Metrics Overview Quiz
Last Updated: September 12, 2026
Which type represents a cumulative count that may reset?
- A) Gauge
- B) Counter
- C) A precomputed p99
- D) A scrape timestamp
Show answer
Answer: B
A Counter accumulates nonnegative increments. Resets can occur when the measured state is recreated. rate() handles observed resets, but cannot recover unobserved increments.
Five methods, twenty routes and ten statuses imply what?
- A) Exactly 1,000 stored series in every deployment
- B) At most 1,000 application label combinations if all combinations are possible
- C) Exactly 1,000 samples per day
- D) No effect on resource use
Show answer
Answer: B
The product is an upper bound. Actual combinations, target/replica labels, histogram buckets and historical churn determine the real series/storage footprint.
Which Pushgateway use is appropriate?
- A) Use one HOSTNAME grouping key for every short-lived Pod and rely on automatic expiry
- B) Use it for a suitable service-level batch with stable grouping, success timestamps and an explicit retirement policy
- C) Treat gateway up=1 as proof that every batch succeeded
- D) Push success timestamps even when the batch fails
Show answer
Answer: B
Pushgateway is not the default for all short-lived jobs and groups have no automatic TTL. Its scrape health is separate from batch freshness. Scraping with honor_labels preserves the pushed job identity.
Which statement about Histogram and Summary is correct?
- A) Summary quantiles are always exact
- B) Averaging instance p99 values gives fleet p99
- C) Compatible classic histogram buckets can be combined; Summary sum/count can be combined for a mean
- D) No Summary data can ever be aggregated
Show answer
Answer: C
Classic buckets are counted by the instrumented producer; Prometheus calculates quantiles at query time. Summary quantiles have algorithm/window-dependent error and are not aggregatable into a fleet quantile, while nonnegative-duration sum/count rates can yield a fleet mean.
Which is NOT the recommended convention for a new Prometheus application metric?
- A) Use a descriptive prefix
- B) Use a unit suffix such as _seconds or _bytes
- C) Prefer camelCase and millisecond units over the usual base-unit convention
- D) Use _total to identify a cumulative counter
Show answer
Answer: C
Prefer descriptive underscore-separated names and base units. _total is a counter marker, not a physical unit. Existing exporter APIs such as node_memory_MemAvailable_bytes retain their published spelling.
Which statement about Prometheus retention is correct?
- A) It can never retain more than 30 days
- B) The default is 15 days without explicit time/size retention settings; longer retention needs suitable configuration and capacity
- C) It does not compress local data
- D) Independent collection replicas are impossible without Mimir
Show answer
Answer: B
Default retention is not a maximum. Local TSDB is not a replicated distributed store; collection redundancy, query deduplication, durability and recovery are separate design decisions.
Which product/storage claim is incorrect?
- A) VictoriaMetrics single-node and cluster deployments have different operating requirements
- B) Traditional CloudWatch metric resolution becomes coarser with age
- C) Mimir object storage guarantees unlimited scale and eliminates every local storage requirement
- D) Datadog metrics use query rollups, so retention does not guarantee original resolution in every graph
Show answer
Answer: C
Object storage is part of Mimir's architecture, not a guarantee of unlimited capacity. Ingest/local resources, query limits, replication and operational capacity still matter. Do not conflate backup destinations or edition-specific features with a product's primary store.
Which approach fails to control metric cardinality?
- A) Use normalized route templates
- B) Avoid user/session IDs as ordinary labels
- C) Group status codes when losing detail is acceptable
- D) Assign a new request_id label value to every request
Show answer
Answer: D
Distinct label values create distinct series, including when values are hashed. Request-specific context belongs in appropriately controlled logs/traces when needed. Cardinality and sensitive-data exposure both require review.
Which Kubernetes metrics role is correctly matched?
- A) node-exporter — Kubernetes API object status
- B) kube-state-metrics — measured container CPU usage
- C) cAdvisor/kubelet metrics — container resource measurements
- D) metrics-server — long-term Prometheus TSDB
Show answer
Answer: C
node-exporter reports host OS metrics; kube-state-metrics exposes API object state; metrics-server serves the Resource Metrics API. Prometheus/vmalert/Mimir rules evaluate alerts, and Alertmanager routes them. vmagent is a collector/forwarder, not a queryable TSDB.
- What makes a cost comparison reviewable?
- A) A product ranking based only on team size
- B) The node count without sample interval or feature assumptions
- C) Measured series/sample volume, retention/resolution, HA/query requirements and current pricing for the selected features
- D) Assuming metric name/value lengths never matter
Show answer
Answer: C
One million actual exported series at 15-second intervals over 30 days implies 172.8 billion samples before delivery filtering/deduplication. Infrastructure, indexes/WAL, replicas, query work, custom-metric allowances and operator effort can change costs. This is a workload calculation, not a provider quote.