I am assigning @dbeltran as this change might conflict with some of the new features he has implemented.
The JobPackager build procedure was not considering the number RUNNING jobs to establish the limit of total jobs allowed to run at the same time. This decision might have been taken (or ignored) because, usually, the queue time is greater than the running time, so it was enough to consider SUBMITTED and QUEUING jobs. In the case of #672 (closed), the queue time was almost none, and the running times were long, this combination caused the overflow of running jobs.
FYI @mcastril