Imported from coralogix/cx-skills (skills/opentelemetry/opentelemetry-collector/SKILL.md). Install upstream with npx skills add coralogix/cx-skills --skill opentelemetry-collector. Copyright stays with the author (Apache-2.0).

OpenTelemetry Collector

The Coralogix-flavored OpenTelemetry Collector — the coralogix exporter, the otel-integration Helm chart for Kubernetes, CDOT for ECS, and the universal installer for standalone hosts. Load this skill when a user is deploying, configuring, or debugging a collector that ships to Coralogix. Most failures come down to a handful of Coralogix-specific defaults that vanilla OpenTelemetry docs don't cover.

When to Use This Skill

High-Signal Answer Rules

For these recurring cases, include the exact corrective detail in the final answer. These are also the authoritative statements of Coralogix-specific defaults — do not contradict them elsewhere in the answer.

Exporter and routing

• domain: is a bare hostname, not a URL. eu2.coralogix.com — not https://ingress.eu2.coralogix.com, not a UI hostname.
• Bracket env vars. ${env:CORALOGIX_PRIVATE_KEY}, not $CORALOGIX_PRIVATE_KEY — unbracketed form silently fails in v0.76+. Minimum exporter block: domain: "<region>.coralogix.com" and private_key: "${env:CORALOGIX_PRIVATE_KEY}".
• Use a dedicated coralogix/resource_catalog exporter for Infrastructure Explorer with the x-coralogix-ingress: metadata-as-otlp-logs/v1 header. The default coralogix exporter won't light up the entity views.
• resourcedetection/resource_catalog crash on daemonset — error can't get K8s Instance Metadata; node name is empty means this processor is on a daemonset agent. It belongs on the opentelemetry-cluster-collector Deployment only. Fix: remove it from the daemonset config. Do not conflate with the coralogix/resource_catalog exporter, which is a separate component.
• No data + transform/OTTL: clearly say to stop trying OTTL; check receiver/exporter connectivity, DNS/TLS/proxy/egress, private key, and region/domain first.

Component stability

• Check stability before production recommendations. Collector components have per-signal stability in their upstream README. alpha is for limited non-critical use, beta is broader but can still break, stable is the production default, and deprecated means avoid new deployments and plan migration. A config that validates is not enough evidence that the component is safe for production.
• When recommending alpha/beta components, qualify the risk. Pin collector versions, test the exact config on staging traffic, read release notes before upgrades, and offer a stable fallback if the workload is business-critical.

Pipeline placement (Kubernetes)

• memory_limiter first, batch last.
• One role owns full k8sattributes extraction — typically gateway; agents use passthrough: true.
• spanmetrics on agent (before sampling), tail_sampling on gateway. Run transactions/groupbytrace/transactions before spanmetrics; never on both agent and gateway simultaneously — this causes double-counting because each tier sees all spans and emits separate metric series that accumulate. tail_sampling on a daemonset agent causes incomplete traces because each agent only sees spans from its own node — a single trace is split across agents and the sampler decides on partial data. Fix: move tail_sampling to a central gateway tier and add a loadbalancing exporter on the agents that routes spans to gateway by trace_id, so all spans for a trace reach the same gateway replica.
• Span Metrics DB label compatibility transforms belong under top-level spanMetrics.transformStatements. Do not put them only under spanMetrics.dbMetrics.transformStatements; that can populate db_calls_total while leaving normal calls_total with blank db_namespace.
• Don't replace service.pipelines wholesale — use extraProcessors/extraReceivers hooks; wholesale overrides silently break resource/metadata (cx.agent.type) and chart upgrades.

