observability-sre-triage

SRE Service Triage

Decide whether a service is healthy, degraded, or unhealthy, and say why. Triage is a hierarchy, not a checklist: SLOs and alerts define whether the service is failing its contract, trace-derived golden signals describe how it is failing, dependencies and infrastructure explain where the failure comes from, and logs supply the sentence you put in the incident channel. Work down the hierarchy until the evidence supports a verdict, then stop.

For authoring and tuning SLO definitions, burn-rate rules, and alert thresholds, use the observability-service-reliability skill. This skill only reads that state. For Kubernetes workload, node, or control-plane diagnosis — restart loops, OOM kill confirmation, node pressure, admission rejections, stuck rollouts — hand off to the observability-k8s-investigation skill. This skill checks whether a Kubernetes-hosted service is saturated; it does not diagnose why the pod or the node behind it is failing.

<!-- begin-partial: preamble -->

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.

<!-- end-partial: preamble -->

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.

Everything here is expressed in ES|QL (POST /_query) or the Kibana Observability APIs. Do not use Query DSL, and do not use the ES|QL KQL search function — express predicates natively (WHERE service.name == "checkout").

Jobs to be done

• Answer "is service X healthy?" with a verdict and the evidence behind it
• Answer "why is service X slow / erroring / quiet?" by localizing the change to the service, a dependency, or its

infrastructure

• Read SLO status, burn rate, and remaining error budget during an incident
• Determine which alerting rules currently apply to a service, including all-services rules
• Funnel a noisy log stream down to the failures that explain the degradation
• Identify which attribute (version, host, pod, region, route) distinguishes the failing or slow subpopulation
• Distinguish a healthy service from a service with no telemetry

Output discipline

Applies to every response produced under this skill.

• Commit to the best-supported conclusion. When the evidence points one way, say so. Do not downgrade confidence to

sound cautious — hedging on unambiguous evidence is a defect, not humility.

• Commit to a verdict: healthy, degraded, or unhealthy, followed by the reason. A triage answer that does not name

one of the three has not done the job.

• State confidence once, in the conclusion. Do not restate it per bullet.
• Do not speculate past the evidence. If the telemetry did not show a cause, it does not go in the answer. Name what

is unknown and stop. Never offer a mechanism ("probably a GC pause", "likely a noisy neighbor") that no signal measured.

• Report absence as absence. Zero rows means the data is missing or not collected; it never means the underlying

condition is healthy. "No dependency metrics" is not "dependencies are fine".

• Do not pad. No restating the question, no narrating which queries were run unless the result mattered, no

summarizing the summary.

• End on the finding. No trailing offers such as "want me to dig deeper?". Actionable follow-ups belong in a

recommendations list, phrased as recommendations, not as questions.

Signal hierarchy

Signals disagree constantly. This ordering decides which one wins.

Conflict rules:

• SLO healthy, latency elevated → degraded but within error budget. The verdict follows the SLO; report the trend as

a risk with the burn rate.

• SLO violated, current-window metrics look fine → trust the SLO and check its window. SLOs are evaluated over hours

or days; a 15-minute ES|QL window can look clean while the budget is already spent.

• Alerts firing, no SLO defined → the alerts are the verdict. Resolve each rule's params to confirm it actually

targets this service before attributing it.

• Logs noisy, golden signals flat → not degraded. High log volume without an error-rate or latency change is a

logging-configuration finding, not a health finding.

• Throughput collapsed, error rate flat → the caller stopped calling. Look upstream before blaming this service.
• Any query returns zero rows → missing data. Say which signal is unavailable and lower the scope of the verdict

accordingly; never convert silence into health.

Routing: symptom to first signal

Data sources

OTel-native data streams, verified against Elasticsearch 9.6.0:

service.name is populated on traces, metrics, and logs, so it is the join key across all three. Use flat OTel field paths in ES|QL (k8s.pod.name, not resource.attributes.k8s.pod.name). When analyzing OTel application metrics, the ES|QL TS (time series) command gives more efficient metric queries. It is GA on Serverless; on Stack it is preview in 9.2 and GA in 9.4, so below 9.4 use FROM with BUCKET instead. TS also rejects COUNT(*) — count a field instead.

The recipes in this skill and its references are written against the OTel-native streams above. A service instrumented with the classic Elastic APM agent ships to traces-apm* and metrics-apm* under different field names (transaction.duration.us, event.outcome), so these recipes return no rows for it. An empty result on a service that is otherwise clearly alive is therefore a scope boundary, not evidence of an outage: check which index family the service actually writes (GET /_cat/indices) and report the ingest path rather than concluding from silence.

ES|QL feature availability

Three features this skill uses are newer than its 8.11 base floor. Check GET / before relying on them: build_flavor: "serverless" means all three are available; otherwise compare version.number against the Stack column. Never report "no data" when the real answer is that the query did not run — say which feature was unavailable and use the fallback.

The Platinum requirement on CATEGORIZE is not a version check. A 9.6 Stack cluster on a Basic or Gold licence fails it exactly as an 8.11 cluster fails FORK, and the error names the licence rather than the syntax. On Serverless the function is GA with no separate licence gate.

