PERF: regression from 0.10.1

[jreback]

f6ac8c0 to 05a737d

Bad news:

Results:
                                            t_head  t_baseline      ratio
name                                                                     
reindex_frame_level_reindex                 1.1343      0.7571     1.4982
reindex_frame_level_align                   1.1690      0.7548     1.5488

Good news:

Results:
                                            t_head  t_baseline      ratio
name                                                                     
frame_get_dtype_counts                      0.0968    208.6380     0.0005
frame_wide_repr                             0.5321    207.0730     0.0026
groupby_first_float32                       3.0386    336.2439     0.0090
groupby_last_float32                        3.2206    335.2571     0.0096
frame_to_csv2                             181.1550   2188.9260     0.0828
indexing_dataframe_boolean                 12.9179    125.7350     0.1027
write_csv_standard                         36.3702    230.9210     0.1575
frame_reindex_axis0                         0.3122      1.0984     0.2842
frame_to_csv_mixed                        340.1070   1091.9340     0.3115
frame_to_csv                              106.5671    219.1830     0.4862
frame_add                                  22.0693     34.1867     0.6456
frame_mult                                 22.0064     33.9859     0.6475
frame_reindex_upcast                       11.4179     17.3291     0.6589
frame_fancy_lookup_all                     14.3931     20.5621     0.7000
series_string_vector_slice                148.3371    195.5922     0.7584

???

frame_reindex_axis1                         2.3612      2.3386     1.0097

Fixed:

indexing_dataframe_boolean_rows             0.2180      0.1953     1.1164

[jreback]

I think that the multi-level reindexing ones are possibly related to the take changes (#2819)
(which help a great deal in the most common cases), I didn't investigate much, that's just the top
on profiling (in both old and new)

bisected?

vbench was meant exactly for that, but it hasn't been updated in a while.
Too bad.

http://pandas.pydata.org/pandas-docs/vbench

[jreback]

can we run it on each commit and get the graph?

That's it's reason for being. I just added test_perf a few months ago for the devs.

[jreback]

agreed....how does it update though?

I think wes goes into the basement and presses a button.

we're getting too chatty on the issue tracker, need an irc channel. or campfire.

[jseabold]

FWIW, I lurk and learn, but I'm sure others hate the noise just as much as I find it helpful. A dev list might be preferable to IRC unless you can commit resources to a bot to archive everything in a publicly searchable way. I don't know how public campfire is.

We're definitely pushing the boundaries of S/N, that's not cool.
Email is inconvenient for realtime, so not really a solution for many things.
I believe freenode has web based logs for all channels, though I'm not sure I consider that a plus.

[jreback]

fyi...i think i found this regression (my issue!)...issue a fix in a bit

[wesm]

I could set up a hipchat for pandas devs if you like, but that wouldn't be public.

[jreback]

#3093 fixes indexing_boolean_data_frame_rows