Cheat sheet & self-check

Pick an answer to see if you got it, and why.

1. Q1. You created a ServiceMonitor, but Prometheus doesn't scrape the target. What's the classic cause with kube-prometheus-stack?

   1. APrometheus needs a restart
   2. BThe Prometheus resource only selects ServiceMonitors matching its selector (by default, the Helm release label) and namespace selector; add the label or relax the selector
   3. CServiceMonitors don't work with Deployments
   4. DThe port must be 9090
   Show answer

   B. Check Prometheus's serviceMonitorSelector/NamespaceSelector. The chart value serviceMonitorSelectorNilUsesHelmValues=false makes it select all.

   From lesson 01 · kube-prometheus-stack baseline

2. Q2. What does kube-state-metrics provide?

   1. ANode CPU metrics
   2. BMetrics about the state of Kubernetes objects (deployments' desired vs available replicas, pod phases, restarts, resource requests)
   3. CApplication traces
   4. Detcd backups
   Show answer

   B. node-exporter covers the machine; kube-state-metrics covers object state from the API; cAdvisor (kubelet) covers container usage.

   From lesson 01 · kube-prometheus-stack baseline

3. Q3. Why use rate() on a counter before summing?

   1. AIt's optional
   2. BCounters only go up (and reset on restart); rate() turns them into per-second increases and handles resets, which can then be aggregated
   3. Csum() doesn't work on counters
   4. DFor formatting
   Show answer

   B. Always rate first, then sum: sum(rate(x[5m])), never rate(sum(x)).

   From lesson 01 · kube-prometheus-stack baseline

4. Q4. What mainly drives Prometheus memory usage?

   1. ANumber of dashboards
   2. BThe number of active series in the head block (plus query load)
   3. CDisk size
   4. DScrape timeout
   Show answer

   B. Each active series has in-memory structures, recent samples and index entries. Double the series ≈ much more memory.

   From lesson 02 · Why Prometheus doesn't scale

5. Q5. A metric has labels method (5 values), status (10), endpoint (50) and pod (40). Roughly how many series can it create?

   1. A105
   2. BUp to 100,000 (5 × 10 × 50 × 40)
   3. C40
   4. D5,000
   Show answer

   B. Cardinality multiplies across labels. Adding one unbounded label (user ID) makes it effectively infinite.

   From lesson 02 · Why Prometheus doesn't scale

6. Q6. Why isn't 'just keep a year of data locally' a good plan for one Prometheus?

   1. APrometheus can't store a year
   2. BLocal disk bounds retention and durability, queries over long ranges get slow, and each cluster's Prometheus has only its own data; there's no global view or dedup of HA pairs
   3. CIt's illegal
   4. DGrafana can't query it
   Show answer

   B. Long-term storage and a global view are exactly what Thanos and Mimir add (lessons 04–07).

   From lesson 02 · Why Prometheus doesn't scale

7. Q7. Why do Thanos and Mimir store long-term metrics in object storage?

   1. AIt's the fastest storage
   2. BIt's cheap, durable and effectively unlimited, and every component can read the same blocks over the network, which decouples storage from compute
   3. CPrometheus requires it
   4. DIt avoids compression
   Show answer

   B. Immutable TSDB blocks fit object storage perfectly; compute (queriers, store gateways) scales separately.

   From lesson 03 · MinIO object storage

8. Q8. What does erasure coding give an S3-compatible cluster like MinIO or Ceph?

   1. AEncryption
   2. BData split into data + parity shards across drives/nodes, so it survives several drive or node failures with less overhead than full replication
   3. CFaster DNS
   4. DAutomatic backups to the cloud
   Show answer

   B. For example, 8 data + 4 parity tolerates 4 lost drives with 1.5× overhead instead of 3× for triple replication.

   From lesson 03 · MinIO object storage

9. Q9. Your team plans to use MinIO's open-source edition. What should you check first?

   1. ANothing
   2. BIts current licensing (AGPLv3), feature set and how builds are distributed, since these changed significantly in 2025; compare with alternatives such as Ceph RGW or a managed S3
   3. COnly the logo
   4. DWhether it supports SQL
   Show answer

   B. Thanos and Mimir only need an S3-compatible API, so choose the store you can support long term.

   From lesson 03 · MinIO object storage

10. Q10. Why are external labels (like cluster) essential with Thanos?

    1. AFor colours in Grafana
    2. BThey identify which Prometheus each series came from, so the global view doesn't mix identical series from different clusters, and replica labels enable dedup
    3. CThey speed up scraping
    4. DThanos ignores them
    Show answer

