[CI] TimeZoneRoundingTests#testRoundingRandom fails

[davidkyle]

https://elasticsearch-ci.elastic.co/job/elastic+elasticsearch+5.6+periodic/1149/console

   > Throwable #1: java.lang.AssertionError
   >    at __randomizedtesting.SeedInfo.seed([1E275DAB660CE5C0:B169E5B952C9109D]:0)
   >    at org.elasticsearch.common.rounding.TimeZoneRoundingTests.assertInterval(TimeZoneRoundingTests.java:589)
   >    at org.elasticsearch.common.rounding.TimeZoneRoundingTests.testRoundingRandom(TimeZoneRoundingTests.java:196)

This is a reproducible failure:

gradle :core:test \
  -Dtests.seed=1E275DAB660CE5C0 \
  -Dtests.class=org.elasticsearch.common.rounding.TimeZoneRoundingTests \
  -Dtests.method="testRoundingRandom" \
  -Dtests.security.manager=true \
  -Dtests.locale=et-EE \
  -Dtests.timezone=America/Grand_Turk

[DaveCTurner]

This reproduces in master (7b7c8c76a76ecb12e60f239ad7a50fe71ea7c149) too.

The offending time zone is America/St_Johns and the failing test is testing rounding in days - specifically rounding the chosen time point 1162091532554ms = 2006-10-29T03:12:12.554Z up and down to the nearest day and checking that the chosen time is between the rounded times.

It isn't, however. It rounds down to 1162002600000ms = 2006-10-28T02:30:00Z and rounds up to 1162089000000ms = 2006-10-29T02:30:00Z, both of which are earlier than the chosen time point.

Of importance, surely, is that America/St_Johns changed their clocks at 2006-10-29T00:01 (local time) from UTC+02:30 to UTC+03:30. The rounded time points are midnight local time as expected.

[DaveCTurner]

Note to future self: because this time zone sets the clocks back at 00:01, it has the first minute of Sunday 2006-10-29 followed by 59 bonus minutes of Saturday 2006-10-28, and the chosen time is within that bonus period. The bug seems to be that we expect every time on a day to precede the following day's midnight within a given timezone, which is one of many reasonable-yet-false assumptions you can make about how time works.

[DaveCTurner]

We concluded in today's Search & Aggs meeting that the current behaviour is desirable and the assertion in the test is at fault. The fix is therefore to change the assertion.

As of TZDB 2017c, the only regular occurrences of these overlapping days since 1950 are the clocks going back in America/St_Johns (Newfoundland, 1987-2010), America/Goose_Bay (Labrador, 1987-2010) and America/Moncton (New Brunswick, 1993-2006) as described above. As a further bonus, in America/St_Johns in 1988 the clocks actually went back 2 hours at 00:01. Eastern Canada is a special place indeed.

There was also a one-off shift in Antarctica/Casey by three hours from UTC+11 to UTC+8 at 2010-03-05 02:00 local time which took it back to 2010-03-04 for an extra hour.

[DaveCTurner]

(also note that Canada/Newfoundland is an alias to America/St_Johns so suppress that too)

[DaveCTurner]

This will be fixed by #28151.