Autoscaling

Enabling

Open the service's Settings → Scaling tab and switch to Autoscale. Set min and max replica counts (max ≥ 2 enables the scaler; min/max equal to 1 disables it). Tune the per-signal scale-up thresholds; scale-down uses 50% of each up-threshold.

Signals

Three signals drive scaling decisions. Any one breaching its up-threshold triggers scale-up; all three must sit below 50% of their up-threshold to trigger scale-down.

# Defaults — every value is per-service tunable in the dashboard.
CPU      avg across replicas > scaleCpuThreshold        (default 80%)
Memory   avg RSS / limit     > scaleMemoryThreshold     (default 80%)
Traefik  RPS into this svc   > scaleRequestsPerSecThreshold (off by default)

# Scale-down requires all three:
CPU avg     < 0.5 × scaleCpuThreshold     (e.g. < 40%)
Memory avg  < 0.5 × scaleMemoryThreshold  (e.g. < 40%)
RPS         < 0.5 × scaleRequestsPerSecThreshold (or RPS threshold unset)

RPS is scraped from Traefik's own Prometheus metrics endpoint — it counts every HTTP request hitting the service's router, regardless of upstream replica. Set a threshold value to opt the service in; leave it unset and the autoscaler falls back to CPU + memory only.

Anti-flap (sustained-reading gate)

The agent doesn't act on a single spike. A scale-up requires 4 consecutive above-threshold readings (≈ 2 minutes at the 30-second tick); scale-down requires 10 consecutive below-threshold readings (≈ 5 minutes). A reading on the opposite side resets the counter, so a noisy burst doesn't accumulate toward an unwanted scale event.

On top of that, only one scale action runs at a time per agent — the in-flight action holds a mutex so a second tick can't double-scale during a 30-second graceful drain.

Hard bounds

• The scaler never exceeds maxInstances or drops below minInstances, even if signals say so.
• A successful scale event resets the relevant counter to 0, so the next decision starts fresh.
• Scale events emit a structured scale_event on the activity feed with the triggering metric value.

When to Use Autoscale vs Fixed

• Autoscale when your traffic is spiky (marketing campaigns, geo-skewed peak hours, queue-backed async workers).
• Fixed when your workload is steady and predictable, or when you want hard budget control on the resource line item.

Billing

Replicas are billed per-second at the service-tier rate. Scaling from 1 → 3 replicas triples the rate while the extra replicas are running and drops back the moment they're released. The dashboard's Billing tab shows the per-resource hourly breakdown.