    B. Without unique external labels, series from different sources collide in the global view.

    From lesson 04 · Thanos: sidecar, store, querier

11. Q11. Two Prometheus replicas scrape the same targets. How does the Thanos querier show one clean series?

    1. AIt averages them
    2. BDeduplication: series that differ only by the replica label are merged, picking samples from one replica and filling gaps from the other
    3. CIt shows both always
    4. DIt deletes one replica
    Show answer

    B. Pass --query.replica-label with the replica label name (prometheus_replica for the Operator).

    From lesson 04 · Thanos: sidecar, store, querier

12. Q12. What does the store gateway do?

    1. AScrapes targets
    2. BServes data from blocks in object storage to queriers via the StoreAPI, using index caches so it doesn't download whole blocks
    3. CUploads blocks
    4. DSends alerts
    Show answer

    B. Sidecars cover recent hours; store gateways cover everything already in the bucket.

    From lesson 04 · Thanos: sidecar, store, querier

13. Q13. How many compactors should run against one bucket?

    1. AOne per querier
    2. BExactly one (per bucket, or per disjoint set of blocks), because concurrent compactors on the same blocks corrupt or duplicate data
    3. CAs many as possible
    4. DZero
    Show answer

    B. The compactor is a singleton by design. Run it as a single-replica StatefulSet with enough disk.

    From lesson 05 · Thanos: compactor, downsampling, hot/cold

14. Q14. Does downsampling reduce storage usage?

    1. AYes, by about 90%
    2. BNot by itself: it adds 5m and 1h copies; it makes long-range queries fast. Storage drops only if you set shorter retention for raw data
    3. CIt deletes raw data immediately
    4. DIt compresses chunks
    Show answer

    B. Keep raw data short, downsampled data long: that's where the savings come from.

    From lesson 05 · Thanos: compactor, downsampling, hot/cold

15. Q15. When are blocks downsampled by default?

    1. AImmediately
    2. B5-minute resolution for blocks older than about 40 hours; 1-hour resolution for blocks older than about 10 days
    3. COnly on request
    4. DAfter one year
    Show answer

    B. Downsampling needs big enough compacted blocks; that's why it happens after these ages.

    From lesson 05 · Thanos: compactor, downsampling, hot/cold

16. Q16. You need 2 years of request-rate trends per service, but raw per-pod series are only useful for 2 weeks. What do you do?

    1. AKeep everything for 2 years
    2. BRecord per-service aggregates with recording rules into a long-retention tier, and keep raw per-pod series in a short-retention tier
    3. CDelete all metrics after 2 weeks
    4. DIncrease the scrape interval
    Show answer

    B. Aggregates have far fewer series; keeping them long costs a fraction of keeping raw data.

    From lesson 06 · Retention by data priority

17. Q17. What's the cheapest place to get rid of a metric nobody uses?

    1. AIn Grafana
    2. BAt scrape time with metric relabeling (or not exposing it at all), so it never enters storage
    3. CIn the compactor
    4. DIn object storage lifecycle
    Show answer

    B. Dropping early saves memory, network, storage and query time everywhere downstream.

    From lesson 06 · Retention by data priority

18. Q18. How do you know which metrics are unused?

    1. AGuess
    2. BCompare stored metric names with those referenced by dashboards, recording/alerting rules and query logs (tools like mimirtool automate this)
    3. CDelete randomly and wait for complaints
    4. DLook at the scrape interval
    Show answer

    B. Usage analysis often finds that a large share of stored series are never queried.

    From lesson 06 · Retention by data priority

19. Q19. How does data get into Mimir?

    1. AMimir scrapes all targets itself
    2. BPrometheus (server or agent mode), the OTel Collector or Alloy remote-write samples to Mimir's distributors
    3. CVia the sidecar only
    4. DBy copying Prometheus disks
    Show answer

    B. Mimir is push-based. Local Prometheus/agents do the scraping; Mimir stores and serves.

    From lesson 07 · Grafana Mimir

20. Q20. What does the ingester replication factor of 3 mean?

    1. AThree copies in object storage
    2. BEach series is written to 3 ingesters, so the loss of one ingester doesn't lose recent data or block writes
    3. CThree tenants
    4. DThree queriers
    Show answer

