Unnecessary churn in task allocation for non-grouped windowed tasks

[lbradstreet]

Download:
results.edn.zip

Then view via the console dashboard:
./onyx-dashboard -e ~/Dropbox/results.edn -j "onyx.job-scheduler/balanced" -s jepsen -o "0.8.5-SNAPSHOT" -f 3803dd05-87ea-4ac3-9df2-7ccdb70815ca

This will filter by a task using windowing/state management 3803dd05-87ea-4ac3-9df2-7ccdb70815ca. Jump to message id 126, and view in replica diff mode by pressing d.

Notice that a leave-cluster event has occured for peer: 91a50764-3f0e-4407-b568-636908a2985c. This peer is not allocated to task 3803dd05-87ea-4ac3-9df2-7ccdb70815ca, and yet peer 622286d3-1ea8-4032-9aad-821a3d676fbc is moved off the task, and peer 5cb4c0b5-a794-4d19-8d90-a4e8698ee529 is placed on it. Peer 5cb4c0b5-a794-4d19-8d90-a4e8698ee529 should be placed on task 145b859a-1a33-4a93-8e78-ccc483066396 where the windowed/state management task was moved to.

Windowed/state managed tasks should possibly be treated specially to reduce churn, however in this case the reallocation caused two peers to switch tasks when only one was necessary.

[lbradstreet]

@MichaelDrogalis this would be a good one to take if you have a bit of time and wanted to do some coding.

[lbradstreet]

This one might be due to previously reported bugs, since we're not on the latest btrplace?

[lbradstreet]

Hrm, I noticed an issue with onyx-console-dashboard. It's fixed to whatever version of onyx is in the dependencies for the main library. I may have to move code from the library into the script. Anyway, it should be good for now because I've fixed it to 0.8.5-SNAPSHOT in the library.

[lbradstreet]

Closed in #511