PERF: Regressions since 0.20.3

[TomAugspurger]

These are fairly reproducible on my machine:

       before           after         ratio
     [eac4d3f7]       [92db5c9c]
+           396μs            746μs     1.88  inference.DtypeInfer.time_uint32
+     2.63±0.02ms       3.93±0.1ms     1.49  indexing.IndexingMethods.time_take_dtindex
+     2.64±0.06ms      3.83±0.06ms     1.45  indexing.IndexingMethods.time_take_intindex

       before           after         ratio
     [eac4d3f7]       [92db5c9c]
+     1.43±0.01ms      1.74±0.07ms     1.22  frame_ctor.FrameConstructorDTIndexFromOffsets.time_frame_ctor('BDay', 1)
+     1.56±0.02ms      1.74±0.07ms     1.12  frame_ctor.FrameConstructorDTIndexFromOffsets.time_frame_ctor('MonthBegin', 2)
+          56.0ms           61.8ms     1.10  frame_ctor.FrameConstructorDTIndexFromOffsets.time_frame_ctor('FY5253Quarter_2', 2)

[jreback]

cc @jbrockmendel if u have any ideas here

[jbrockmendel]

I'll take a look at this.

[jbrockmendel]

Best guess is the change in _to_i8 (previously in _libs.index, now in tslib). Old:

cdef inline _to_i8(object val):
    cdef pandas_datetimestruct dts
    try:
        return val.value
    except AttributeError:
        if util.is_datetime64_object(val):
            return get_datetime64_value(val)
        elif PyDateTime_Check(val):
            tzinfo = getattr(val, 'tzinfo', None)
            # Save the original date value so we can get the utcoffset from it.
            ival = _pydatetime_to_dts(val, &dts)
            if tzinfo is not None and not is_utc(tzinfo):
                offset = get_utcoffset(tzinfo, val)
                ival -= tslib._delta_to_nanoseconds(offset)
            return ival
        return val

New:

cdef inline _to_i8(object val):
    cdef pandas_datetimestruct dts
    try:
        return val.value
    except AttributeError:
        if is_datetime64_object(val):
            return get_datetime64_value(val)
        elif PyDateTime_Check(val):
            return Timestamp(val).value
        return val

[jbrockmendel]

Hmm in %timeit reverting this makes a pretty big difference for the ind.get_loc(0) bench, but pretty small (though at least in the right direction) differences for time_take_dtindex and time_take_intindex

[jbrockmendel]

Does asv have something analogous to git bisect?

[jorisvandenbossche]

Maybe https://asv.readthedocs.io/en/latest/commands.html#asv-find ?

[TomAugspurger]

Sorry, those commit hashes in the original post look entirely incorrect... Let me rerun them.

[TomAugspurger]

Updated:

