DOC: DateOffset.is_anchored track down intention, fix docstring

[jbrockmendel]

def is_anchored(self) -> bool:
        # TODO: Does this make sense for the general case?  It would help
        # if there were a canonical docstring for what is_anchored means.

The docstring on the base class is_anchored pretty clearly doesn't make sense in the non-base class methods, so I suspect it is wrong. It has never been clear to me exactly what the intention of is_anchored is, but my best guess is something like offsets that are "anchored" on a particular day of the week or month of the year or etc.

The task here is to dig into the GH history to determine if the intent and correct behavior was ever clear. If so, to fix the implementation and/or docstring (and then remove the comment). The method isn't used outside the tests and I doubt many users use it, so OK to deprecate if it isn't useful or doesn't make sense.

I suspect something like it can make sense and be useful for DatetimeIndex set operations to determine when we can use RangeIndex fastpaths, xref #44025.

For historical digging/grepping it may be useful to know that this used to be isAnchored.

[tpackard1]

I'll start looking into the GH history and based on what I have found, then make a PR if it is warranted or come back here with my findings if intent and correct is ambiguous.

[tpackard1]

take

[tpackard1]

Hi @jbrockmendel, I have been looking through the pandas historical releases code and have gone back as far as v0.18.0 where isAnchored was still in use and the code for `DateOffset.isAnchored is as follows:

    def isAnchored(self):
        return (self.n == 1)

The above provides even less detail information about what the use case would be for the class DateOffset. And @jbrockmendel your comment that, "The method isn't used outside the tests and I doubt many users use it, ...", seems to be ringing true as I cannot determine a use case for this method, so I think depreciation may be the route to take but I will keep going back in GH history if you think it is necessary? Or I will go ahead and submit a PR to depreciate it if that is the correct course.

[jbrockmendel]

Thanks for taking a look @tpackard1. I'm going to defer to @natmokval on this.

[natmokval]

@jbrockmendel, sorry for taking a while to get into this.

I dug into the GH history and I found the first appearance of isAnchored

commit c6b236db73ff81007909be6406f0e484edc4a9eb (HEAD)
Author: Wes McKinney <wesmckinn@gmail.com>
Date:   Wed Aug 5 03:30:16 2009 +0000

We define isAnchored in pandas/core/datetools.py

class Week(DateOffset):
    def isAnchored(self):
        return (self.n == 1 and self.dayOfWeek is not None)

class BQuarterEnd(DateOffset):
   def isAnchored(self):
        return (self.n == 1 and self.startingMonth is not None)

and for classes DateOffset / BDay / BMonthEnd the definition is the same:

def isAnchored(self):
    return (self.n == 1)

isAnchored is called only once in constructor of the class DateRange in pandas/core/daterange.py (link)
In case we have anchored offset we call classmethod getCachedRange

if offset.isAnchored() and not isinstance(offset, datetools.Tick):
    index = cls.getCachedRange(fromDate, toDate, periods=periods,
                                           offset=offset)

[natmokval]

The first time isAnchored appeared in 2009, see commit c6b236db73ff81007909be6406f0e484edc4a9eb
Definitions of isAnchored were quite similar to ones we use now. The idea behind: offsets with n=1 are used more often than offsets with n > 1, so we didn’t want to compute offsets with n=1 every time. To reuse them we used cache.
We called isAnchore only in constructor of the class DateRange in order to save this offset to cachedRange

Then in 2012, see commit 0cc1a3d0bc4617440e3df9e3e55b19cb6fb51a2d all definitions isAnchored were moved to offsets.py and we used isAnchored in _should_cache, which we called in constructor of class DatetimeIndex.

Then in 2019, see commit 1d36851ffdc8391b8b60a7628ce5a536180ef13b we deprecated isAnchored and started to use is_anchored instead (there was no connection between is_anchored and cache since then). We used is_anchored in def first and in def intersection

[natmokval]

Seems like now we don't use is_anchored as it was supposed to be used at the beginning.
I checked: is_anchored is called only in tests, several test_is_anchored, and one test_constructor_timestamp.

I think we can either deprecate is_anchored or correct documentation for is_anchored. I suggest to deprecate it. What do you think @jbrockmendel?

[jbrockmendel]

Agreed let's deprecate this.