CI/TST: slow tests

[jreback]

xref #30641

from the new arm64 job, but indicative of general slowness: https://travis-ci.org/github/pandas-dev/pandas/jobs/685955999

• let's make these much faster
• we can still mark some as slow if needed
• let's change the https://github.com/pandas-dev/pandas/blob/master/ci/run_tests.sh#L23, durations to 20 or 30

This issue can be closed by multiple PRs

========================== slowest 10 test durations ===========================
88.88s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[tzlocal()-30S-Y]
65.15s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[tzlocal()-30S-10M]
22.33s call     pandas/tests/io/parser/test_multi_thread.py::test_multi_thread_path_multipart_read_csv[python]
21.54s call     pandas/tests/io/parser/test_multi_thread.py::test_multi_thread_path_multipart_read_csv[c_high]
21.03s call     pandas/tests/io/parser/test_multi_thread.py::test_multi_thread_path_multipart_read_csv[c_low]
19.46s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['dateutil/US/Pacific'-30S-Y]
15.75s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['dateutil/Asia/Singapore'-30S-10M]
15.71s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['dateutil/Asia/Singapore'-30S-Y]
15.61s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['Asia/Tokyo'-30S-Y]
15.59s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['US/Eastern'-30S-Y]
14.27s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['Asia/Tokyo'-30S-10M]
13.18s call     pandas/tests/groupby/transform/test_transform.py::test_cython_transform_frame[shift-args3-<lambda>]
13.11s call     pandas/tests/test_sorting.py::TestSorting::test_int64_overflow_moar
12.98s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['US/Eastern'-30S-10M]
12.92s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['dateutil/US/Pacific'-30S-10M]
12.78s call     pandas/tests/io/sas/test_sas7bdat.py::TestSAS7BDAT::test_iterator_loop
12.76s call     pandas/tests/io/sas/test_xport.py::TestXport::test1_basic
12.18s call     pandas/tests/groupby/transform/test_transform.py::test_cython_transform_frame[shift-args2-<lambda>]
11.97s call     pandas/tests/indexing/multiindex/test_chaining_and_caching.py::test_indexer_caching
11.95s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[pytz.FixedOffset(300)-30S-Y]
11.28s call     pandas/tests/io/parser/test_multi_thread.py::test_multi_thread_string_io_read_csv[python]
11.19s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[datetime.timezone(datetime.timedelta(0, 3600))-30S-Y]
10.60s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[tzutc()-30S-Y]
10.45s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[<UTC>-30S-Y]
10.35s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[datetime.timezone.utc-30S-Y]
10.19s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[datetime.timezone(datetime.timedelta(-1, 82800), 'foo')-30S-Y]
10.12s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[tzutc()-30S-10M]
10.05s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit['UTC'-30S-Y]
9.99s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[pytz.FixedOffset(-300)-30S-Y]
9.58s call     pandas/tests/resample/test_datetime_index.py::test_nearest_upsample_with_limit[<UTC>-30S-10M]```

[jreback]

updated the above list for top 30: https://travis-ci.org/github/pandas-dev/pandas/jobs/686095405 is the build (one of the standard ones)

[dsaxton]

For test_nearest_upsample_with_limits from pandas/tests/resample/test_datetime_index.py is it necessary to parametrize over so many timezones? Removing that fixture causes a speedup for me from around 5 minutes to less than a second. Looks there was a timezone specific issue, but I suppose a smaller set could be tested? #33895

[jreback]

The issue is here: https://github.com/pandas-dev/pandas/pull/33939/files#diff-dd427a6bbdfef7e333ac320adb15614dR539

its not the tz's at all, rather upsampling creates an enormous array

e.g.

In [10]: ts = pd.Series(1, index=pd.date_range("1/1/2000", periods=3, freq='Y'))                                                                              

In [11]: ts.resample('1min').nearest(limit=2)                                                                                                                 
Out[11]: 
2000-12-31 00:00:00    1.0
2000-12-31 00:01:00    1.0
2000-12-31 00:02:00    1.0
2000-12-31 00:03:00    NaN
2000-12-31 00:04:00    NaN
2000-12-31 00:05:00    NaN
                      ... 
2002-12-30 23:55:00    NaN
2002-12-30 23:56:00    NaN
2002-12-30 23:57:00    NaN
2002-12-30 23:58:00    1.0
2002-12-30 23:59:00    1.0
2002-12-31 00:00:00    1.0
Freq: T, Length: 1051201, dtype: float64

In [12]: ts.resample('10s').nearest(limit=2)                                                                                                                  
Out[12]: 
2000-12-31 00:00:00    1.0
2000-12-31 00:00:10    1.0
2000-12-31 00:00:20    1.0
2000-12-31 00:00:30    NaN
2000-12-31 00:00:40    NaN
2000-12-31 00:00:50    NaN
                      ... 
2002-12-30 23:59:10    NaN
2002-12-30 23:59:20    NaN
2002-12-30 23:59:30    NaN
2002-12-30 23:59:40    1.0
2002-12-30 23:59:50    1.0
2002-12-31 00:00:00    1.0
Freq: 10S, Length: 6307201, dtype: float64

so should split this test up into 2, e.g. no real need to upsample Y to 10s

[avirlrma]

take

[OlivierLuG]

I've made a PR to remove the fixtures @pytest.mark.slow when the tests were fast.