       before           after         ratio
       [888c4ec3]       [36c309ed]
!     4.01±0.02ms           failed      n/a  frame_ctor.FrameConstructorDTIndexFromOffsets.time_frame_ctor('LastWeekOfMonth', 1)
!     4.39±0.01ms           failed      n/a  frame_ctor.FrameConstructorDTIndexFromOffsets.time_frame_ctor('LastWeekOfMonth', 2)
!      9.17±0.1ms           failed      n/a  frame_methods.Formatting.time_frame_repr_wide
!      19.7±0.2ms           failed      n/a  frame_methods.Formatting.time_html_repr_trunc_mi
!      14.4±0.2ms           failed      n/a  frame_methods.Formatting.time_html_repr_trunc_si
!      10.8±0.1ms           failed      n/a  frame_methods.Formatting.time_repr_tall
!         117±2ms           failed      n/a  frame_methods.Formatting.time_to_html_mixed
!      14.1±0.1ms           failed      n/a  frame_methods.Formatting.time_to_string_floats
+      10.3±0.5ms            4.83s   469.35  inference.DtypeInfer.time_timedelta64_2
+     8.94±0.09ms            3.03s   339.45  inference.DtypeInfer.time_datetime64
+      29.8±0.4ms            5.99s   201.19  binary_ops.Timeseries.time_timestamp_ops_diff2
+        80.7±1ms            9.27s   114.89  binary_ops.TimeseriesTZ.time_timestamp_ops_diff2
+     3.36±0.03ms          278±2ms    82.71  timeseries.SeriesArithmetic.time_add_offset_delta
+      60.2±0.4ms            3.91s    64.91  inference.DtypeInfer.time_timedelta64_1
+      9.64±0.2ms          297±5ms    30.83  timeseries.SeriesArithmetic.time_add_offset_fast
+     6.18±0.09ms          122±1ms    19.77  categoricals.Categoricals3.time_rank_string_cat
+      36.6±0.4ms          576±1ms    15.75  packers.packers_read_stata.time_packers_read_stata
+     45.8±0.01ms            616ms    13.44  packers.packers_read_stata_with_validation.time_packers_read_stata_with_validation
+      28.5±0.2ms          308±2ms    10.84  packers.STATA.time_write_stata
+      40.5±0.4ms        323±0.9ms     7.97  packers.STATA.time_write_stata_with_validation
+      34.1±0.4ms          107±1ms     3.15  join_merge.Align.time_series_align_int64_index
+        631±10μs      1.51±0.01ms     2.39  indexing.Int64Indexing.time_getitem_array
+     7.13±0.04ms      16.4±0.09ms     2.30  timeseries.DatetimeIndex.time_infer_freq_none
+     3.62±0.05ms       7.34±0.1ms     2.03  indexing.Int64Indexing.time_getitem_lists
+     2.44±0.04μs      4.25±0.05μs     1.74  indexing.DataFrameIndexing.time_get_value
+         285±6μs         490±10μs     1.72  indexing.Int64Indexing.time_getitem_list_like
+      19.2±0.1ms       30.8±0.8ms     1.60  sparse.sparse_array_constructor.time_sparse_array_constructor_object_non_nan_fill_value_1percent
+      32.2±0.4ms       48.7±0.1ms     1.51  sparse.sparse_array_constructor.time_sparse_array_constructor_object_non_nan_fill_value_10percent
+     2.47±0.03ms      3.72±0.04ms     1.51  indexing.IndexingMethods.time_take_dtindex
+     2.48±0.07ms      3.73±0.03ms     1.50  indexing.IndexingMethods.time_take_intindex
+      6.12±0.1ms      9.04±0.05ms     1.48  binary_ops.Ops2.time_frame_int_div_by_zero
+     5.33±0.01ms      7.65±0.01ms     1.43  timeseries.DatetimeIndex.time_infer_freq_business
+        564±20ms          792±7ms     1.40  timeseries.SeriesArithmetic.time_add_offset_slow
+     2.95±0.09ms      4.05±0.03ms     1.37  strings.StringMethods.time_get
+     2.68±0.02ms      3.66±0.06ms     1.37  period.Algorithms.time_value_counts_pseries
+     2.09±0.03ms      2.87±0.02ms     1.37  period.Algorithms.time_drop_duplicates_pseries
+     6.84±0.07μs       8.89±0.1μs     1.30  timestamp.TimestampOps.time_replace_across_dst
+         252±2ms          322±5ms     1.28  rolling.SeriesRolling.time_rolling_quantile_median_l
+      7.91±0.1μs       10.1±0.1μs     1.28  timestamp.TimestampOps.time_replace_tz
+       252±0.6ms          322±1ms     1.28  rolling.DataframeRolling.time_rolling_quantile_median_l
+     1.25±0.01ms      1.58±0.01ms     1.26  categoricals.Categoricals.time_constructor_datetimes_with_nat
+         1.30±0s            1.61s     1.24  groupby.GroupBySuite.time_unique('float', 10000)
+      10.1±0.2μs       12.5±0.2μs     1.23  timestamp.TimestampProperties.time_is_quarter_end
+         842±4ms          1.02±0s     1.22  groupby.GroupBySuite.time_unique('int', 10000)
+         113±1ms          135±1ms     1.19  rolling.SeriesRolling.time_rolling_quantile_median
+         111±1ms          132±2ms     1.19  rolling.DataframeRolling.time_rolling_quantile_median
+     3.17±0.04ms      3.77±0.05ms     1.19  index_object.SetOperations.time_datetime_difference
+      9.66±0.1ms       11.4±0.1ms     1.18  groupby.GroupBySuite.time_unique('int', 100)
+      35.5±0.5μs       41.5±0.4μs     1.17  indexing.IntervalIndexing.time_getitem_scalar
+      14.5±0.2ms       16.9±0.4ms     1.17  groupby.GroupBySuite.time_unique('float', 100)
+      59.9±0.3ms       69.8±0.3ms     1.17  hdfstore_bench.HDF5.time_write_store_table
+        34.7±1ms       40.3±0.5ms     1.16  strings.StringMethods.time_join_split_expand
+     3.98±0.05ms      4.59±0.05ms     1.15  groupby.Transform.time_transform_ufunc
+         227±2μs          261±6μs     1.15  indexing.MultiIndexing.time_series_xs_mi_ix
+      10.5±0.1μs       12.1±0.2μs     1.15  indexing.Int64Indexing.time_iloc_scalar
+     1.68±0.01ms         1.93±0ms     1.14  indexing.MultiIndexing.time_multiindex_get_indexer
+         251±4μs          287±2μs     1.14  indexing.MultiIndexing.time_multiindex_large_get_loc
+     6.53±0.03μs      7.47±0.07μs     1.14  timestamp.TimestampOps.time_replace_None
+      41.3±0.7μs       46.8±0.5μs     1.13  indexing.Int64Indexing.time_loc_scalar
+      84.1±0.7μs       94.4±0.6μs     1.12  series_methods.series_clip.time_series_dropna_datetime
+           508ms            568ms     1.12  groupby.groupby_multi_index.time_groupby_multi_index
+       206±0.8ms          231±2ms     1.12  timeseries.ToDatetime.time_iso8601_tz_spaceformat
+         199±2ms        221±0.5ms     1.12  packers.Packers.time_packers_read_csv
+         606±3μs          674±5μs     1.11  groupby.GroupBySuite.time_value_counts('float', 100)
+       265±0.9μs          294±3μs     1.11  indexing.MultiIndexing.time_frame_xs_mi_ix
+      8.49±0.1ms         9.39±1ms     1.11  eval.Eval.time_mult('numexpr', 1)
+     8.07±0.04μs      8.91±0.09μs     1.10  timeseries.DatetimeIndex.time_timestamp_tzinfo_cons
+         209±2ms          231±3ms     1.10  groupby.GroupBySuite.time_describe('float', 100)
+          73.8ms           81.4ms     1.10  packers.JSON.time_write_json
+      58.2±0.7μs       64.0±0.7μs     1.10  indexing.IntervalIndexing.time_loc_scalar
+      31.3±0.2μs       34.4±0.6μs     1.10  indexing.Int64Indexing.time_ix_scalar

