Load balancing configuration

[Geal]

Right now we have two structures governing load balancing, one to choose the algorithm (RoundRobin, Random, LeastConnections, PowerOfTwo), and one to define load balancing parameters (only the weight for now, and it is not even used).

Here's what I want to change:

• algorithms: round robin, random, least loaded, power of two, sticky session
• weight parameter can be used by all algorithms
• load measurement parameter is how we measure the load of a server. An enum of: least connections, least active requests, latency EWMA, peak EWMA (latency wighted by the number of active requests)
• sticky id parameter

[Geal]

5535555

[Geal]

we're still missing the request latency measurement per backend

[FlorentinDUBOIS]

Status after the 2.0.x rewrite and the UDP datapath (#1274):

• Algorithms: ✅ ROUND_ROBIN, RANDOM, LEAST_LOADED, POWER_OF_TWO, plus flow-affine HRW and MAGLEV (lib/src/load_balancing.rs). Sticky sessions exist as Cluster.sticky_session + per-backend sticky_id rather than as a named LB algorithm.
• sticky_id parameter: ✅ (AddBackend.sticky_id).
• Weight: partial — honored by RANDOM (WeightedIndex), HRW (score share) and MAGLEV (slot share) via backend_weight() (clamped ≥ 1); ROUND_ROBIN, LEAST_LOADED and POWER_OF_TWO still ignore it.
• Load metric: partial — LoadMetric offers CONNECTIONS, REQUESTS, CONNECTION_TIME (peak-EWMA of backend connect time). A per-backend request-latency measurement (latency EWMA) is still missing, as noted in the comment above.

Keeping this open to track the two remaining gaps: universal weight support and a request-latency load metric.