observability-k8s-investigation

Kubernetes Investigation

Diagnose Kubernetes issues using OTel telemetry collected via EDOT (Elastic Distribution of OpenTelemetry) and the kube-stack collector. Correlate cluster state, pod runtime metrics, K8s events, application logs, and APM to identify root cause across the workload, node, and control-plane layers.

Environment Configuration

This skill executes Elasticsearch operations through the elastic CLI. If the elastic CLI is not installed, tell the user what it is needed for. Do not guess credentials, call the HTTP API directly, or attempt other workarounds.

This skill references operations in HTTP-shorthand form (e.g., GET /, GET /_cat/indices, GET /{index}/_mapping, GET /{index}/_settings/index.mode, POST /_query). The Operations table at the end of this document maps each shorthand to the equivalent elastic CLI command — always use the CLI rather than calling the HTTP API directly.

Analysis without cluster access

The CLI check above gates querying the cluster — it does not gate analysis. When the user has already supplied the evidence in their question (metric values, counts, status reasons, log lines, alert payloads, configuration), reason from that evidence and deliver the conclusion.

When you genuinely do need data the user has not provided, still say what you would check and how — name the specific query, index, and field that would settle the question — and then ask for CLI setup. An answer that names the check is useful without a cluster; one that only asks for setup is not.

Every ES|QL query in this skill and in references/query-recipes.md runs via POST /_query. Alert state is read with GET kbn:/api/alerting/rules/_find. Field-presence checks use GET /<index>/_mapping or GET /_field_caps. The Operations table maps each to its elastic CLI equivalent.

Scope

In scope: OTel-receiver-namespaced indices (metrics-kubeletstatsreceiver.otel-*, metrics-k8sclusterreceiver.otel-*, logs-k8seventsreceiver.otel-*, logs-k8sobjectsreceiver.otel-*) and OTel semantic conventions (k8s.pod.name, k8s.namespace.name, k8s.container.restarts).

Out of scope:

• The legacy Elastic Agent Kubernetes integration (metrics-kubernetes.*, logs-kubernetes.*, kubernetes.* fields). Being deprecated — do not author queries against these paths.
• APM-layer analysis (service SLO breaches, transaction error rates, upstream dependency health). Different domain — once a K8s root cause is ruled in or out, hand off to the observability-sre-triage skill, which owns SLO status and burn rate, active alerting rules, throughput, latency, error rate, dependency health, and log funnelling. That is also the skill to use when the workload turns out not to be Kubernetes-hosted at all.
• Cluster provisioning, capacity planning, cost optimization. Different domain.

Guidelines

These apply to every investigation. When in doubt, re-read them before writing the synthesis.

Absence of evidence is not evidence. Do not confabulate from empty results. If log queries return 0 rows, logs are likely not collected or the pod has no recent lines — this does not mean "dependency unavailable" or any other specific failure mode. Report no_logs_available and weight remaining signals accordingly.

Empty dependency data ≠ upstream healthy. Services without APM instrumentation (load generators, workers) emit no destination metrics. Report insufficient_dependency_data, not "upstreams OK."

Co-symptoms are not causes. Two services degrading simultaneously usually share an upstream, not a causal link. Only attribute causation when (a) one service's degradation clearly precedes the other's, and (b) the delta is large (>5× error rate, >3× latency).

OOMKilled ≠ memory leak by default. The limit might simply be undersized for the workload's working set. Tell them apart by the shape of the memory curve: a monotonic climb to the limit that resets on each restart, with load flat against the prior week and no recent deploy, is the leak signature — commit to it at high confidence. Reach for a 7-day same-hour baseline when the shape is ambiguous — spiky, diurnal, or load-correlated — not as a precondition for every OOMKilled finding.

Error-termination ≠ application bug by default. Check k8s.container.cpu_limit_utilization first. CFS throttling driving liveness probe timeouts is the most common misdiagnosis in this space.

Average CPU hides throttling. A pod can look healthy at 40–60% average cpu_limit_utilization while being throttled severely at p99. Linux enforces CPU limits in 100ms periods; bursty workloads reach quota mid-period and stall. Look at max and p95, not only the average.

Restart count is boolean, not a counter. k8s.container.restarts is pulled directly from the K8s API and can be pruned by the kubelet at any time, so the absolute value is unreliable. Treat it as == 0 (no recent restarts) versus > 0 (recently restarting); do not derive backoff timing or "linear versus exponential" patterns from it. Confirm the restart pattern via K8s Killing / BackOff events instead.

