just need to figure out why this one test fails.

[jreback]

@jtratner / @cpcloud hows this coming?

[cpcloud]

trying to repro now ...

[cpcloud]

@jreback So for frame/series ops check ne version and ignore it and for eval raise?

[cpcloud]

@jtratner Do you want me to take this?

[jreback]

I think just have to make _NUMEXPR_INSTALLED=False if the version is say < 2.0 (in compuatation/expressions.py, then a similar check in computation/eval.py where you try to import numexpr

so yes I would raise if you try to eval, but just falls back like for evaluate)

[cpcloud]

i don't think it should fall back for eval ... i think it should raise .. since that would make it use the python engine if the numexpr version is too low and the user will never know

[cpcloud]

oh i c ... that's wat u said .. .misread it

[jreback]

@cpcloud I wouldn't even attempt to use the Python eval...just raise the ImportError (for pd.eval/df.query)
no point....let's make it simple, I think @jtratner was trying to fall back, but not necessary (and for .evaluate in expressions, it already falls back gracefully).

In fact you can just import computation/expressions and use _NUMEXPR_INSTALLED and done

[cpcloud]

I can't repro @twiecki's issue ... tried with all past versions of numexpr ...

[cpcloud]

that said, the truediv keyword isn't supported for v < 2.0 so we still have to do this for eval

[jreback]

ok...let's just add an install of say numexpr 1.4 on say the 2.6 build...and just have the tests check that eval/query raise....no biggie

yeh < 2.0 should be effectively not installed

if yu want to get fancy you can raise a different message and say your numexpr is too old...(and of course if the import fails entirely say you need it)

[cpcloud]

also i have gotten similar weirdness with modules importing but not having certain attributes before. some issue with importing pkg_resources ... pytz had or has this issue sometimes... i used to get it with vbench

[jreback]

yep....sort of like import but require this version, otherwise raise...I think you have to do it sort of manually

[jreback]

ok...you are doing a new PR for this I assume? going to close @jtratner one....(as I think he tried to refactor a lot)

[jreback]

@cpcloud ref/close these #5857, #5855, #5845.

[cpcloud]

le checklist:

• numexpr 1.4 on 2.6 build, check for raising in eval query
• require version on expressions/evaluate (ndframe ops)

[jtratner]

heh, I just got this to work I think.

[jreback]

#6109 looks fine..and ready to go...don't worry about it...