Autoscaling and cold starts
Desired replicas and ready capacity are different states.
Predict first
Before you read: predict
A catalog worker group has 10 Ready replicas, complete metrics averaging 54 units per replica, and a target of 50. Limits and rate policies do not bind; the default tolerance applies.
Under the simplified HPA model, what replica recommendation results?
The 30-second version
Autoscaling requests capacity; readiness follows later. A simplified HPA multiplies current replicas by the metric-to-target ratio and rounds upward, unless tolerance suppresses the change. Sampling, boot and warm-up delay usable capacity. Downscale stabilization retains recent higher recommendations. Scheduled preparation or initialized reserves can cover known demand; backlog metrics can reveal waiting work that CPU misses.
Desired replicas and ready capacity are different states.
How it works
Step 1 of 7
Analogy: Preparing ice sheets
An ice arena accepts a tournament booking before the players arrive. The manager can commission another sheet, but commissioning does not mean skaters can use it immediately. Staff must prepare and inspect the surface. If the manager only checks demand periodically, a group arriving just after inspection waits until the next inspection even before preparation starts. A ready reserve sheet can avoid some preparation work, but keeping it ready consumes resources.
| Ice-rink operation | Autoscaling concept |
|---|---|
| Periodic attendance check | Metric sampling |
| Commissioned sheet | Desired replica |
| Prepared, inspected surface | Ready worker |
| Tournament booking | Scheduled preparation |
| Maintained reserve sheet | Initialized capacity |
The mapping concerns readiness, not rink construction speed. Physical ice sheets do not resize like processes, and a reservation is not a universal guarantee of immediate service. Different services reserve different resources. AWS Lambda reserved concurrency allocates a concurrency allowance; provisioned concurrency initializes execution environments. An EC2 warm pool keeps instances beside an Auto Scaling group, potentially stopped or hibernated, so activation can still have work remaining. These distinctions matter before treating every product called a reserve as equivalent. AWS concurrency Warm pools
Try it
Answer first, then check. Nothing is saved.
Question 1 of 5
8 replicas are running. Recommendation history contains 8 at virtual second 30 and a raw recommendation of 2 at second 120. At second 120, default downscale stabilization applies; other policies do not bind.
Which recommendation survives?
Question 2 of 5
A catalog Lambda release needs a configured baseline that can respond without allocation-time initialization delay.
Which capacity feature supplies that operating condition?
Question 3 of 5
A worker is requested at virtual second 15. Its boot phase takes 30 virtual seconds and its subsequent warm-up takes 15. The endpoint returns 200 once ready and 503 while warming.
What status sequence appears when checked at seconds 15, 45 and 60?
Question 4 of 5
Demand arrives just after virtual second 0. The controller samples every 15 seconds; after detection, two requested catalog workers each take 30 seconds to boot and 15 seconds to warm.
When can those workers first answer?
Question 5 of 5
A catalog controller is configured as metric: queue-depth with target: 10 messages/replica. The observed queue depth is 31; the dashboard also shows CPU 5 against target 50, and a previous recommendation was 8.
What desired replica count does the configured metric produce?
Go deeper
Pro tip
Keep metric identity and target units beside a scaling trace so reviewers can interpret the observation. Kubernetes metrics
Sources
Read and fetched 2026-09-30; full snapshots with URL/date/SHA256 are in this lesson's sources directory: Kubernetes HPA, KEDA 2.18 cron, KEDA 2.18 SQS, KEDA 2.18 Prometheus, AWS EC2 warm pools, AWS Lambda provisioned concurrency, AWS Lambda concurrency.
P2 passage ledger, read 2026-09-30: A uses AWS provisioned concurrency: “These execution environments are ready to respond immediately to incoming function requests.” B uses AWS concurrency, in its reserved-concurrency discussion: “However, none of these environments come pre-initialized.” C uses that page's request table: “This is the first request; no execution environment instances are available.”
Related lessons
Checked against the tools and docs as of 2026-09-30.