SIGINT during test run leads to uneven test distribution on future runs

[Andrew Laski]

If a testr run is interrupted it records a test time of 0 for tests that didn't run. This causes issues when tests are being grouped for parallel runs since all of the 0 run time tests end up grouped into the same bucket which can grow disproportionately large. An example split I've seen is 14 14 10600 13 for concurrency 4.

[Robert Collins (lifeless)]

Oh nice catch, we should be not doing that.

[Corey Wright (coreywright)]

note: .testrepository/times.dbm needs to be deleted if created by an incomplete test run with < 0.0.19.

what i comprehend from the code change (commit 87e0de0) and validated through testing...

the change directly affects the writing of the database and indirectly the running of the tests, so if a database was written with < 0.0.19 during an interrupted test run, then the problem will still exhibit itself even after an upgrade as the "invalid" data will still be read from the database by 0.0.19. (not to be misinterpreted as me suggesting testrepository should start versioning and upgrading its database.)

[Corey Wright (coreywright)]

@Robert Collins

thanks for the fix!