lazy connection support

[justinreddit]

We have a setup where our service acts as a proxy that fans out requests across a set of clusters, individual processes may not serve requests that span a large number of these clusters. It looks like rueidis maintains connections to every node and we're starting to observe pretty heavy goroutine overhead.

Have there been thoughts about only establishing connections as needed?

[rueian]

rueidis should do lazy connection already. It only has connections that it talked to. However, when refreshing the cluster topology, it will keep querying the topology from every 4 nodes until it receives a successful response. This is currently not configurable.

If you serve concurrent requests, auto-pipelining will kick in and start additional goroutines. These could be the reasons why you saw heavy goroutine overhead. Would you mind sharing the goroutine profile?

You may try DisableAutoPipelining=true to avoid additional goroutines for auto-pipelining.