PERF: define is_all_dates to shortcut inadvertent copy when slicing an IntervalIndex

[qwhelan]

We get a few orders of magnitude speedup in IntervalIndex slicing by simply overriding the base class definition of is_all_dates, like all other Index derivatives also do. The root cause of the performance degradation is as follows:

• When slicing a Series, a new Series is created for the result.
• The last step in Series.__init__() is Series._set_axis(), which in turn calls .is_all_dates on the new Index
• The base definition of Index.is_all_dates is:

@cache_readonly
def is_all_dates(self):
    if self._data is None:
        return False
    return is_datetime_array(ensure_object(self.values))

which seems harmless at first glance. However, this eventually invokes IntervalArray.__array__, which is a pure Python for-loop creating Interval objects and leading to the performance regression here.

As the value of IntervalIndex.is_all_dates appears to always be False, even in the case of datetime-like left/right values, we simply override to return that value and shortcut the inadvertent copy described above.

Benchmarks

       before           after         ratio
     [ce62a5c1]       [a8f5e90b]
     <interval_index_fix~1>       <interval_index_fix>
        3.42±0.6s          446±0μs    ~0.00  indexing.IntervalIndexing.time_getitem_list
         101±50μs          103±0μs     1.02  indexing.IntervalIndexing.time_getitem_scalar
-      3.80±0.08s          355±0μs     0.00  indexing.IntervalIndexing.time_loc_list
          140±0μs          123±0μs    ~0.88  indexing.IntervalIndexing.time_loc_scalar

Speed up of ~10704x for time_loc_list

• closes PERF: Slowdown in IntervalIndex.get_loc #23576
• tests added / passed
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry (I don't think this regression has been in a release)

[pep8speaks]

Hello @qwhelan! Thanks for submitting the PR.

• There are no PEP8 issues in the file pandas/core/indexes/interval.py !

• There are no PEP8 issues in the file pandas/tests/indexes/interval/test_interval.py !

[codecov]

Codecov Report

Merging #23591 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master   #23591      +/-   ##
==========================================
+ Coverage   92.25%   92.25%   +<.01%     
==========================================
  Files         161      161              
  Lines       51237    51239       +2     
==========================================
+ Hits        47269    47271       +2     
  Misses       3968     3968

Flag	Coverage Δ
#multiple	90.63% <100%> (ø)	⬆️
#single	42.29% <50%> (ø)	⬆️

Impacted Files	Coverage Δ
pandas/core/indexes/interval.py	94.68% <100%> (+0.01%)	⬆️

---

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 ce62a5c...a8f5e90. Read the comment docs.

[TomAugspurger]

Thanks! I opened #23598 to investigate removing this implicit casting, but this will be nice if we don't do #23598 for 0.24.0

[jreback]

note that we should actually change the default for Index suclasses i think

[qwhelan]

Thanks @TomAugspurger !