Process: triage a degraded service

1. Fix the service and the window. Resolve the service name and the time range from the request. Use the user's time

range — do not silently assume the last hour when the complaint is historical. If no range is given, use the last hour and say so. Confirm the service actually exists in telemetry with a COUNT(*) BY service.name over traces-*.otel-* via POST /_query; if the name does not appear, resolve the ambiguity before querying further.

Decision: which service and window every later query is scoped to. Data: distinct service.name values in range.

1. Read SLO status and burn rate. List SLOs with GET kbn:/api/observability/slos and fetch the ones bound to this

service with GET kbn:/api/observability/slos/{id}. Read status, current SLI, burn rate, and remaining error budget.

Decision: does an agreed contract exist, and is it being violated? If yes, the verdict is already determined and the remaining steps only explain it. If no SLO covers this service, say so once and fall through to step 3.

1. Determine which alerting rules apply to this service, and which of them are firing. Call

GET kbn:/api/alerting/rules/_find with per_page=100&filter=alert.attributes.enabled:true, paging with page if total exceeds what you received. Then filter the response client-side. *Do not query `.alerts indices to determine active state** — the Alerting API response is the source of truth. Fetch a rule's full definition with GET kbn:/api/alerting/rule/{id} when its params` are needed.

Do not narrow this call server-side. The _find filter parameter is KQL over saved-object attributes, and params is not among them — filter=alert.attributes.params.serviceName:<name> returns zero rules on a cluster that has them. Narrowing by search=apm&search_fields=tags, by alertTypeId, or by consumer is worse: it drops rules on a naming convention or a rule-type allowlist, and the rules it drops are disproportionately the all-services ones. See references/slo-and-alerts.md for the measured failure.

From the fetched set, evaluate both rules whose params.serviceName matches the service and rules where params.serviceName is absent, because the latter are all-services rules that apply to it too. Read execution_status.status on each: active means the rule's last run produced alerts, ok means it ran and produced none, and error means it is not evaluating at all — a blind spot, not a pass.

Decision: what covers this service, and is any of it currently firing? Data: rule params.serviceName, rule type, and execution status.

1. Check ML anomalies, if any jobs exist. List jobs with GET /_ml/anomaly_detectors and confirm they are running

with GET /_ml/anomaly_detectors/_stats — a stopped job produces no anomalies, which is not the same as no anomaly. Pull scored records with GET /_ml/anomaly_detectors/{id}/results/records.

Decision: did latency, throughput, or error rate deviate from its learned baseline, and when? Use the anomaly window to narrow steps 5 and 6.

1. Measure the golden signals. Run ES|QL over traces-*.otel-* for throughput, latency (avg, p95, p99), and error

rate, bucketed over the window and compared against the immediately preceding window of equal length. See references/apm-signals.md.

Decision: is the service actually changed relative to itself, and in which dimension? Data: request count, latency percentiles, and failure ratio for the current and prior windows.

1. Localize: dependencies, then subpopulation, then infrastructure.

• Dependencies — aggregate metrics-service_destination.1m.otel-* by span.destination.service.resource for

call volume, average latency, and failure rate. If this query returns zero rows for the service, the service is not APM-instrumented for dependencies; report insufficient dependency data and do not claim upstreams are healthy.

• Subpopulation — when only part of the traffic is bad, compare the failure or slow rate per candidate attribute

against the overall rate to find which attribute is over-represented. See references/apm-signals.md.

• Infrastructure — read the resource attributes on the service's spans (k8s.pod.name, container.id,

host.name) first, then branch on what they contain. Pod and namespace attributes mean the service is Kubernetes-hosted: check k8s.container.cpu_limit_utilization and k8s.container.memory_limit_utilization in metrics-kubeletstatsreceiver.otel-*. A host.name with no pod attributes means the service runs on a VM or bare host, where every k8s.* field is empty: check system.cpu.utilization, system.memory.utilization, and system.cpu.load_average.1m in metrics-hostmetricsreceiver.otel-* instead. OOM kills, CPU throttling, and host saturation degrade APM health directly. See references/apm-signals.md.

• Recent change — a deploy is the most common cause of a step change. Search deploy annotations for the service

with GET kbn:/api/apm/services/{serviceName}/annotation/search over the incident window, and compare the failure or latency rate by service.version in the subpopulation query. An annotation inside the onset window is a strong correlation; confirm it plausibly explains the symptom before attributing.

Decision: is the cause inside this service, in something it calls, in one slice of its instances, under it, or in a change that landed?

When the Kubernetes branch shows saturation, restarts, or an OOM kill, the mechanism is established and the remaining diagnosis — why the pod is being killed, whether the node is under pressure, whether a rollout is stuck — belongs to the observability-k8s-investigation skill. Hand off rather than continuing here.

1. Explain with logs. Scope logs by service.name, or by trace.id when a specific failing trace is in hand, and

run the funnel until the remaining set is small enough to read. See references/log-investigation.md. Logs confirm and articulate the cause; they do not overturn steps 2 and 3.