On the other hand, I have gathered these data for tests that take a lot of time to finish. It may be the functions to look to ?

pandas/tests/filename(::class)::function (first lines only)	Nb. of call	Avg. Time	Total time
groupby/test_value_counts.py::test_series_groupby_value_counts	1568	0,29	460,19
window/moments/test_moments_consistency_rolling.py::test_rolling_apply_consistency	1807	0,16	297,66
groupby/test_nunique.py::test_series_groupby_nunique	49	5,23	256,3
window/moments/test_moments_consistency_rolling.py::test_rolling_consistency	1807	0,06	110,99
indexing/multiindex/test_indexing_slow.py::test_multiindex_get_loc	1	99,42	99,42
window/moments/test_moments_consistency_rolling.py::test_rolling_consistency_series	757	0,13	97,43
io/test_stata.py::TestStata::test_stata_119	1	91,81	91,81
window/moments/test_moments_consistency_rolling.py::test_rolling_consistency_cov	1814	0,047	86,29
frame/test_arithmetic.py::TestFrameFlexArithmetic::test_floordiv_axis0_numexpr_path	2	28,185	56,37
window/moments/test_moments_consistency_ewm.py::test_ewm_consistency	862	0,057	49,15
window/moments/test_moments_consistency_rolling.py::test_rolling_consistency_var	1767	0,024	43,96
test_sorting.py::TestMerge::test_int64_overflow_issues	1	42,4	42,4
window/moments/test_moments_consistency_rolling.py::test_rolling_consistency_std	1807	0,0197	35,73
plotting/test_boxplot_method.py::TestDataFrameGroupByPlots::test_grouped_box_return_type	3	11,4	34,43
plotting/test_frame.py::TestDataFramePlots::test_errorbar_plot	2	15,27	30,54
computation/test_eval.py::TestAlignment::test_medium_complex_frame_alignment	4	7,5	30
frame/methods/test_duplicated.py::test_duplicated_do_not_fail_on_wide_dataframes	1	28,54	28,54
plotting/test_frame.py::TestDataFramePlots::test_errorbar_timeseries	2	12,55	25,1
plotting/test_misc.py::TestDataFramePlots::test_andrews_curves	3	7,08666666666667	21,26
computation/test_eval.py::TestAlignment::test_complex_series_frame_alignment	5	4,064	20,32
plotting/test_hist_method.py::TestDataFramePlots::test_hist_df_legacy	2	9,46	18,92
plotting/test_frame.py::TestDataFramePlots::test_plot	2	9,25	18,5
window/moments/test_moments_consistency_ewm.py::test_ewm_consistency_series_data	361	0,0477285318559555	17,2299999999999
computation/test_eval.py::TestEvalNumexprPandas::test_complex_cmp_ops	24	0,69625	16,71
plotting/test_boxplot_method.py::TestDataFramePlots::test_boxplot_legacy1	2	8,255	16,51
plotting/test_boxplot_method.py::TestDataFrameGroupByPlots::test_grouped_box_layout	2	7,96	15,92
window/moments/test_moments_consistency_expanding.py::test_expanding_consistency	216	0,0641203703703706	13,8500000000001
plotting/test_datetimelike.py::TestTSPlot::test_tsplot	2	6,72	13,44
plotting/test_datetimelike.py::TestTSPlot::test_ts_plot_with_tz	26	0,488076923076923	12,69
plotting/test_hist_method.py::TestSeriesPlots::test_hist_layout_with_by	2	6,34	12,68
plotting/test_hist_method.py::TestDataFrameGroupByPlots::test_grouped_hist_layout	3	4,10666666666667	12,32
plotting/test_series.py::TestSeriesPlots::test_hist_layout_with_by	2	6,16	12,32

[jreback]

most of the above list are now fixed by #34413; the csv multithread test could use some work on reducing the time

[TomAugspurger]

Cleared the milestone. When will we be able to close this?

[mroeschke]

Since it appears these timings were run a while ago, it's probably better that we open targeted issues for slow tests so it's more actionable. Closing.