This happens when incoming documents have inconsistent type that causes replica shard allocation failure. While the cluster keeps trying to allocate the failed shards, tribe node memory usage grows unbounded. Eventually GC can't reclaim any space and tribe node hangs.