Cardinality and redaction

• Rule of 100 for metric dimensions. If an attribute can have more than about 100 unique values, do not add it as a Span Metrics dimension. Keep IDs, raw URLs, emails, request IDs, session IDs, and trace_id on spans/logs instead, or normalize them before metrics are generated.
• url.full, k8s.pod.name, and k8s.pod.ip are dangerous Span Metrics dimensions. Prefer http.route, low-cardinality host/operation labels, and stable service/resource labels. If a customer insists on url.full, sanitize it before spanmetrics consumes the span and make the risk explicit.
• aggregation_cardinality_limit is a guardrail, not a fix. For Helm spanMetrics.aggregationCardinalityLimit / collector aggregation_cardinality_limit, use it to collapse overflow series, but still remove or normalize high-cardinality labels.
• Redact before export and before metrics derivation. PII/secrets should be removed, masked, or hashed in the collector pipeline before spanmetrics, batch, and coralogix. For broad URL-like span names or URL attributes, explicitly recommend redactionprocessor with the literal keys url_sanitizer and sanitize_span_name, and include allow_all_keys: true unless intentionally using an explicit allowed_keys whitelist; otherwise unspecified attributes are dropped. Warn that broad sanitizers can over-sanitize and validate before/after examples. Use the opentelemetry-ottl skill for targeted transforms such as SHA256, replace_pattern, replace_all_patterns, delete_key, and nil-safe guards.

Platform-specific rules

• GKE Autopilot Warden: Set logsCollection.storeCheckpoints: false; disable coralogix-ebpf-profiler, hostMetrics, hostEntityEvents, resourceDetection on agent; disable resourceDetection on cluster-collector. Use gke-autopilot-values.yaml.
• ECS EC2 daemonset localhost: Apps must target the EC2 host IP (not localhost); daemonset needs networkMode: host. Remove ecs from resourcedetection.detectors — it stamps the collector's own container ID.
• ECS Fargate startup loss: Add sidecar healthCheck + dependsOn: [{containerName: otel-collector, condition: HEALTHY}] on the app. Use the CDOT image.
• Standalone installer: Recommend otel-installer/one-liner with both CORALOGIX_PRIVATE_KEY and CORALOGIX_DOMAIN.
• Infrastructure Explorer (Kubernetes): kubernetesResources and hostEntityEvents are enabled by default in the chart — do not disable them. kubernetesResources must stay on the opentelemetry-cluster-collector only (enabling it on the agent crashes with can't get K8s Instance Metadata; node name is empty). Use a dedicated coralogix/resource_catalog exporter with x-coralogix-ingress: metadata-as-otlp-logs/v1.
• Resource Catalog daemonset crash: resourcedetection/resource_catalog belongs on the opentelemetry-cluster-collector Deployment only; remove it from daemonset agents.
• Full k8sattributes: Exactly one role should do full extraction; set passthrough: true on the others.
• OpAMP supervisor endpoint: It is different from exporter domain: and needs the full URL, e.g. https://ingress.eu2.coralogix.com/opamp/v1.
• Windows extensions: [opamp] fails on old image pins: The K8s Windows sub-preset defaults to coralogixrepo/opentelemetry-collector-contrib-windows:0.92.0, which predates OpAMP on Windows — enabling extensions: [opamp] there causes the collector to refuse to start. Fix: bump the image to ≥ v0.130. When bumping the image is not an option (e.g. locked in a production freeze), use the -Supervisor wrapper instead — this runs opampsupervisor as a separate Windows Service and works regardless of collector version.
• Java multiline stack traces not merging (Kubernetes): CRI tags every log line as F (full/final) — the standard P→F recombine never triggers. Use firstEntryRegex on the filelog recombine operator to detect new entries by timestamp pattern.
• spanNameReplacePattern escaping: There are two layers: single-quote or block scalar for YAML/OTTL backslashes, and write backreferences as $$1/$$2 because the collector envprovider expands $...; verify with helm template.
• Target Allocator debugging: Port-forward svc/coralogix-opentelemetry-targetallocator on 8080; inspect /jobs and /scrape_configs; then check RBAC, selectors, and watched namespaces.

