Read a recommendation snapshot

A snapshot records resource advice, desired replicas, decisive factors, sources, version, and expiry. It explains what to configure and why; execution records show what was actually applied.

Allow time for capacity to become ready

Scaling only when demand arrives may be too late. Look-ahead decisions account for collection, prediction, control loops, Pod readiness, and node preparation where needed, then evaluate the corresponding future window.

Scale-up and scale-down use different perspectives. Reductions require caution across future demand, live load, and readiness rather than reacting to a brief dip.

Multi-model consensus

Available candidates participate in the look-ahead window. Confidence combines scores with recent tracking error. A configured weighted quantile combines predictions before live load and safeguards settle the decision.

Read consensus alongside the final replica count, constraints, and actual use. Without eligible predictions, use schedules, fallback, or unchanged values according to module rules.

Resource and replica constraints

  • CPU, memory, throughput, or JVM pressure may be decisive.
  • Replica changes respect minimums, maximums, hard ceilings, steps, and cooldowns.
  • OOM floors and memory peaks provide protection beyond smoothed point forecasts.
  • Configured or estimated per-Pod capacity needs comparison with workload load tests.
Advice, writes, and readiness are separate stages

A submitted replica change does not mean the new Pods are Ready. Inspect actual replicas, readiness, and subsequent resource behavior.

Forecast far enough for capacity to arrive

For modules planning replica scaling, the effective horizon accounts for the module horizon, run interval, and readiness lead time. Default lead-time coverage is twice the lead time, capped at 24 hours, with the reason shown in recommendation details. Once enough scale-up outcomes exist, lead time uses measured readiness for that module and allows for recent late arrivals.

Candidates that substantially lag recent actuals are excluded from consensus when healthy alternatives exist. Configurable emergency scale-up can cover observed demand above both forecasts and current capacity. This path only scales up and still respects replica ceilings.