Pool distributes load unevenly across connections.

Bug Description

Pool distributes load unevenly across connections, prioritizing them by creation time. See here.

Where your connections are load-balanced across instances (pods/containers), this results in overuse of some instances and underuse of others.

Reproducible By

1. Create a service, ServiceA, with an endpoint that takes 100ms - 200ms to respond.
2. Using Kubernetes, deploy 5 pods with instances of ServiceA.
3. In another service, ServiceB, create a Pool of connections to ServiceA, referencing it by Kubernetes Service address (https://service-name), such that we'll be load-balancing connections through kube-proxy.
4. Send 1 request to ServiceB every 50ms for 1 minute.
5. Observe balanced connections across ServiceA pods, but imbalanced request distribution and resource usage across ServiceA pods.

Thinking through, this may be otherwise observable plainly through internal diagnostics around amount of requests processed by each Client or something like that.

Expected Behavior

Roughly balanced request distribution and resource usage across Clients and ServiceA pods.

Environment

MacOS, Node v18

Additional context

There are a number of ways we could resolve this intelligently. Most simply/crudely though, we could track index of last used client and pick back up next scan from that index to roughly round-robin.

Something like:

...
const totalClients = clients.length;
let clientIndex = (this.lastUsedIndex + 1) % totalClients;

for (let i = 0; i < totalClients; i++) {
  const client = clients[clientIndex];

  if (!client[kNeedDrain]) {
    this.lastUsedIndex = clientIndex;
    return client
  }

  clientIndex = (clientIndex + 1) % totalClients;
}

Start with the pool connection-selection logic in lib/dispatcher/pool.js around lines 75-87, then review the linked reproduction scenario and any diagnostics for per-client request counts. The change is complete when requests and resource use are roughly balanced across available clients under the described workload, with regression coverage for the observed imbalance.