Skip to content

Modal Cold Start Spike

Cold starts are spiking — new Modal containers are starting frequently, causing high lifecycle.cold_start metric values and intermittent latency spikes on first requests.

Symptoms

  • lifecycle.cold_start counter increases rapidly in metrics
  • First-request latency spikes (> 2s) visible in http.server.request.duration histogram
  • Modal dashboard shows frequent container restarts
  • No application errors — subsequent requests to the same container are fast

Diagnosis

1. Identify the cause of container cycling.

modal app logs <app-name> | grep "cold_start\|container\|start"

Containers restart due to: OOM kills, unhandled exceptions in @enter, scale-to-zero + scale-out cycles, or Modal infrastructure events.

2. Check memory usage.

If containers are OOM-killed, lifecycle.cold_start spikes coincide with lifecycle.oom_kill events (if recorded). Check Modal dashboard for memory metrics.

3. Check setup_telemetry() timing.

setup_telemetry() is called in @enter (the Modal equivalent of lifespan). If it throws (e.g. invalid OTLP endpoint), the container may restart in a loop.

Recovery

OOM kills: Increase container memory in modal_app.py:

@app.function(memory=2048)   # MB

setup_telemetry() crashing: setup_telemetry() is designed to never raise — it degrades to no-op mode and logs a warning. If it is raising, check the OTel SDK version and the OTEL_EXPORTER_OTLP_ENDPOINT format.

Scale-to-zero thrash: Set min_containers=1 in modal_app.py to keep at least one warm container:

@app.function(min_containers=1)

Prevention

Call record_lifecycle_event("cold_start") at the top of @enter. This populates lifecycle.cold_start in metrics and makes cold start frequency visible before it becomes a latency problem.