RedisClusterClient default connection congestion

[pcchan]

Following up from my question in the Google Group.

As mentioned in the documentation, connections created with RedisClusterClient will always create a default connection to the node with the lowest latency. However, this may cause congestion when one node has a lower latency. If all clients experience the lowest latency with a particular node, then all clients will pick that node for its default connection.

In my case I'd be happy to use randomly selected nodes. That gives me a more predictable distribution of connections. Perhaps some environments prefer selecting by latency, but maybe lettuce could expose the choice as a client configuration?

[mp911de]

Thanks for the ticket. The default connection is created to the first node able to connect. RedisClusterClient starts to its connection attempts with the first entry of the Partitions collection (ordered by latency). Randomly selected nodes will work, but it's not a deterministic approach.

There are two other options which both seem reasonable to me:

1. Sticking to the order of seed nodes: The connection attempts are performed in the order the seed nodes are provided. This approach gives full control to the user.
2. Using the client count: Lettuce obtains the client connection count for each cluster node while topology discovery (can be disabled) and Partitions is ordered by the client count. Obtaining the number of client connections beyond the seed nodes can be disabled with the upcoming 4.2 version which is useful when running large clusters.

[mp911de]

The default will be to connect to the node with the least clients. For now RedisClusterClient will expose getSocketAddressSupplier(Function<Partitions, Collection<RedisClusterNode>> sortFunction) which can be overriden to specify an own function. lettuce ships with TopologyComparators that expose predefined sorting methods (sortByUri, sortByClientCount and predefinedSort). Applying a sorting function requires to subclass RedisClusterClient for now as this is the least invasive approach.

predefinedSort allows to use the initial seed nodes as predefined ordering and apply it to the dynamic topology results from Partitions. Nodes that are not found in the initial seed nodes are ordered by the URI (lexicographical comparison of the host and port string, i.e. 127.0.0.1:6379 < 127.0.0.1:6380). The block of nodes not found in the predefined order comes after the block of nodes with a predefined order.

[mp911de]

Implemented.