Skip to main content
In Lasso RPC Core v0.5.0, request execution supplies performance observations without requiring a separate benchmark request for every user call. Background health and head probes are separate upstream traffic. Measure their cost alongside your workload when qualifying a deployment.

​
Routing evidence and reporting

Recent routing evidence and dashboard summaries serve different purposes: Routing evidence is local to the node and scoped to the route, transport and a bounded registered workload key. A lifetime average or a high aggregate success rate does not substitute for recent qualified evidence on the relevant route. Provider aliases, shared upstreams and multiple nodes also affect how counts should be interpreted; do not add overlapping summaries as independent work.

​
How strategies use evidence

Fastest orders reliability-qualified candidates by recent successful mean latency, with successful p95 and identity tie-breakers. Unqualified candidates remain as fallbacks. Missing and stale evidence do not establish reliability. Latency weighted uses relative weights (best_mean / candidate_mean)^beta and exponential-race ordering -log(U) / weight. Reliability is a qualification boundary, not a success-rate multiplier. There is no weight floor or hidden exploration share. Available latency priors can still order unqualified fallbacks; unmeasured channels are shuffled. Load balanced starts from a randomized order and gives distinct physical instances a bounded first pass for replay-safe calls within each health tier. It does not rank by a benchmark score. Priority uses configured order. All strategies apply health tiering and live admission. The supported latency control is LW_BETA (positive number, default 3.0). See routing strategies for the full released contract.

​
Interpreting measurements

Keep chain, profile, transport, workload, source revision, time window and sample count with any latency result. Separate successful-attempt latency from end-to-end request latency: retries, selection and internal work can make them differ. Use attempted_channels and executed_channel in request metadata to understand a specific result. A percentile computed from one sample set cannot be reconstructed by averaging percentiles from other nodes. Report local percentiles or retain compatible raw samples for an aggregate calculation. Missing observations are unknown, not zero latency or successful service.

​
Retention and overhead

In v0.5.0, BenchmarkStore periodically trims raw per-chain samples. A burst can exceed that threshold before the next trim, and aggregated provider/method score entries have no fixed cardinality cap in this release. It is not a durable complete request journal. Monitor process and ETS memory for your configured profile, method, and provider mix. The released BenchmarkStore source defines the exact behavior; this page does not assign a fixed per-call overhead or memory footprint to every deployment. The historical routing-overhead report is a dated engine benchmark. Preserve its source and environment attribution; it is not v0.5.0 hosted capacity or an SLA. Versions and evidence links the current release acceptance evidence. Implementation references: Fastest, LatencyWeighted, and routing evidence.