Clients send batched statahead RPCs (MDS_BATCH) to OUT_PORTAL, so they're handled by the MDT's OUT threads (mdt_out, and ost_out on combined MDS/OSS nodes). A batched getattr can wait on another MDT. If that MDT is restarting, the OUT threads fill up waiting for it, and the restarting MDT's recovery reads, which also need those threads, never get served. Recovery hangs.