Common Workflows

1. Triage a "no data reaching Coralogix" report

Work through these steps in order before touching any pipeline configuration:

Step 1 — Prove the collector is running and exporting

# Kubernetes: check exporter metrics
kubectl exec -n <namespace> <collector-pod> -- wget -qO- http://localhost:8888/metrics \
  | grep -E 'otelcol_exporter_(sent|send_failed|enqueue_failed|queue)'
# Success: otelcol_exporter_sent_* > 0 and climbing
# Failure indicator: otelcol_exporter_send_failed_* > 0 — proceed to Step 2

Step 2 — Verify DNS and TLS reach the ingestion endpoint

# From inside the collector pod / host
nslookup ingress.<domain>          # e.g. ingress.coralogix.com
curl -v https://ingress.<domain>   # expect 400/401, NOT a TLS or connection error

If DNS fails → network/VPC/proxy issue, not a collector config issue. If TLS fails → certificate bundle or proxy MITM — check NO_PROXY / HTTPS_PROXY env vars.

Step 3 — Confirm the private key is expanded correctly

# Kubernetes: inspect the live env
kubectl exec -n <namespace> <collector-pod> -- env | grep CORALOGIX
# The key must appear as a 36-char UUID-like string, not the literal "${env:...}" text
# Literal text → bracket syntax wrong, or Secret not mounted

Step 4 — Check the exporter domain: value

In the running config (/etc/otelcol-contrib/config.yaml or kubectl get cm), verify:

• domain: is a bare hostname such as eu2.coralogix.com — no https:// prefix, no UI hostname (app.coralogix.com is wrong)
• private_key: resolved to the actual key (Step 3)

Step 5 — Enable debug logging for one minute

service:
  telemetry:
    logs:
      level: debug

Look for Exporting failed or grpc status lines. A StatusUnauthenticated confirms a key/region mismatch. A context deadline exceeded suggests egress/proxy or ingress-side latency — also check coralogix.timeout (default 5s; increase to 30s).

Step 6 — Inspect pipeline wiring only after Steps 1–5 pass

If export is healthy but data is missing in the Coralogix UI: check receiver connectivity, processor filters (filter processor dropping everything), and that the pipeline is wired in service.pipelines. Full symptom → root-cause table: references/ops-troubleshooting.md.

2. Bring up a new Kubernetes cluster with otel-integration

Use references/setup-kubernetes.md for install flow, per-platform variants, Target Allocator, and chart-specific failure modes. Use references/preset-kubernetes.md when the question is about Helm presets or Infrastructure Explorer.

3. Bring up ECS Fargate (sidecar mode)

Use references/setup-ecs-fargate.md. The fragile pieces are sidecar health checks, dependsOn: HEALTHY, essential flags, and keeping the ecs detector enabled only for sidecar mode.

4. Diagnose memory_limiter firing constantly

Use references/ops-memory-performance.md. Compare Go heap metrics to RSS before changing pod limits or memory_limiter settings.

5. Reduce Span Metrics cardinality or sanitize PII

Use references/data-safety-cardinality.md. Keep the answer layered: prevent bad labels at instrumentation, normalize/sanitize before spanmetrics, and only then discuss collector/backend cardinality limits.

Limitations

• OTTL authoring — use the opentelemetry-ottl skill.
• OpAMP / Fleet Manager internals — preset-fleet-management.md covers only the collector-config overlap (endpoint shape, values-vs-UI precedence, Windows image pitfall); deep supervisor/CDOT work is out of scope.
• SDK instrumentation problems — use the opentelemetry-instrumentation skill.
• Upstream infrastructure (DNS, TLS, proxies, IAM/IRSA, NAT, VPC endpoints) — diagnose to the boundary, then escalate.

References

Upstream links:

• Coralogix OpenTelemetry docs
• telemetry-shippers (Helm charts, ECS task defs, installer)
• integration-definitions (UI wizards)
• OpenTelemetry Collector (upstream)
• OTel Collector Contrib