1. State the verdict. Healthy, degraded, or unhealthy, with the reason and one statement of confidence, followed by

recommendations. Name any signal that was unavailable.

Examples

"Is checkout healthy?" — resolve the window, read its SLOs, then the active rules including all-services rules, then throughput, latency percentiles, and error rate over traces-*.otel-* against the prior window. If the availability SLO is at 99.2% against a 99.5% target with a burn rate above 1, the verdict is unhealthy on SLO violation, and the golden signals are the explanation, not the verdict.

"Why is the frontend slow?" — compare p95 and p99 for the current window against the previous window of equal length. If service-level latency rose while per-destination latency in metrics-service_destination.1m.otel-* is flat, the added time is inside the service; if one destination's average response time rose in step with it, the dependency is the cause and the frontend is a victim.

"Only some checkout requests fail" — run the subpopulation comparison: failure rate grouped by service.version, k8s.pod.name, host.name, and cloud.region alongside the overall failure rate. An attribute value whose failure rate is several times the overall rate, on a volume large enough to matter, is the correlated attribute. On live data, grouping frontend server spans by route showed a 3.8% slow rate for POST against a 0.9% overall rate — a 4x lift that localizes the problem to write paths.

"The cart service logs look bad" — run the funnel over logs-*.otel-* scoped to service.name == "cart": get trend, total, samples, and message categorization in one FORK, then add NOT ... LIKE exclusions for each dominant pattern and re-run with the full accumulated filter until fewer than 20 patterns remain. High log volume alone is not a health verdict — check the golden signals before calling the service degraded.

"Is the payment service's upstream healthy?" — query metrics-service_destination.1m.otel-* for it. Zero rows means the service does not emit dependency metrics. Report that dependency data is unavailable for this service and give the verdict from the signals that do exist; do not report the upstreams as healthy.

"An alert fired on api-gateway" — fetch the enabled rules with no server-side narrowing, then match in memory on params.serviceName == "api-gateway" and on rules with no params.serviceName, reading execution_status.status to see which are firing. Read the firing rule's threshold from GET kbn:/api/alerting/rule/{id}, then query the same metric over the same window in ES|QL to confirm the rule is describing a real change rather than a threshold that is set too tight.

Guidelines

• Work the signal hierarchy in order and stop when the evidence supports a verdict. Do not run every query in this

document on every request.

• Anchor to SLO status and burn rate when SLOs exist. When they do not, fall back to alerts, ML anomalies, throughput,

latency, error rate, dependencies, infrastructure, and logs — and say that no SLO covers the service.

• Use the Alerting API for active-alert state. Never query .alerts* indices for it. Always evaluate both

service-scoped rules and rules with no params.serviceName.

• Fetch alerting rules unnarrowed and filter client-side. _find cannot filter on params, tag search drops rules that

do not follow a naming convention, and executionStatus.status:active returns only rules that are firing right now — each of those silently hides the all-services rules the bullet above requires.

• Always use the user's time range. Compare every metric against the immediately preceding window of equal length —

absolute numbers without a baseline do not support a verdict.

• Zero rows is missing data. Say which signal is unavailable rather than treating silence as a pass.
• Scope every query by service.name and a bounded @timestamp range, and cap output with LIMIT. Prefer coarse

buckets when only a trend is needed.

• Prefer event.outcome == "failure" for failed spans; status.code == "Error" is equivalent on OTel traces but is

null on successes, so it cannot be counted directly.

• Filter server-side traffic with kind == "Server" when measuring a service's own throughput and latency, so client

spans do not double-count.

• Treat log.level and severity_text as hints, never as filters you rely on. On real OTel data most log records carry

no level at all and those that do disagree on case and vocabulary (INFO, Information, SEVERE, Normal). In particular never write `log.level == "error"` — the lowercase ECS vocabulary is not what the OTel SDKs emit, so it returns zero rows with no error even on a service that is logging errors, and reports the service healthy. Use the normalized numeric severity_number >= 17 if you need a severity predicate at all.

• Logs explain; they do not decide. Never issue a verdict whose only support is log content.
• Do not invent field names. If a field might not exist in this deployment, confirm the data stream exists with

GET /_resolve/index/{pattern} before building on it.

• Establish where the service runs before checking saturation. Kubernetes and host telemetry share no field names, so a

Kubernetes query against a VM-hosted service returns zero rows and says nothing about whether it is saturated.

• Pass --drop-null-columns on POST /_query when a result is mostly empty columns. Infrastructure metrics are sparse

by nature — limit utilization is absent wherever no limit is declared — and the flag collapses the noise while listing the suppressed column names under all_columns, so nothing is hidden.

Operations

The SLO find command takes a KQL query string because that is the API's contract; it is not an exception to the ES|QL rule for data queries.

The annotation search route rejects a request that omits environment, so pass ENVIRONMENT_ALL when the service's environment is not known. Only the search direction is in scope: this skill is read-only, so the companion create-annotation operation is deliberately not bound.