Reproduction case:

Create a concurrent matrix job with a user defined axis that does a 'sleep 120' as its build step. Launch several of these jobs, enough that all available executors are taken up and there are still builds in the queue. Some of these builds will abort, with a console message similar to:

[...]
9 completed with result SUCCESS
24 completed with result SUCCESS
23 completed with result SUCCESS
2 completed with result SUCCESS
10 appears to be cancelled
10 completed with result ABORTED
25 appears to be cancelled
25 completed with result ABORTED
18 appears to be cancelled
18 completed with result ABORTED
13 appears to be cancelled
[...]

For my test case, I have 26 slaves, 82 executors, 25 sub-jobs. I can reproduce reliably if I launch 5 or more top level jobs at once.