verifying for correctness and then I'll start digging into them.

[TomAugspurger]

+      10.3±0.5ms            4.83s   469.35  inference.DtypeInfer.time_timedelta64_2
+     8.94±0.09ms            3.03s   339.45  inference.DtypeInfer.time_datetime64

are real. _maybe_box_datetimelike looks suspicious (this is the RC):

ncalls	tottime	percall	cumtime	percall	filename:lineno(function)
2000000	6.369	3.185e-06	6.57	3.285e-06	common.py:170(_maybe_box_datetimelike)
2	0.6152	0.3076	7.185	3.593	base.py:813()
2000723/2000456	0.2013	1.006e-07	0.203	1.015e-07	~:0()

http://pandas.pydata.org/speed/pandas/#inference.DtypeInfer.time_datetime64 points to ~ eef810ef (cc @jreback)

[TomAugspurger]

DtypeInfer:

+      10.3±0.4ms            4.78s   464.27  inference.DtypeInfer.time_timedelta64_2
+     8.71±0.07ms            2.93s   336.02  inference.DtypeInfer.time_datetime64
+        63.4±2ms            3.84s    60.54  inference.DtypeInfer.time_timedelta64_1

SeriesArthmetic.time_add_offset*

560±20ms       before           after         ratio
     [888c4ec3]       [36c309ed]
+     3.38±0.08ms          298±9ms    87.90  timeseries.SeriesArithmetic.time_add_offset_delta
+     10.00±0.3ms          294±3ms    29.37  timeseries.SeriesArithmetic.time_add_offset_fast
+        560±20ms          803±2ms     1.43  timeseries.SeriesArithmetic.time_add_offset_slow

These are handled by #17980

[TomAugspurger]

categoricals.Categoricals3.time_rank_string_cat_ordered

5.12±0.08ms       before           after         ratio
     [888c4ec3]       [36c309ed]
+      7.22±0.5ms          139±6ms    19.25  categoricals.Categoricals3.time_rank_string_cat

Seems like this is ranking the objects, instead of codes.

This is handled by #17982

[TomAugspurger]

packers.*stata.*

46.1±0.01ms       before           after         ratio
     [888c4ec3]       [36c309ed]
+      36.5±0.3ms          601±6ms    16.47  packers.packers_read_stata.time_packers_read_stata
+     46.1±0.01ms            575ms    12.48  packers.packers_read_stata_with_validation.time_packers_read_stata_with_validation
+      30.2±0.7ms            346ms    11.48  packers.STATA.time_write_stata
+          42.5ms            339ms     7.96  packers.STATA.time_write_stata_with_validation

Fixed by #17980