Investigate exposing Arrow StringBuilder to numba

[chmp]

The numba implementation should be much slower, than arrow's StringBuilder.

Question: what can numba inteface with? The docs mention cffi.

[xhochy]

What is the use case where we need to the builder in Numba?

[chmp]

Whenever you want to return an array of strings and don't know what size the result will be.

[xhochy]

Seems like it is not easily possible to call C++ code from Numba. We have a GLib-based C-API in Arrow: https://github.com/apache/arrow/blob/master/c_glib/arrow-glib/array-builder.h#L891 This may be useful here.

[alendit]

I've looked into it recently: the fletcher's NumbaStringArrayBuilder is around 10 times slower than Arrow's StringBuilder for code like this for the current version of numba:

@numba.njit(nogil=True)
def build_string(chars):
   sb = fr._numba_compat.NumbaStringArrayBuilder(10 ** 6, 10 ** 7)
   for str_idx in range(0, 10 ** 6):
      for c_idx in range(0, 10):
         sb.put_byte(chars[str_idx * 10 + c_idx])
      sb.finish_string()
   sb.finish()

This comes from the issues with inlining and optimization of numba's jitclasses.

Good news is that I'm working on a set of patches which will bring down the runtime of the code above from 400ms to 40ms on my machine, i.e. on par wit this C++ code. The relevant issues are numba/numba#2166, numba/numba#3305. This PR fixes the inlining numba/numba#3531. This PR numba/llvmlite#429 to llvmlite will allow linking custom optimization passes, and finally there is local code which optimizes ref counting across the CFG.

TL;DR It can take a while, but there is not reason why numba wouldn't be as fast as native C++ for this task.

[xhochy]

@alendit That sounds promising. Thank you for taking a look into this!

[xhochy]

This project has been archived as development has ceased around 2021.
With the support of Apache Arrow-backed extension arrays in pandas, the major goal of this project has been fulfilled.