PREDICTIVE WARM-UP OF MICROSERVICES USING HIDDEN SEMI-MARKOV MODELS

Resumen

Predictive warm-up of microservices reduces cold-start penalties by anticipating incoming requests
and preemptively initializing service instances. We introduce a strategy that leverages a hidden semiMarkov model calibrated with empirical startup and idle durations, together with a Bayesian decision
framework, to trigger warm-up actions. We hypothesize that this approach can significantly reduce the
95th-percentile request latency (P95) while keeping infrastructure costs comparable to or lower than
those of conventional reactive autoscaling. To test this, we collect invocation traces from a representative Kubernetes-based deployment, fit the HSMM to capture idle sojourn times and probabilistic transitions driven by upstream calls, derive a decision threshold that balances residual cold-start risks against
unnecessary warm-ups, and evaluate the approach through a discrete-event simulation. Experimental
results demonstrate up to a 48.7% reduction in P95 latency relative to reactive policies, together with a
2.0% decrease in computed usage, confirming that precise probabilistic modeling and Bayesian decision-making can deliver substantial performance gains with minimal cost overhead.