[FEA] TZiF reader to support timezone-aware operations

[shwina]

Related: #10047, #2477.

It is desirable to be able to do timezone-aware operations in cuDF.

One relatively simple approach is to load the IANA time zone database into cuDF as a table, and use existing algorithms to implement timezone-aware operations. This works quite well (see "Additional Context" below).

Unfortunately, the timezone database is typically not distributed or available in a format that libcudf can consume. It is usually distributed and present on user systems as a collection of TZiF files.

It would be great if we had a way to read TZiF files into cudf/libcudf. The first pass at this doesn't even need to be GPU-accelerated.

Additional context

To prototype this approach, I used a .csv version of the tzdb distributed by the third-party website timezonedb.com. It looks something like this:

             zone_name                    time_start  gmt_offset  dst
0     America/New_York 1883-11-18 16:59:59.000000000      -17762    0
1     America/New_York 1883-11-18 17:00:00.000000000      -18000    0
2     America/New_York 1918-03-31 07:00:00.000000000      -14400    1
3     America/New_York 1918-10-27 06:00:00.000000000      -18000    0
4     America/New_York 1919-03-30 07:00:00.000000000      -14400    1
...                ...                           ...         ...  ...
1156  America/New_York 1913-04-16 06:25:26.290448384      -18000    0
1157  America/New_York 1913-08-20 07:25:26.290448384      -14400    1
1158  America/New_York 1914-04-15 06:25:26.290448384      -18000    0
1159  America/New_York 1914-08-19 07:25:26.290448384      -14400    1
1160  America/New_York 1915-04-14 06:25:26.290448384      -18000    0

With the tzdb loaded into a table:

• tz_convert, or converting to/from UTC and between timezones, is a binary search (cudf::upper/lower_bound)
• tz_localize, or identifying nonexistent/ambiguous timestamps, is a binning problem (cudf::label_bins)

You can see my implementations of these operations here.

Perf comparison with Pandas:

# skipping some details in the code below:

psr = pd.Series(pd.date_range("1970-01-01", "2022-01-01", freq="1T"))
sr = cudf.from_pandas(psr)

%timeit psr.dt.tz_localize("America/New_York", ambiguous="NaT", nonexistent="NaT")
2.11 s ± 779 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

%timeit sr.dt.tz_localize("America/New_York", ambiguous="NaT", nonexistent="NaT")
31.4 ms ± 509 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

%timeit psr.dt.hour
2.12 s ± 8.72 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

%timeit sr.dt.hour
1.58 ms ± 9.46 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

[shwina]

After a conversation with @bdice:

As it turns out, we already have a TZiF reader in https://github.com/rapidsai/cudf/blob/branch-22.10/cpp/src/io/orc/timezone.cpp. It needs a bit of refactoring to return a cudf::table, but nothing major. I'll try to use this with the tz-experiment branch I've been prototyping with and report back here.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[shwina]

Closing in favor of #12813