Skip to main content
Source

This page is generated from devops-agent/eks-operation-review/references/observability.md. Edit the source, not this page.

Observability

Purpose​

Assess observability across three layers: control plane, data plane (nodes), and workloads — covering metrics, logs, and alerting.

Checks to Execute​

4.1 — EKS Control Plane Logging​

What to check:

  • Which of the 5 log types are enabled (api, audit, authenticator, controllerManager, scheduler)
  • CloudWatch log group existence and retention policy

How to check:

  1. Describe cluster → logging.clusterLogging → check each entry for enabled: true and which types. If eks:DescribeCluster returns 403/Forbidden → mark the control-plane-logging signal UNKNOWN (do not rate RED "logging disabled" on a forbidden describe).
  2. Use CloudWatch tools to check log group /aws/eks/{cluster-name}/cluster retention (the >= 30-day audit-log recommendation is a skill-defined heuristic, not an AWS-published requirement; no retention policy = logs kept forever at cost). If logs:DescribeLogGroups returns 403/Forbidden → mark the retention signal UNKNOWN (do not conclude "no retention policy defined"); a successful read with no retentionInDays means retention is genuinely undefined. Do NOT route an all-5-types-on + forbidden-retention case to UNKNOWN; per the 403 retention-signal rule in the Rating block, cap it at AMBER-with-note (logging confirmed enabled, retention unverifiable).

Rating:

  • 🟢 GREEN: All 5 log types enabled with defined retention policy — AND both the log-type read (eks:DescribeCluster) and the retention read (logs:DescribeLogGroups) SUCCEEDED (an unverifiable retention signal cannot award GREEN — see 403 retention-signal rule)
  • 🟡 AMBER: Audit enabled but not all 5 log types on, or all 5 types on but retention confirmed undefined (successful log-group read, no retentionInDays), or all 5 types on but the retention read returned 403 (logging confirmed enabled, retention unverifiable — see 403 retention-signal rule)
  • 🔴 RED: Control plane logging completely disabled, or audit logs specifically disabled (confirmed via a successful eks:DescribeCluster — this RED survives a 403 on the retention read; see 403 retention-signal rule)
  • ⬜ UNKNOWN: eks:DescribeCluster forbidden AND that forbidden read was the sole discriminator (log-type state unknowable and nothing else yields a color)
  • 403 retention-signal rule: a confirmed RED survives a 403 on a different read — a "logging disabled / audit disabled" RED observed via a successful eks:DescribeCluster stays RED even if the retention logs:DescribeLogGroups read returned 403 (CONFIRMED FLOOR); a 403 on a different signal never downgrades it to UNKNOWN. GREEN requires the retention policy confirmed via a successful log-group read, so an all-5-types-on + retention-403 case must NOT award a clean GREEN and must NOT route to UNKNOWN — cap it at AMBER with the note "logging enabled but retention could not be verified". Whole-check UNKNOWN only when eks:DescribeCluster was forbidden AND that was the sole discriminator (no successfully-read signal yields a color).
  • Evaluation order: assess RED first (on successfully-read signals); if not RED, assess AMBER; otherwise GREEN. Keeps the bands exhaustive and non-overlapping.

Key talking point: EKS control plane logging is OFF by default. The audit log is your security camera for every API call.

4.2 — Metrics Collection & Dashboards​

What to check:

  • CloudWatch Container Insights add-on (amazon-cloudwatch-observability)
  • Prometheus pods (labels: app.kubernetes.io/name=prometheus or app=prometheus)
  • Grafana pods
  • kube-state-metrics deployment (critical for cluster state visibility)
  • node-exporter DaemonSet
  • ADOT add-on
  • Third-party monitoring DaemonSets (Datadog, New Relic, Dynatrace)

How to check:

  1. Describe addon amazon-cloudwatch-observability. If eks:DescribeAddon/eks:ListAddons returns 403/Forbidden → mark the Container-Insights signal UNKNOWN (do not conclude the add-on is absent).

  2. List pods with label app.kubernetes.io/name=prometheus across all namespaces

  3. List pods with label app.kubernetes.io/name=grafana

  4. List pods with label app.kubernetes.io/name=kube-state-metrics. For steps 2-4: Pod (core v1) is a core API so 404 is not expected; a 403/Forbidden on any of these pod lists → mark that signal UNKNOWN (do not conclude the component is absent); an empty successful list means none exist.

  5. List DaemonSets across all namespaces (catches node-exporter and third-party agents). If 403/Forbidden → mark the metrics-collection signal UNKNOWN (do not conclude absent); DaemonSet (apps/v1) and Pod (core v1) are core APIs so 404 is not expected; an empty successful list means none exist. Do not rate RED for 'no metrics collection' on a forbidden list — route to UNKNOWN.

