BUG/PERF: perf issues in object groupby aggregations (GH7555)

[jreback]

related #7555

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
groupby_object_last                          |  15.2340 | 4175.6846 |   0.0036 |
groupby_object_first                         |  15.6957 | 4166.6920 |   0.0038 |
groupby_transform_ufunc                      |   5.4990 |   5.7740 |   0.9524 |
groupby_indices                              |   6.7443 |   7.0153 |   0.9614 |
groupby_multi_different_functions            |  12.6866 |  12.9040 |   0.9832 |
groupby_multi_series_op                      |  13.2560 |  13.4466 |   0.9858 |
groupby_multi_cython                         |  15.0924 |  15.2720 |   0.9882 |
groupby_multi_different_numpy_functions      |  11.1883 |  11.2376 |   0.9956 |
groupby_first_float32                        |   3.4907 |   3.4963 |   0.9984 |
groupby_last_float32                         |   3.6677 |   3.6586 |   1.0025 |
groupby_frame_apply_overhead                 |   9.2754 |   9.2440 |   1.0034 |
groupby_first                                |   3.5706 |   3.5577 |   1.0036 |
groupby_pivot_table                          |  17.2007 |  17.1343 |   1.0039 |
groupby_transform                            | 174.6823 | 173.8520 |   1.0048 |
groupby_frame_singlekey_integer              |   2.4533 |   2.4397 |   1.0056 |
groupby_frame_nth                            |   2.7863 |   2.7630 |   1.0085 |
groupby_simple_compress_timing               |  35.6836 |  35.3360 |   1.0098 |
groupby_sum_booleans                         |   1.2880 |   1.2727 |   1.0121 |
groupby_transform2                           | 162.5650 | 160.2323 |   1.0146 |
groupby_apply_dict_return                    |  39.7460 |  39.1597 |   1.0150 |
groupby_multi_size                           |  23.2656 |  22.9087 |   1.0156 |
groupby_datetimes_last                       |  12.0660 |  11.8797 |   1.0157 |
groupby_series_simple_cython                 | 192.2614 | 188.9360 |   1.0176 |
groupby_last                                 |   3.8017 |   3.7254 |   1.0205 |
groupby_int_count                            |   4.6777 |   4.5680 |   1.0240 |
groupby_mixed_first                          |  12.3363 |  12.0014 |   1.0279 |
groupby_frame_median                         |   7.6403 |   7.4034 |   1.0320 |
groupby_frame_apply                          |  45.0197 |  43.6123 |   1.0323 |
groupby_object_nth                           | 430.5693 | 416.2673 |   1.0344 |
groupby_datetimes_nth                        | 428.6557 | 414.3833 |   1.0344 |
groupby_frame_cython_many_columns            |   4.3843 |   4.2194 |   1.0391 |
groupby_multi_python                         | 142.0707 | 135.8123 |   1.0461 |
groupby_nth_float32                          |  63.5343 |  59.9463 |   1.0599 |
groupby_nth_float64                          |  64.0316 |  59.8179 |   1.0704 |
groupby_multi_count                          |   9.0540 |   8.2114 |   1.1026 |
-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------

Ratio < 1.0 means the target commit is faster then the baseline.
Seed used: 1234

Target [e28ec0d] : BUG/PERF: perf issues in object groupby aggregations (GH7555)
Base   [69bb0e8] : Merge pull request #7556 from onesandzeroes/expand-grid

DOC: Cookbook recipe for emulating R's expand.grid() (#7426)

[cpcloud]

hm sorry for the churn here, i would've thought that float64 output for object would be faster than object output

[cpcloud]

moral of the story: i should vbench more often

[cpcloud]

can you show the vbench?

[jreback]

no the problem with the float64 was that it was blowing up on object dtypes themselves (e.g. strings), forcing a slow python path for those.

[cpcloud]

ah ok

[cpcloud]

is it just me or are these kind of hard to test...

[jreback]

yeh...didn't have vbenches for them, that was a problem
further the exception messages are trapped....hmmm let me add that to #7569

I added for more dtypes