The latest version cannot handle situations where the Redis master node goes down.

[conan8737]

When the Redis master fails, it triggers an automatic ​failover​. Requests sent to the old master node will time out. At this point, eredis_cluster should refresh the cluster mapping, but currently it treats this situation as a pool_busy error - only retrying the operation without actually refreshing the cluster mapping. Therefore, we can only wait for TCP timeout, which would take an excessively long time.

-spec transaction(PoolName::atom(), fun((Worker::pid()) -> redis_result())) ->
    redis_result().
transaction(PoolName, Transaction) ->
    try
        poolboy:transaction(PoolName, Transaction)
    catch
        exit:{timeout, _GenServerCall} ->
            %% Poolboy checkout timeout, but the pool is consistent.
            {error, pool_busy};
        exit:_ ->
            %% Pool doesn't exist? Refresh mapping solves this.
            {error, no_connection}
    end.

poolboy:transaction involves two steps:

1. Acquire a connection​ from the connection pool
2. Send the message​ to the Redis node using the connection

Currently, timeouts in ​both steps​ are uniformly treated as pool_busy errors. This should be differentiated:

When ​message sending times out, should refresh the cluster mapping instead of just retrying?

[zuiderkwast]

Thanks for the report. Can you submit a PR to fix it?

I don't actively work on eredis_cluster anymore because we use ered now. It is faster and it doesn't use a connection pool. Try it if you want: https://github.com/Ericsson/ered.