When running multiple emulator instances on a single node each within their own Xvnc session, at the end of the job the plugin seems to end up killing the adb daemon. The other emulators lose connection to the daemon, and even though a background task may start adb again, emulators may not successfully reconnect to adb. This is particularly problematic when running large multi-configuration android-emulator jobs on a single node with a large degree of parallelism. 2 concurrent executors seems to occasionally cause problems but with 4 the results are not even worthwhile to review.
Reducing the number of executors on the node to 1 (or otherwise weighting the job to consume an entire system) does work-around the problem.