    403 → RED guard (all reads): the RED "no metrics collection at all" band is reachable only when every read above (steps 1-5) succeeded and returned nothing. If ANY of these reads was forbidden (403), RED is unreachable — route to UNKNOWN.

Rating:

  • 🟢 GREEN: Metrics collection + kube-state-metrics + dashboards (Container Insights or Prometheus+Grafana or third-party) — AND every read (steps 1-5) that establishes a GREEN precondition SUCCEEDED (an unconfirmable good signal caps at AMBER-with-note, never GREEN — see 403 floor-and-cap rule)
  • 🟡 AMBER: Partial stack (e.g., Container Insights but no kube-state-metrics, or Prometheus without Grafana), or a GREEN-worthy stack where one contributing read returned 403 (a confirmed good signal present but a different precondition unverifiable — see 403 floor-and-cap rule)
  • 🔴 RED: No metrics collection at all — reachable ONLY when all reads (steps 1-5) succeeded and each returned nothing; if any read was forbidden, this RED is unreachable (route to UNKNOWN only per the floor-and-cap rule)
  • ⬜ UNKNOWN: a metrics-source read (steps 1-5) was forbidden AND that forbidden read was the sole discriminator, so no successfully-read signal yields a color — do NOT rate RED, and do NOT downgrade a signal that a successful read already colored
  • 403 floor-and-cap rule: a confirmed AMBER survives a 403 on a different read — e.g. Container Insights confirmed present (successful step 1) + kube-state-metrics confirmed absent (successful step 4) is a CONFIRMED AMBER and STAYS AMBER even if the DaemonSet list (step 5) returned 403; the 403 only marks its own signal (node-exporter/third-party agents) UNKNOWN and never downgrades the confirmed AMBER (CONFIRMED FLOOR). Still evaluate successfully-read signals RED-first then AMBER. GREEN requires all preconditions confirmed by successful reads, so a 403 on a GREEN-contributing read caps at AMBER-with-note ("...could not verify X") when other signals are GREEN-worthy — never GREEN. Whole-check UNKNOWN only when the 403 was the sole discriminator (no successfully-read signal yields a color).
  • Evaluation order: assess RED first (on successfully-read signals); if not RED, assess AMBER; otherwise GREEN. Keeps the bands exhaustive and non-overlapping.

Investigate manually: Whether dashboards are actively used (viewership) is not observable via API — a present dashboard resource does not prove it is watched; suggest the user confirm.

4.3 — Centralized Log Aggregation for Workloads​

What to check:

  • Fluent Bit DaemonSet (labels: app.kubernetes.io/name=fluent-bit or k8s-app=fluent-bit)
  • Fluentd DaemonSet
  • CloudWatch agent DaemonSet in amazon-cloudwatch namespace
  • Application log groups in CloudWatch

How to check:

  1. List DaemonSets with Fluent Bit labels across all namespaces. If 403/Forbidden → mark the log-collection signal UNKNOWN (do not conclude absent); DaemonSet (apps/v1) and Pod (core v1) are core APIs so 404 is not expected; an empty successful list means none exist. Do not rate RED for 'no centralized log collection' on a forbidden list — route to UNKNOWN.
  2. List DaemonSets in amazon-cloudwatch namespace
  3. Use CloudWatch tools to check for application/workload log groups under /aws/containerinsights/{cluster-name}/… (specifically the application, host, and dataplane groups). Note: /aws/eks/{cluster-name}/cluster is the CONTROL-PLANE log group, not workload logs — do not use it to assess application logging. If logs:DescribeLogGroups returns 403/Forbidden → mark the retention signal UNKNOWN (do not conclude "no retention policy"); a successful read with no retentionInDays means retention is genuinely undefined. Do NOT route a shipper-present + forbidden-retention case to UNKNOWN; per the 403 retention-signal rule in the Rating block, cap it at AMBER-with-note (shipper confirmed present, retention unverifiable).

Rating:

  • 🟢 GREEN: Log shipper deployed, logs centralized with retention policy — AND both the shipper DaemonSet read (step 1) and the retention read (step 3) SUCCEEDED (an unverifiable retention signal cannot award GREEN — see 403 retention-signal rule)
  • 🟡 AMBER: Log shipper exists but retention confirmed absent (successful log-group read, no retentionInDays), or log shipper confirmed present but the retention read returned 403 (shipper confirmed present, retention unverifiable — see 403 retention-signal rule)
  • 🔴 RED: No centralized log collection — teams rely on kubectl logs (confirmed via a successful DaemonSet list returning no shipper — this RED survives a 403 on the retention read; see 403 retention-signal rule)
  • ⬜ UNKNOWN: the DaemonSet list (step 1) returned 403/Forbidden AND that forbidden read was the sole discriminator (shipper presence unknowable and nothing else yields a color)
  • 403 retention-signal rule: a confirmed RED survives a 403 on a different read — a "no centralized log collection" RED observed via a successful DaemonSet list (step 1) returning no shipper stays RED even if the retention logs:DescribeLogGroups read (step 3) returned 403 (CONFIRMED FLOOR); a 403 on a different signal never downgrades it to UNKNOWN. GREEN requires the retention policy confirmed via a successful log-group read, so a shipper-present + retention-403 case must NOT award a clean GREEN and must NOT route to UNKNOWN — cap it at AMBER with the note "log shipping enabled but retention could not be verified". Whole-check UNKNOWN only when the DaemonSet list was forbidden AND that was the sole discriminator (no successfully-read signal yields a color).
  • Evaluation order: assess RED first (on successfully-read signals); if not RED, assess AMBER; otherwise GREEN. Keeps the bands exhaustive and non-overlapping.

