The problem is that we submit a reindex task per cluster node to the ClusteredTaskQueue. But tasks in that queue can be run by any node and they will only run on one node. So whichever node first picks up the task will run it, which is most likely not the node that submitted the task. If the task is picked up by a node which currently is reindexing it will just be ignored, so in most cases you end up with at least one node not being reindexed.
This is an architectural problem and we need to use some kind of local task queue for reindexing tasks.