@xiezhq-hermann
The background of the problem we described:
We use HiRadixCache in the scenario of PD separation, write_back strategy. The local radix tree will send update events when nodes are added and deleted in rank 0, and the global radix tree will be adjusted according to the update events. When the request comes, we first match according to the global radix tree, and decide to choose P nodes and D nodes according to the number of prefix matches and load. We found that the number of matches in the global tree is sometimes much larger than the number of matches in the local number under the premise of distinguishing between instances. It looks like the host indices is not matched.
In the process of troubleshooting the problem, we encountered the following problems:

1、pending_nodes is not used

pending_nodes is not used, this will cause the parent node not to be placed in the heap. Maybe assigned a value here:

 if self.cache_controller.write_policy == "write_back": 
     num_evicted += self.write_backup(x) 
     ----> pending_nodes.append(x) <----

2、token_to_kv_pool_allocator not release device_indices if write_policy is write_back

token_to_kv_pool_allocator not release device_indices if write_policy is write_back, maybe it should be released in the writing_check function

    def writing_check(self):
        ...
        for _ in range(queue_size.item()):
            ack_id = self.cache_controller.ack_write_queue.get()
            self.dec_lock_ref(self.ongoing_write_through[ack_id])
            ----> self._evict_write_through(self.ongoing_write_through[ack_id]) <----
            del self.ongoing_write_through[ack_id]

3、inc_lock_ref function in writing_check causes all parent nodes to be locked.

This function of inc_lock_ref called by write_check wiil lock node parents so that parent cannot be added to the heap