    B. Recent (not yet uploaded) data lives in ingesters; replication protects it. Object storage has its own durability.

    From lesson 07 · Grafana Mimir

21. Q21. Why are per-tenant limits important in Mimir?

    1. ALicensing
    2. BThey stop one tenant's cardinality explosion or ingestion spike from overloading the shared cluster (series limits, ingestion rate, query limits)
    3. CThey speed up Grafana
    4. DThey're only for billing
    Show answer

    B. Multi-tenancy without limits is a shared failure domain.

    From lesson 07 · Grafana Mimir

22. Q22. Which is a strong reason to pick Thanos with sidecars?

    1. ANative multi-tenancy with per-tenant limits
    2. BYou want to keep existing Prometheus servers as they are and add a global view and long-term storage incrementally, without changing the write path
    3. CYou need no object storage
    4. DYou want AGPL licensing
    Show answer

    B. Sidecars bolt on to what you have. Mimir changes the architecture to push-based central storage.

    From lesson 08 · Thanos vs Mimir

23. Q23. Which is a strong reason to pick Mimir?

    1. AYou need downsampling
    2. BYou're building a central, multi-tenant metrics service for many teams with strict per-tenant limits and very high scale
    3. CYou can't run object storage
    4. DYou have only one small cluster
    Show answer

    B. Multi-tenancy, limits and horizontal scale of every component are Mimir's core strengths.

    From lesson 08 · Thanos vs Mimir

24. Q24. What's a real risk of the sidecar model for remote sites?

    1. ANone
    2. BThe central querier must reach every sidecar for recent data; if a site is unreachable, its last hours are missing from global queries until blocks upload
    3. CSidecars can't upload
    4. DQueries become AGPL
    Show answer

    B. Push models (Receive/Mimir with agents) buffer and send; pull models need connectivity at query time.

    From lesson 08 · Thanos vs Mimir

25. Q25. What is Prometheus federation good for?

    1. ACopying all raw data to a central server
    2. BPulling a small set of already-aggregated series (recording rules) from many Prometheus servers into a higher-level one
    3. CLong-term storage
    4. DDeduplicating HA pairs
    Show answer

    B. Federating raw series at scale overloads both sides. Use it for aggregates; use remote write or global query for the rest.

    From lesson 09 · Federation & multi-cluster

26. Q26. Why run Prometheus in agent mode on edge clusters?

    1. AIt stores more data
    2. BIt uses fewer resources (no local querying, rules or long-term TSDB), buffers in a WAL, and remote-writes everything to the central store
    3. CIt has a better UI
    4. DIt doesn't need scraping
    Show answer

    B. Edge clusters then only need to scrape and forward; queries and alerting happen centrally.

    From lesson 09 · Federation & multi-cluster

27. Q27. An edge site loses its uplink for 6 hours. With remote write, what's at risk?

    1. ANothing, ever
    2. BSamples older than what the WAL keeps (by default only a few hours) can't be sent after reconnecting; size the WAL/retention for expected outages or accept gaps
    3. CThe site's pods restart
    4. DCentral Mimir deletes the tenant
    Show answer

    B. Remote write reads from the WAL. Plan buffering for your realistic outage lengths, and alert on growing pending samples.

    From lesson 09 · Federation & multi-cluster

28. Q28. 500,000 active series scraped every 30 seconds produce roughly how many samples per second?

    1. A500
    2. BAbout 16,700
    3. C500,000
    4. D15 million
    Show answer

    B. 500,000 / 30 ≈ 16,667 samples per second.

    From lesson 10 · Capstone: cost, HA & the platform view

29. Q29. Why run a separate, small 'meta-monitoring' Prometheus?

    1. ATo double the data
    2. BTo watch the main monitoring stack from outside it, so a failure of the main stack is still detected and alerted
    3. CIt's required by Thanos
    4. DFor Grafana plugins
    Show answer

    B. Monitoring can't reliably report its own death. Combine with a dead man's switch (Watchdog) to an external service.

    From lesson 10 · Capstone: cost, HA & the platform view

30. Q30. What makes a capacity estimate trustworthy in a design review?

    1. ABig numbers
    2. BStated assumptions (series, interval, bytes/sample, replication, retention), a measured baseline from a real cluster, headroom, and a plan to re-measure
    3. CVendor benchmarks only
    4. DRounding everything to powers of two
    Show answer

    B. Reviewers can challenge assumptions; measurements anchor them.

    From lesson 10 · Capstone: cost, HA & the platform view