Investigate manually: Log format (structured vs unstructured logging) — not observable via API; requires sampling actual log records.

4.4 — Alerting Defined and Actionable​

What to check:

  • CloudWatch Alarms related to EKS/ContainerInsights
  • Prometheus Alertmanager pods
  • PrometheusRule resources (alert definitions)

How to check:

  1. List pods with label app.kubernetes.io/name=alertmanager. If 403/Forbidden → mark the alerting signal UNKNOWN (do not conclude absent); Pod (core v1) is a core API so 404 is not expected; an empty successful list means none exist. Do not rate RED for 'no alerting configured' on a forbidden list — route to UNKNOWN. Note alerting can also be satisfied by CloudWatch alarms (step 3): a 403 on this k8s alertmanager list must not RED if CloudWatch alarms are present or the alarm read itself failed.
  2. List PrometheusRule resources. If 404/NotFound (CRD not installed) → Prometheus Operator not deployed, rate alerting based on CloudWatch only. If 403/Forbidden → mark UNKNOWN.
  3. Use CloudWatch tools to list alarms with ContainerInsights namespace. If cloudwatch:DescribeAlarms returns 403/Forbidden → mark the CloudWatch-alerting signal UNKNOWN (do not conclude "no CloudWatch alarms"); a successful read returning zero alarms means none exist. Do not rate RED "no alerting configured" when this alarm read was forbidden — route to UNKNOWN, since CloudWatch alarms may exist and be unreadable.

Rating:

  • 🟢 GREEN: Alerts cover critical scenarios (node, pod, capacity) — AND every alerting-source read (steps 1-3) that establishes a GREEN precondition SUCCEEDED (an unconfirmable good signal caps at AMBER-with-note, never GREEN — see 403 floor-and-cap rule)
  • 🟡 AMBER: Some alerts exist but incomplete coverage, or a GREEN-worthy alerting posture where one contributing read returned 403 (a confirmed good signal present but a different precondition unverifiable — see 403 floor-and-cap rule)
  • 🔴 RED: No alerting configured — reachable ONLY when all three reads succeeded and each was conclusively empty: alertmanager pod list succeeded and returned zero pods, PrometheusRule returned 404 (CRD absent) or succeeded with zero rules, AND cloudwatch:DescribeAlarms succeeded and returned zero alarms. If any of these reads was forbidden (403), this RED is unreachable (route to UNKNOWN only per the floor-and-cap rule).
  • ⬜ UNKNOWN: an alerting-source read was forbidden — alertmanager list 403 (step 1), PrometheusRule 403 (step 2), or cloudwatch:DescribeAlarms 403 (step 3) — AND that forbidden read was the sole discriminator, so no successfully-read signal yields a color; do NOT rate RED, and do NOT downgrade a signal that a successful read already colored
  • 403 floor-and-cap rule: a confirmed AMBER survives a 403 on a different read — e.g. CloudWatch alarms confirmed present with incomplete coverage (successful step 3) is a CONFIRMED AMBER and STAYS AMBER even if the PrometheusRule read (step 2) returned 403; the 403 only marks its own signal (Prometheus alert rules) UNKNOWN and never downgrades the confirmed AMBER (CONFIRMED FLOOR). Still evaluate successfully-read signals RED-first then AMBER. GREEN requires all preconditions confirmed by successful reads, so a 403 on a GREEN-contributing read caps at AMBER-with-note ("...could not verify X") when other signals are GREEN-worthy — never GREEN. Whole-check UNKNOWN only when the 403 was the sole discriminator (no successfully-read signal yields a color).
  • Evaluation order: assess RED first (on successfully-read signals); if not RED, assess AMBER; otherwise GREEN. Keeps the bands exhaustive and non-overlapping.

Investigate manually: Runbooks linked to alerts, and whether on-call actually monitors the alerts — neither is observable via cloudwatch:DescribeAlarms or cluster state; requires reviewing alert/notification configuration and on-call process directly.

Minimum viable alert set: NodeNotReady, PodCrashLooping, PodPendingTooLong, HighAPIServerLatency, IPExhaustion.