[tz] Adding a "test" region?
paul at ganssle.io
Fri Jan 26 16:53:13 UTC 2018
On 01/26/2018 12:45 AM, Brian Inglis wrote:
> The official release must then remain the main stable compatibility release for
> consumption by the major downstream distributors and users.
I *very strenuously* disagree with this. I think in general the default should be the latest version of the database, and any compatibility release should be something you have to opt in to. Generally people are not going to pick anything except the main release, and they are not going to make any changes unless that breaks. The whole point of a second release is that if some change (possibly necessary) breaks enough downstream consumers, it might be worth providing a release without these changes so that they don't have to manually patch all new releases themselves. People who are not broken will not seek out the alternate source because nothing breaks for them - and people writing new software will likely start to rely on the same undocumented features that the old, broken software uses.
That said, I am also not entirely convinced that a second "compatibility" release is warranted. Even the Ireland case did not break backwards compatibility in any meaningful way - as has been said many times, zic has had the capability to handle negative DST offsets for decades, this is just the first example of the actual *use* of this feature. I think the main reason this is controversial is that this is, in some sense, an "opt in" change, because we can choose to define the "standard" offset as the one that allows all other offsets in the zone to be defined with positive SAVE values. In most other cases where major software has made inappropriate assumptions about the nature of the data, we will not get to "opt out", and if a "compatibility" release is provided, it will necessarily contain unambiguously inaccurate data. Rather than maintain a release that is the lowest common denominator of all the inaccuracies that are required to allow all the broken software out there to work, I think it should probably be incumbent upon people deploying the broken software to add in these inaccuracies to the data *a la carte* to the data before consuming it, so that they can have the *most* accurate data that their software can handle.
> On 2018-01-24 15:24, Paul G wrote:
>> On 01/24/2018 04:55 PM, Paul.Koning at dell.com wrote:
>>> No, please do NOT do that. Anyone who depends on an automated or semi-automated system to track real changes will get completely messed up by such a tactic. Let the real database contain real data. A test database is fine; a test zone in the real data is not.
>> Do you have examples of this? I'm not really sure what problem it would cause to have well-namespaced fictional zones in the real database. What kind of automated processes would be affected by there being some zones that don't represent real zones? I could imagine there being some applications that don't know about these test zones that erroneously display them to users, but why would users set their time zone to `TestZone/Test_Monthly_DST` or something? Doesn't seem hugely risky to me, but I would not be surprised to find that there are some applications where this could be a deal-breaker.
> See Hawai'i Missile Alert technical articles or search Risks list for "test in
> production" or "test data in production database" articles.
> How much do you want to further delay downstream distribution of changes?
> As someone else mentioned about Debian, the current release on most major
> distributions is 2017c, often on ...-update or even ...-security repos to get
> changes distributed!
> Package maintainers and QA have enough work to do to get stable releases thru
> their processes without introducing deliberate errors that get a release put on
> the back burner until the *volunteers* have time to deal with broken packages,
> or release a package that breaks downstream applications before it gets
> detected, complained about, and reverted.
> Distros are already pretty selective about what they package from a release, as
> looking at the selection of files available in a few distros' tzdata binary and
> source packages will show.
> The non-standard (non-existent) structure of the release probably requires
> package scripts to at least move each file into a directory structure acceptable
> to the distro build and packaging tools, add build and packaging control files,
> and/or patch distributed files to conform to their tools and processes, then get
> the builds and packages to a state where they can pass their QA processes to
> become officially released binary and source packages.
> As a result, additions to a release are likely to be ignored until that breaks a
> process or fails QA e.g. moving version from Makefile to a separate file.
> QA processes are likely to involve regression testing all downstream packages
> dependent on tzdata, so any change that causes a downstream package issue can
> halt tzdata packaging until the issue can be reported, investigated, diagnosed,
> fixed, tested, and either resolved or the test bypassed for minor issues.
> Unfortunately, not all (no?) libraries use the tzcode reference implementation,
> as they also have to support i18n and l14n facilities, so QA will presumably run
> libc, coreutils, and other package tests across multiple locales, increasing the
> possibility of regressions.
>> That said, one of the problems this would be designed to solve is that apparently a lot of people are already *not* testing against the master branch or doing their own tests, since we hear about a lot of problems only *after* new releases come out. Having the default be "you don't get the test zones" would at least partially defeat the purpose of such an initiative.
> As some major distributors have previously mentioned, their orgs have policies
> that they can only use official releases from official sources, and do not have
> either the authorization or resources to test against the experimental github
> The result is that most significant complaints are made by major distributors
> only after official releases.
> Some users perform their own QA against the experimental github repo and they
> contribute reports and patches that will avoid some problems when an official
> release is generated.
>>> Of course it would be possible for the tools that track changes to ignore the test entry, but that's an incompatible change. And that also only works if the test zone(s) have rigorously defined name patterns fixed for all time.
>> I assume it's fine to have rigorously defined name patterns. I was thinking of something designed from the beginning to be filtered out (it could, for example, *only* generate zones in TestZones/, and/or there can be a "testzones" file that lists all the fictional zones).
>> Since it seems like Paul is already leaning towards moving to a 2-track
> release ("main" release and "compatibility" release that is more stable with
> respect to some of these features), it would make sense to distribute the test
> zones *only* as part of the "main" release (and whether the `TestZone/`
> region/folder is generated at all can be opt-in in `zic`, since this is mostly
> intended for downstream *compilers*, not for the much more stable compiled
> binaries). Presumably anyone not ready to handle the test zones can opt-out
> easily enough by using the compatibility releases.
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 833 bytes
Desc: OpenPGP digital signature
More information about the tz