01.06 · Lecture
Feedback, Lag, Overshoot
Define a feedback loop - an input, an output, a measurement of that output, and an error signal - and say what the loop is trying to hold steady.
A feedback loop compares a measured output with a target, turns the gap into an error signal, and changes an input to keep a chosen variable steady. In deployment systems, lag between action and measurement can make corrections arrive late, causing overshoot, oscillation, wasted capacity, or overload.
What this lesson answers
- what is a feedback loop in deployment
- why autoscaling overshoots after traffic spikes
- how lag affects production control loops
Notes
Feedback, Lag, Overshoot — Feedback loops exist to keep a system variable near a target by measuring output and correcting error; without them, load, latency, temperature, or queue depth can drift until the system becomes unstable or fails.
Key Concepts: - A feedback loop has 4 parts: input, output, measurement, and error signal, where . - The loop is trying to hold a controlled variable steady, such as CPU utilization at , p95 latency under , or queue depth below messages.
Common questions
- What is the error signal in a feedback loop?
- The error signal is the difference between the target value and the measured output. It tells the controller both direction and size of correction. In an autoscaler, that might mean comparing current utilisation or queue depth with the desired operating point, then changing replica count, concurrency, or admitted work.
- Why does lag make autoscaling unstable?
- Lag means the system is acting on information from the past. A pod, worker, or function may take time to start, and metrics may arrive late. During that delay, the controller can keep applying corrections that were sensible earlier but are now too large, producing overshoot or repeated scale up and scale down cycles.
- How is negative feedback different from positive feedback?
- Negative feedback reduces the gap between measured behaviour and the target, such as adding capacity when utilisation is too high. Positive feedback amplifies a change instead. Immediate aggressive retries during an outage are a common production example, because they add more load to a system that is already failing.