Prefer to report uncertainty over manufacturing confidence. If the evidence is ambiguous, the synthesis should say so. Competing hypotheses are a valid output.

Equally, do not manufacture uncertainty. The rule above is about ambiguous evidence, not about tone. When the pivotal signal is present and corroborated, commit to it at high confidence. Hedging an unambiguous finding down to "medium" is as much a defect as overclaiming.

Deliver the synthesis and stop. State confidence once, in the HYPOTHESIS line — not again per bullet. Do not narrate which queries were run unless a result changed the conclusion, and do not restate the alert back to the reader. End on RECOMMENDED NEXT STEPS or DOWNSTREAM IMPACT; never close with an offer such as "want me to look further?". Follow-up work belongs in the recommendations list, phrased as a recommendation.

Indices and fields

Where to look

Key fields

Flat OTel paths work in ES|QL. Prefer the flat form for readability; the nested resource.attributes.* form is for raw log documents only.

Container-level against pod-level limit utilization

Read limit utilization at the container level. k8s.container.cpu_limit_utilization and k8s.container.memory_limit_utilization are the default; the pod-level pair is a different measurement, not a synonym.

The receiver emits the two families on separate documents in the same data stream: pod-level fields appear on documents that carry no k8s.container.name, and container-level fields appear only on documents that do. A STATS ... BY k8s.container.name therefore returns null for every pod-level field, and the reverse holds too. Measured over one hour on a live 9.6.0 cluster: of 42,240 documents without a container name, 360 carried k8s.pod.cpu_limit_utilization and none carried the container field; of 18,240 documents with a container name, 900 carried the container field and none carried the pod field.

Availability differs too. Container-level utilization is emitted for each container that declares the limit, while the pod-level aggregate requires every container in the pod to declare it. Across two live clusters over three hours, no pod carried the pod-level field without also carrying the container-level one, while 27 pods carried the container-level field with the pod-level field absent — every one of them a multi-container pod in which only some containers declared limits. A pod-level throttling check on a sidecar-injected pod silently returns null.

Use the pod-level fields only when the question is genuinely about the pod as a whole — total consumption against the sum of its containers' limits — and only after confirming they are populated. observability-sre-triage applies the same rule, so the two skills return the same answer for the same pod.

Field availability

Several fields above are off by default in stock kube-stack collectors and require explicit configuration. Verify presence with GET /<index>/_mapping or GET /_field_caps before relying on them; if absent, fall back as noted and call out the substitution in the synthesis.

If a fall-back is used, note it in the synthesis (for example, (via memory.usage; limit_utilization not collected)) so the reader knows the signal is indirect.

ES|QL gotchas

Before writing queries, know these. Each of them silently produces wrong answers rather than failing loudly.

VALUES() returns scalar for single distinct value, array for multiple. Templating that assumes array shape (for example, | first) extracts the first character of the string when scalar. Use MV_FIRST(VALUES(...)) or handle both.

VALUES() is newer than this skill's base floor. It is GA on Serverless, but on Stack it is preview from 8.14.0 and GA only in 9.4.0, and it does not exist at all below 8.14. Check GET / before using it: build_flavor: "serverless" means it is available, otherwise read version.number. Where it is not available, move the field into the BY clause instead of aggregating it — one row per distinct value carries the same information:

| STATS restarts = MAX(k8s.container.restarts), phase = MAX(k8s.pod.phase)
    BY term_reason = k8s.container.status.last_terminated_reason
|

PERCENTILE does not work on OTel histogram type (as of 8.15). For APM duration percentiles, use AVG on the aggregate_metric_double summary field (AVG(transaction.duration.summary) divides sum by value_count). For true percentiles, fall back to Kibana Query DSL.

COUNT(agg_metric_double) returns value_count (events), not doc count. SUM(field) gives the sum component; AVG(field) gives sum/value_count. Do not use SUM(transaction.duration.summary) as an event-count proxy — it returns total duration.

K8s metrics use flat OTel field paths in ES|QL. k8s.pod.name, not resource.attributes.k8s.pod.name. The nested form is for raw log documents.

Failure-mode taxonomy

The classification vocabulary — pivotal signal and corroborating checks for each mode across the workload, node, control plane, autoscaling and networking layers, plus what to do when two modes fit — lives in references/failure-modes.md. Read it before classifying.

