Update hiredis-cluster to 0.14.0 from 0.11.0

[plainbanana]

fix: #45

[plainbanana]

Need for Fork Behavior Change and Considerations

To ensure fork-safe code and reuse the cluster topology obtained by the parent process in the child process, several adjustments are made at Redis::Cluster::Fast using hiredis-cluster 0.11.0 .

Previous Behavior (0.10.0)

• Topology updates were performed synchronously when commands were issued.
• Right after fork, the event_base would have no unissued handlers, only holding the connection.
  • Issuing event_reinit followed by redisClusterAsyncDisconnect was sufficient to handle this.

Previous Behavior (0.11.0)

• Topology updates were performed asynchronously.
• Right after fork, the event_base may have unissued handlers (e.g., CLUSTER NODES).
• [BUG] Event loop occasionally hangs after redisClusterAsyncDisconnect under high connection error conditions

Upcoming Behavior (0.14.0)

• Topology updates will be performed asynchronously.
• Right after fork, the event_base may have unissued handlers (e.g., CLUSTER NODES).
• redisClusterAsyncDisconnect related bug is fixed 🎉
  • When redisClusterAsyncDisconnect is issued, the HIRCLUSTER_FLAG_SHUTDOWN flag will be set, and reconnection can only be achieved via redisClusterConnect2.

Two Options for Reconnection

1. Issue event_reinit and redisClusterAsyncDisconnect, then use redisClusterConnect2 to clear the shutdown flag and reconnect.
2. Create a new event_base and asynchronous context, then reconnect to the cluster.

The second option seems preferable.

In either option, the cluster topology will be updated. However, considering that there may still be unissued handlers from the asynchronous CLUSTER NODES command, it seems safer to create a new context entirely.

In option (2.), if there are remaining handlers for the asynchronous CLUSTER NODES, issuing event_base_free could result in dangling pointers.

In either option, To avoid creating dangling pointers, all remaining handlers must be executed. It is necessary to execute event_base_dispatch. Because the HIRCLUSTER_FLAG_SHUTDOWN flag is set, the topology update will not be performed and will return immediately.

Since there’s no significant difference, either method should work.

[plainbanana]

[MEMO]

I implemented (1.) as it allows for a more simple solution.

Issue event_reinit and redisClusterAsyncDisconnect, then use redisClusterConnect2 to clear the shutdown flag and reconnect.

Since there’s no significant difference, either method should work.