SRE Capacity Planning & Scaling 2 — Questions and Answers
Question 1: Which metric is most useful for predicting when a database will run out of disk space?
- Current disk usage percentage
- Linear regression on historical growth rate (Correct answer)
- Peak query latency
- Number of active connections
Correct answer: Linear regression on historical growth rate
Linear regression on historical growth trends provides a time-based forecast, allowing teams to predict when capacity will be exhausted.
Question 2: An SRE observes that CPU utilization spikes to 95% every Monday morning for 30 minutes. What is the best scaling strategy?
- Permanently provision double the CPU capacity
- Configure scheduled auto-scaling to add capacity before Monday morning (Correct answer)
- Ignore it since the spike is brief
- Vertically scale the instance class permanently
Correct answer: Configure scheduled auto-scaling to add capacity before Monday morning
Scheduled auto-scaling proactively adds resources before a known predictable load pattern, avoiding over-provisioning at other times.
Question 3: What does 'headroom' refer to in capacity planning?
- The latency budget for a single request
- Reserved capacity above current usage to absorb unexpected demand (Correct answer)
- The number of redundant nodes in a cluster
- Memory allocated to OS kernel operations
Correct answer: Reserved capacity above current usage to absorb unexpected demand
Headroom is the buffer of spare capacity maintained above current utilization to handle unexpected traffic bursts without degrading service.
Question 4: A service's p99 latency degrades significantly when CPU utilization exceeds 70%. What should the SRE set as the scaling threshold?
- 90%
- 70% or below (Correct answer)
- 50%
- 100%
Correct answer: 70% or below
Scaling should trigger at or before the 70% threshold to prevent latency degradation, ensuring performance SLOs are maintained.
Question 5: Which load testing approach best simulates realistic production traffic for capacity planning?
- Synthetic constant-rate load at 2× peak
- Replay of recorded production traffic traces (Correct answer)
- Random request generation at maximum throughput
- Testing with a single large payload repeatedly
Correct answer: Replay of recorded production traffic traces
Replaying recorded production traffic captures real-world request distributions, user patterns, and payload sizes for accurate capacity modeling.
Question 6: What is the primary risk of setting auto-scaling cooldown periods that are too short?
- Instances never scale down
- Thrashing, where the system repeatedly scales up and down (Correct answer)
- Inability to handle burst traffic
- Increased provisioning costs only
Correct answer: Thrashing, where the system repeatedly scales up and down
Short cooldown periods cause scaling thrashing, where instances are repeatedly added and removed in rapid succession, destabilizing the system.
Question 7: In the context of capacity planning, what is 'demand forecasting'?
- Measuring current system throughput
- Predicting future resource needs based on business and traffic growth trends (Correct answer)
- Calculating the cost of running a service
- Alerting on real-time capacity exhaustion
Correct answer: Predicting future resource needs based on business and traffic growth trends
Demand forecasting uses historical trends, business growth projections, and seasonality to predict future infrastructure resource requirements.
SRE Capacity Planning & Scaling 2
Which metric is most useful for predicting when a database will run out of disk space?