string getitem methods are slow

[hayd]

related #2802

It seems that str[1] is significantly slower than .apply(lambda x: x[1])

See this So answer http://stackoverflow.com/a/18473330/1240268

[cpcloud]

Couple of reasons:

1. the mapped function is actually lambda x: x[i] if len(x) > i else np.nan
2. isnull is called to compute a mask for mapping over in lib.map_infer_mask (which is in inference.pyx)

looks like the perf hit is about 2x, might be able to squash that by moving string methods to cython

[hayd]

Ah, that'll do it (I guess it's only sometimes apply doesn't care about errors?).

maybe cythonizing these is the way forward, I guess even with object dtype you get some perf improvement.

[jreback]

these methods could be much faster (there is an issue out there about this) if you basically push everything to use native c calls (eg stuff like strcmp and such) or maybe add a nice c library in the mix

just cythonizing doesn't help much

but this would be a bit of work

[cpcloud]

wonder if this is worth looking into: http://bstring.sourceforge.net/

[jtratner]

If you're going to c level, better to use a c library that handles strings
/ unicode for you so we don't have to worry as much about all the gotchas
with c strings.

[cpcloud]

darn bstring doesn't support unicdoe

[cpcloud]

Converting these functions to C without breakage is going to be very difficult. You'll probably have to use ICU and have a compatibility layer between Cython (PyICU might make this a bit easier) and ICU.

We definitely cannot use the C standard library string functions since they don't handle Unicode.

[brandon-rhodes]

Is anyone working on this currently? Would I be duplicating effort if I were to look into possible quick wins for at least getting these slow str routines a bit faster?

[jreback]

@brandon-rhodes don't think so. would be great!

prob DO need some asv benchmarks for these.