Signal interpretation

Memory

• Monotonic rise over 30–60 min → leak. Check GC metrics for the language: JVM jvm.gc.duration, Go process.runtime.go.gc.pause_ns, Node v8js_gc_duration. Rising GC frequency/pause with stable live-set is the canonical leak signature.
• Diurnal / load-correlated spikes → load-driven, not leak. Consider HPA tuning or limit increase.
• Hits 1.0, then restart → OOMKilled confirmed. Exit code 137 (SIGKILL) in app logs consistent.

CPU

• cpu_limit_utilization > 1.0 sustained → CFS throttling. Node has spare CPU; the pod is quota-blocked.
• Symptoms of throttling (not the throttle metric itself): liveness probe timeouts, p99 latency 4–16× p50, queue backpressure upstream, Error-reason container terminations.
• Average can look healthy while p95 is throttled. Do not trust average alone.

Restart patterns

• restarts > 0 recently → workload has been restarting. Don't read magnitude into the count (see Restart count is boolean); confirm the pattern from K8s Killing / BackOff event timestamps in logs-k8seventsreceiver.otel-*.
• Restarts correlated with memory pressure (memory_limit_utilization → 1.0) → OOMKilled path.
• Restarts without memory/CPU pressure → probe misconfig, app bug, or startup dependency failure. Pull events for Unhealthy and Killing.

Termination reasons

• OOMKilled → memory path.
• Error → non-zero exit. Check app logs; if empty/minimal, check CPU throttling before attributing to app logic.
• Completed → ran to completion. Normal for Jobs/CronJobs/init containers; anomalous otherwise.
• ContainerCannotRun → runtime/image/exec issue. Check image pull events.

Investigation flow

An investigation is not a checklist. The sections below describe a typical arc — compress, skip, or revisit them based on what you find. Terminate as soon as you have enough evidence to synthesize at a known confidence. Chasing signals past the point of diminishing returns is a failure mode, not thoroughness.

Orient

Resolve the target: k8s.pod.name, k8s.namespace.name, optionally k8s.deployment.name and service.name. If no time window is given, default to the last hour for pod-level investigations, last 2 hours for event correlation, last 6 hours for ongoing/unresolved incidents.

If the alert payload already tells you the failure mode (for example, it fires specifically on OOMKilled), note that and skip classification; move to confirmation and baseline comparison.

Characterize

Get the shape of the workload's recent behavior: restart count, termination reasons, phase, utilization. One or two queries usually suffice.

VALUES() needs Serverless or Stack 8.14+ (GA 9.4). On an older Stack cluster use the BY-clause rewrite in ES|QL gotchas rather than dropping the termination reason from the query.

Group by container: a sidecar-injected pod has several, and only the ones that declare limits report utilization. If every column comes back null, no limits are declared — fall back as described under Field availability rather than reading null as idle.

Classify

Use the taxonomy in references/failure-modes.md. The pivotal signal should match; the "Investigate" column tells you what corroboration to seek.

When two modes fit, note both and proceed with the one that has the stronger pivotal signal. You can revise during corroboration.

Corroborate

Pull the evidence your classification predicts you'll find. Typical sources:

K8s events for the namespace and window:

Namespace is a log attribute on this receiver, not a resource attribute. Filter attributes.k8s.namespace.name, not the flat k8s.namespace.name. The flat form is mapped on this data stream, so a query using it parses and executes and returns zero rows with no error — on a Stack 9.4.4 cluster the flat field was populated on 0 of 832 events while the attributes. form carried the namespace on all of them, so the flat filter discarded 300 matching BackOff, Killing and Unhealthy events. The flat k8s.* paths do work on metrics-kubeletstatsreceiver.otel-* and metrics-k8sclusterreceiver.otel-*, where namespace is a resource attribute; the events receiver is the exception. Confirm with COUNT(attributes.k8s.namespace.name) against COUNT(*) before trusting an empty event result.

Application logs if available — look at the 200 most recent lines before the termination timestamp. If absent, flag no_logs_available; do not invent a log pattern.

APM if the pod runs an instrumented service — resolve service.name from pod resource attributes for later correlation. SLO / latency / error-rate analysis itself is APM-layer work and out of scope for this skill.

Baseline comparison — for utilization-based findings, compare current values to 7-day-prior at the same hour-of-day. "High memory" is meaningful only relative to what's normal for this workload.

Check for upstream cause (conditional)

Only pursue if the symptom pattern suggests it. Threshold: upstream error rate >5× baseline or latency >3× baseline, AND degradation started before the symptom on the target service. Co-symptoms do not establish causation.

