Card 1 of 20
Front. Failover Load. Three equal zones each run at 50% capacity. What happens to utilization if one fails and traffic redistributes evenly?
These cards are original study material. They are not questions from a named employer or a live hiring loop.
All 20 cards in this deck
1.Failover Load
Three equal zones each run at 50% capacity. What happens to utilization if one fails and traffic redistributes evenly?
The remaining two carry 1.5 times their previous load, reaching 75%, assuming linear capacity and no additional failover overhead. Validate that this remains below the latency knee.
2.Headroom
Why is average CPU headroom insufficient evidence that a service can survive losing a zone?
Memory, connections, hot shards, network, and downstream capacity may be limiting, and redistribution may be uneven. Test the complete failure workload.
3.Queue Growth
A queue receives 1,200 jobs per second and completes 1,000. How fast does backlog grow?
By 200 jobs per second, or 12,000 per minute, while those rates persist. Increasing queue storage does not resolve the capacity deficit.
4.Drain Time
A backlog holds 60,000 jobs; workers complete 1,500 per second while 1,000 new jobs arrive. What is ideal drain time?
60,000 / (1,500 - 1,000) = 120 seconds, assuming stable rates. Use net spare throughput rather than total worker throughput.
5.Concurrency Law
A stable endpoint completes 400 requests per second with mean time in system 0.2 seconds. What average concurrency follows?
Little's Law gives 400 × 0.2 = 80 in-flight requests within that measurement boundary. It does not determine a safe peak or p99 concurrency.
6.Service Variance
Why can two workloads with the same mean service time produce different tail latency?
Service-time and arrival variability affect queueing. A few very expensive operations can block many short ones; preserve distributions in capacity tests.
7.Load Knee
Why operate below the throughput point where latency rises sharply?
Near saturation, small bursts or slowdowns create large queues. The margin should cover realistic variability and failure load rather than a universal utilization percentage.
8.Goodput
Why can total requests per second rise while useful service capacity falls?
Retries, rejected calls, and timed-out work inflate throughput. Measure successful operations completed within their deadlines and account for wasted work separately.
9.Closed Loop
Why can fixed-concurrency clients understate the demand an overloaded service would face?
They issue fewer new requests as responses slow. Use an arrival-driven workload when external arrivals continue independently of completion.
10.Generator Proof
What must a load test verify before reporting the server's maximum throughput?
That generators delivered the intended arrival rate and were not limited by CPU, sockets, network, or client scheduling. Measure offered load separately from accepted and completed load.
11.Cold Start Scale
Why can autoscaling fail to absorb a short traffic spike even with an accurate metric?
Detection, scheduling, image pull, startup, and warmup take time. Maintain headroom or pre-scale when the spike is shorter than the scale-out response.
12.Scaling Signal
When is queue age a better scaling signal than average CPU?
When workers are blocked on I/O or the user deadline depends on backlog delay. Confirm scaling workers can increase the actual bottleneck's throughput.
13.Scale Down
Why can aggressive scale-down create a repeating overload cycle?
Capacity disappears before delayed metrics or queued work reflect its need, causing another scale-up after startup delay. Use stabilization and drain semantics based on observed dynamics.
14.Connections
A deployment doubles pods temporarily, each with 40 database connections. What capacity must be reviewed?
The database's aggregate connection and query-concurrency budget during surge, plus other clients. Per-pod defaults can exhaust a shared dependency during an otherwise healthy rollout.
15.Cache Failure
Why should capacity tests include a degraded or cold cache?
Backend demand and latency can change drastically on misses. A system sized only for a warm hit rate may collapse during deployment, eviction, or cache outage.
16.Hot Shard
Cluster-wide utilization is low but one shard times out. What does adding generic capacity risk missing?
Skew concentrates work on one owner. Inspect key distribution and shard limits; repartition or change the data access pattern when extra idle nodes cannot serve that shard.
17.Fanout Tail
Why does a request that waits for every one of many dependencies have a worse latency tail?
Completion is governed by the slowest required result. Measure joint behavior and shared causes; independence assumptions can badly underestimate correlated delays.
18.Quota Leadtime
Why should capacity planning include procurement and quota lead time?
A predicted shortage cannot be fixed instantly if hardware, region capacity, or service quotas take time. Trigger action early enough to cover forecast uncertainty and approval delays.
19.Unit Economics
How can a capacity change improve throughput but still be a poor operational tradeoff?
It may cost more per useful operation or worsen failure behavior. Compare cost, latency, successful throughput, and headroom under the same workload.
20.Capacity Retest
Which changes should trigger revalidation of a service's capacity model?
Material changes in request mix, data size, dependencies, runtime, hardware, or failure topology. A historic requests-per-second number is conditional on its measured workload and environment.