use liboffsets to_offset function

[jbrockmendel]

No description provided.

[codecov]

Codecov Report

Merging #21693 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master   #21693   +/-   ##
=======================================
  Coverage    91.9%    91.9%           
=======================================
  Files         154      154           
  Lines       49657    49657           
=======================================
  Hits        45638    45638           
  Misses       4019     4019

Flag	Coverage Δ
#multiple	90.28% <ø> (ø)	⬆️
#single	41.9% <ø> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update dc45fba...b2fa7cb. Read the comment docs.

[jreback]

thanks!

[qwhelan]

While investigating general Period slowness, an asv find bisection identified this PR as introducing a ~56x slowdown on Windows 10:

asv find v0.23.4..24ab22f75a groupby.Datelike.time_sum
[...]
·· For pandas commit dc45fbaf <master~690>:
·· Uninstalling from conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
·· Building dc45fbaf <master~690> for conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
·· Installing dc45fbaf <master~690> into conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
·· Benchmarking conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
··· Running (groupby.Datelike.time_sum--).
··· groupby.Datelike.time_sum                                                                                                                                                   ok
··· =============== ==========
        grouper
    --------------- ----------
      period_range   46.9±8ms
       date_range    3.12±0ms
     date_range_tz   3.91±1ms
    =============== ==========

· Testing ------------O-----------------------------------------------
·· For pandas commit d105ce26 <master~689>:
·· Uninstalling from conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
·· Building d105ce26 <master~689> for conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
·· Installing d105ce26 <master~689> into conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
·· Benchmarking conda-py3.6-Cython-matplotlib-numexpr-numpy-openpyxl-pytables-pytest-scipy-sqlalchemy-xlrd-xlsxwriter-xlwt
··· Running (groupby.Datelike.time_sum--).
··· groupby.Datelike.time_sum                                                                                                                                                   ok
··· =============== ===========
        grouper
    --------------- -----------
      period_range   2.63±0.3s
       date_range     3.12±0ms
     date_range_tz    3.12±0ms
    =============== ===========

· Greatest regression found: d105ce26 <master~689>

My suspicion is the moving of from pandas.tseries.frequencies import to_offset from period.pyx module scope into offset.pyx in function scope is leading to this slowdown on Windows.

[jbrockmendel]

Thanks for tracking this down. A couple questions: is there anything weird about your Windows environment, e.g. running on top of cygwin or something? Or more generally since I don't use Windows, any other reason why your system might be special? Long-shot since you're already doing a lot here: can you reproduce these results on non-Windows?

If you re-run the same asv a few times, are the results pretty consistent? Often asv results can have a lot of noise.

[qwhelan]

• Nothing unusual from a Windows perspective: relatively clean Windows 10 install with benchmarks run under PowerShell. Only standard antivirus, with benchmarks excluded from AV checks.
• None that come to mind, actually did a fresh install a few months back due to upgrading the SSD and strictly used this machine for remote desktop until recently (so very little installed on it).
• Actually, it also looks like the trend is visible on my Ubuntu 18.04 machine, so seems like it might be a cross-platform issue. Haven't run a bisection yet, but could this evening.

And yes, sadly consistent.

[jbrockmendel]

@jreback I thought we had made the "pxdfiles" obsolete when moving to cythonize. Am I remembering wrong?

[jbrockmendel]

Never mind, thought I was commenting on an open PR, not an old one.