If metrics-service_destination.1m.otel-default has no rows for the service, report insufficient_dependency_data — not "upstreams healthy."

Check for recent change (conditional)

SuccessfulCreate / Pulled events in the last 2 hours often correlate with deploys. logs-k8sobjectsreceiver.otel-* shows configmap/secret/deployment spec changes. A change within 15 minutes of the symptom onset is a strong correlation, but still a correlation — verify it plausibly explains the mode you've classified.

Synthesize and stop

Synthesize as soon as you have enough evidence to support a hypothesis at known confidence. You do not need to complete every preceding section — investigation terminates when either:

• You have a high-confidence hypothesis with corroboration, or
• You have a low/medium-confidence hypothesis and further queries are unlikely to change the picture (for example, logs are unavailable, APM isn't instrumented, no recent changes found).

Synthesis

Default structure:

Scale. The whole synthesis runs 250–400 words. HYPOTHESIS is two or three sentences. EVIDENCE is three to five single-line bullets, each citing a concrete value rather than re-explaining it. RECOMMENDED NEXT STEPS is two or three single-line items — the ones you would actually do first, not everything that could be done. DOWNSTREAM IMPACT is one or two sentences. A well-evidenced alert fits inside this comfortably; length is not thoroughness, and the on-call reader scanning mid-incident will not get past the first screen.

When two hypotheses are live: replace HYPOTHESIS with COMPETING HYPOTHESES; list both, say which you lean toward and why, and list the evidence that would disambiguate them.

When no incident is found (symptom resolved, or alert appears spurious): say so directly. ALERT FIRED BUT SYSTEM APPEARS HEALTHY is a valid output. List what you checked and what you didn't find.

Confidence calibration

Start at high and downgrade based on what's missing:

• Downgrade to medium if: primary signal is clear but corroboration is missing (no logs, no APM, no baseline comparison possible). Or: two modes fit and you can't disambiguate.
• Downgrade to low if: only a single signal supports the hypothesis, signals conflict, or the mode requires evidence you couldn't fetch.

Never return high when application log data was absent and the hypothesis depends on application behavior. Absence of evidence does not corroborate a hypothesis.

Query recipes

Ready-made queries for the most-restarting-pods, CPU-throttling, node-memory-pressure, admission-denial, and firing-alert paths live in references/query-recipes.md.

Examples

"Why is my pod CrashLoopBackOff-ing?"

Characterize first: get restart count, termination reason, memory and CPU utilization.

• If last_terminated_reason == "OOMKilled" and memory utilization reached 1.0 → memory path. Corroborate with 7-day baseline: monotonic rise over days = leak; spiky = load-driven. Check GC metrics if language is known.
• If last_terminated_reason == "Error" and cpu_limit_utilization > 1.0 → CPU throttling path. Corroborate with liveness probe config (initialDelaySeconds, timeoutSeconds) and K8s events for Unhealthy.
• If last_terminated_reason == "Error" and CPU is fine → application-logic path. Pull recent logs before termination.
• If last_terminated_reason == "ContainerCannotRun" → image/exec path. Check K8s events for Failed pull events.

Synthesize with appropriate confidence. If logs were unavailable on the Error path, downgrade to medium and say so.

"Is my rollout stuck?"

Authoritative signal: k8s.deployment.available < k8s.deployment.desired for > 10 minutes.

Diagnose the constraint:

• K8s events on the new ReplicaSet: FailedCreate → admission rejection (quota, webhook, PSP). FailedScheduling → no node fits.
• New-pod utilization: all at 0% memory → never started (image pull failure); high CPU with low memory → slow startup hitting readiness probe.
• HPA state: stable current_replicas < desired_replicas under load → unready-pod dampening.

"Alert fired but everything looks healthy"

Possible and worth naming explicitly. Check:

• Has the symptom resolved? Compare current utilization/restart rate to the alert trigger point.
• Was the alert a transient spike that's already decayed?
• Is the alert tuned appropriately (for example, a too-short evaluation window)?

Output: ALERT FIRED BUT SYSTEM APPEARS HEALTHY with what you checked. Recommend alert tuning if the pattern is recurrent.

Related

• Workflow: K8s CrashLoopBackOff Investigation — alert-triggered automated version of the pod-level path above. Runs deterministic ESQL + branches; this skill provides the interpretation layer the workflow lacks.
• Forge genome library: 16 K8s failure scenarios (OOMKill cascade, CPU throttling, probe misconfig, node NotReady, admission webhook block, and so on) validating this skill's coverage.

Operations