[C++] Timezone database configuration and access

[asfimport]

Note: currently timezone database is not available on windows so timezone aware operations will fail.

We're using tz.h library which needs an updated timezone database to correctly handle timezoned timestamps. See installation instructions.

We have the following options for getting a timezone database:

1. local (non-windows) OS timezone database - no work required.

2. arrow bundled folder - we could bundle the database at build time for windows. Database would slowly go stale.

3. download it from IANA Time Zone Database at runtime - tz.h gets the database at runtime, but curl (and 7-zip on windows) are required.

4. local user-provided folder - user could provide a location at buildtime. Nice to have.

5. allow runtime configuration - at runtime say: "the tzdata can be found at this location"

   For more context see: ARROW-12980 and PEP 615.

Reporter: Will Jones / @wjones127
Assignee: Will Jones / @wjones127
Watchers: Rok Mihevc / @rok

Related issues:

• [R] fix behaviour when converting timestamps with "" as tzone #30005 (blocks)
• [Python] Use tzdata timezone database on Windows #31472 (relates to)
• [C++] Support cast from timestamp[UTC] to string on Windows #30186 (is related to)
• [R] clean-up date time unit testing once tzdb is available on Windows #31487 (is related to)

PRs and other links:

• ARROW-13168: [C++][R] Enable runtime timezone database for Windows #12536

_{Note: This issue was originally created as ARROW-13168. Please see the migration documentation for further details.}

[asfimport]

Dewey Dunnington / @paleolimbot:
Just a +1 for a "runtime configuration" option. In R we have the tzdb package. Currently it only provides the text format of the IANA database but we could use that approach if we need something different (maintained sepaerately to keep it up to date). I'm less familiar with Python but I imagine something similar exist there, too.

list.files(tzdb::tzdb_path("text"))
#>  [1] "africa"            "antarctica"        "asia"             
#>  [4] "australasia"       "backward"          "backzone"         
#>  [7] "calendars"         "checklinks.awk"    "checktab.awk"     
#> [10] "CONTRIBUTING"      "etcetera"          "europe"           
#> [13] "factory"           "iso3166.tab"       "leap-seconds.list"
#> [16] "leapseconds"       "leapseconds.awk"   "LICENSE"          
#> [19] "Makefile"          "NEWS"              "northamerica"     
#> [22] "README"            "southamerica"      "theory.html"      
#> [25] "version"           "windowsZones.xml"  "ziguard.awk"      
#> [28] "zishrink.awk"      "zone.tab"          "zone1970.tab"     
#> [31] "zoneinfo2tdf.pl"

[asfimport]

Rok Mihevc / @rok:
Nice to see interest on this @paleolimbot! I will probably pick it up next.

I'm not sure we want to depend on R/Python libraries being present from C++, but I'll look into this option, thanks!

[asfimport]

Will Jones / @wjones127:
It looks like that tzdb is a wrapper around this C++ library: https://github.com/HowardHinnant/date

I also saw that same library recommended in this StackOverflow answer by a Microsoft employee about Windows TZ libraries.

[asfimport]

Rok Mihevc / @rok:
Indeed it appears to be @wjones127. It also seems to be doing the bundling approach - including the timezone data with the install - https://github.com/r-lib/tzdb/tree/main/inst/tzdata

We also use date.h and tz.h (https://github.com/HowardHinnant/date#summary) in Arrow for almost all time related things.

[asfimport]

Dewey Dunnington / @paleolimbot:
For sure you don't want to depend on it being present from C+! From our end on the R bindings, though, it means we can support Windows through the tzdb R package (if C+ lets us point at a directory with a database in the right format).

[asfimport]

Rok Mihevc / @rok:
Ah, I see your point. Neat! Is tzdb package always present?

[asfimport]

Dewey Dunnington / @paleolimbot:
We can force it to be present at build-time, force it to be present at install time or force it to be present at runtime (probably what we'd do).

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
The same is the case for python: https://pypi.org/project/tzdata/

Ideally we would be able to specify the path at runtime, I think.

[asfimport]

Alessandro Molina / @amol-:
@pitrou I think that given that Python, R and C++ would have to be aligned in regard of how they can load(or embed) timezone database to work on Windows, it would be helpful to have your opinion. Personally I think that a reasonable solution would be to pass to C++ the "path" to the directory containing the database. So that both R and Py and pass their own dir and C++ can embedd files and provide them at install time. What do you think?

[asfimport]

Antoine Pitrou / @pitrou:
I think a runtime setting is the best solution here - if easily implemented. This is a pattern already in place for TLS certificates, and could be moved to a more general global-config routine:

https://github.com/apache/arrow/blob/master/cpp/src/arrow/filesystem/filesystem.h#L518-L536

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
It might need some tweaking of the vendored date.h to allow this (I seem to remember from when we started adding timezone related kernels that it doesn't support a runtime configuration out of the box)

[asfimport]

Will Jones / @wjones127:
I started to look at this then got busy with something else. My thought was we might be able to achieve setting this at runtime using set_install and reload_tzdb() (See these lines). Not 100% sure this will work; there seems to be some issue that this supports the text format and not some binary format (I have some more investigating to do).

[asfimport]

Antoine Pitrou / @pitrou:
date.h's set_install() seems to support the text form of the IANA database, which is also what R provides.

However, on Python, pytz provides the binary form of the IANA database, which date.h currently doesn't support on Windows.

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
I suppose we mainly care about tzdata and not pytz for Python (going forward, since that is what the standard library uses), however that seems to also use the binary form (https://pypi.org/project/tzdata/)

[asfimport]

Antoine Pitrou / @pitrou:
Yeah, let's start with R-only if that's easier.