It's possible for concurrent, sharded mapReduces to fail with DEAD plan executors when there's a collision in temporary namespaces across multiple mongos processes.
This bug is intermittently triggered by the concurrency suite.
This seems to be the sequence of events:
1 - A mongos process issues a drop command, on all shards, on a tmp.mrs namespace after finishing the mapReduce.shardedfinish command (in cluster_map_reduce_cmd.cpp).
2 - At the same time, another mongos process tries to initialize a ParallelSortClusteredCursor on the very same tmp.mrs namespace as part of another mapReduce.shardedfinish command.
3 - The drop invalidates cursors on the tmp.mrs namespace, which leads to a DEAD plan executor and a failed mapReduce command.
Relevant log lines: