Debug E2e Pipeline

Debug E2E Pipeline

When To Use

Use this skill when the user wants to debug a failing Prow e2e pipeline for cloud-provider-azure. Trigger on:

• Prow build URLs (prow.k8s.io/view/..., gcsweb.k8s.io/..., storage.googleapis.com/kubernetes-ci-logs/...)
• References to job names like cloud-provider-azure-*-capz
• Requests to "debug the e2e failure", "check why CI failed", "look at the build log", or "analyze the pipeline"

Inputs

Evidence Discipline

Follow these rules when analyzing failures and presenting findings.

• Distinguish facts from inferences. A fact is something directly visible in a log line, timestamp, or artifact. An inference is a conclusion drawn from facts. Always label inferences explicitly — say "this suggests" or "I infer," never state an inference as though it were directly observed.

• Beware single-snapshot bias. A single kubectl get nodes output or one status dump captures one moment in time. Do not treat a point-in-time observation as permanent state. Look for corroborating evidence across the full timeline: were pods scheduled to the node later? Do CAPI Machine conditions show it healthy at a later time? Do controller logs reference it as Ready?

• Require corroboration for strong claims. Before asserting a root cause, require at least two independent pieces of evidence that support the claim and verify that no available evidence contradicts it. If you only have one data point, say so.

• State what's missing. If key evidence is unavailable, say so explicitly rather than filling gaps with assumptions. Missing evidence is not confirming evidence.

• Classify confidence in findings. When presenting root cause analysis, label each finding:

  • Confirmed: supported by multiple independent pieces of evidence with no counter-evidence.
  • Likely: supported by evidence but with gaps — state what additional evidence would confirm it.
  • Uncertain: plausible hypothesis but insufficient evidence — state what you would need to investigate to confirm or rule it out.

Workflow

The workflow has two phases: investigation (delegated to a sub-agent) and review (done by you). The sub-agent gets a fresh context window dedicated to log analysis; you review its structured findings and present to the user.

Phase 1 — Investigation (sub-agent)

Delegate the investigation to a sub-agent. For multiple jobs, spawn one sub-agent per job in parallel — each gets a dedicated context for its job. Pass each sub-agent the context listed below, then tell it to follow the investigation steps that follow.

Context to pass the sub-agent:

• The pipeline URL
• The absolute skill directory path (for running the fetch script)
• The output directory path (default: _artifacts/e2e-debug/{JOB}/{BUILD}/ in the repo root)
• Any user-provided context about what to investigate
• The Evidence Discipline rules from this document
• Any specific investigation questions you have based on the user's request

Replace <SKILL_DIR> with the absolute path to this skill directory.

Step 1 — Fetch and Parse Artifacts

python3 <SKILL_DIR>/scripts/fetch_artifacts.py "<PIPELINE_URL>"

This downloads key artifacts from GCS, parses JUnit and build-log, and outputs a concise summary JSON to stdout. The full manifest is written to manifest.json inside the output directory (default: _artifacts/e2e-debug/{JOB}/{BUILD}/).

The stdout summary contains the most important fields for quick triage:

• status, duration, prow_url
• test_counts: {total, passed, failed, skipped}
• failed_tests: [{name, duration, message}]
• build_log_summary.ginkgo_summary: Ginkgo result line
• build_log_summary.ginkgo_result: "Test Suite Failed" / "Test Suite Passed" / null
• build_log_summary.node_info: raw kubectl get nodes -o wide table

If ginkgo_result is null, tests never produced results. This does not necessarily mean cluster provisioning failed — check build_log_summary.node_info and the build log to determine where the job stalled. Common causes: cluster failed to provision, tests timed out during initialization (e.g. log watcher setup), or pods were stuck Pending for hours. Only conclude "provisioning failure" if node_info is absent or shows NotReady nodes.

If the download fails (404, network error), report the error and ask the user to verify the URL.

Step 2 — Understand the Job

Prefer reading output files directly using your file-reading tools over shell commands for post-processing.

Read the full manifest.json from disk. Start with job_summary:

• job_summary.command / job_summary.args: the entrypoint script the job runs
• job_summary.extra_refs: which repos are cloned (e.g. CAPZ, windows-testing)
• job_summary.env: key environment variables — in particular CLUSTER_TEMPLATE (determines cluster topology and credential provider support), KUBERNETES_VERSION, and AZURE_LOADBALANCER_SKU
• job_summary.presets: preset labels that configure the job
• job_summary.repos: repo versions under test
• job_summary.duration: how long the job ran

This tells you what the job does — which script it runs, which cluster template it uses, and how it's configured. The cluster template determines whether the cluster has OOT credential provider, dual-stack networking, machine pools vs machine deployments, etc. To find which template was used:

1. Check job_summary.env for CLUSTER_TEMPLATE
2. If not set, search build-log.txt for "Using cluster template:"
3. The rendered resources in available_artifacts (e.g. */resources/default/KubeadmControlPlane/*.yaml) show exactly what was applied to the cluster

Check build_log_summary.node_info to see the cluster topology — OS versions, node roles, and container runtime versions. This is critical for platform mismatch issues (e.g. Windows Server 2019 images on 2022 nodes).

Also note available_artifacts — a flat list of all artifact paths available in GCS for this build. Use this as a table of contents: fetch any of these on demand via fetch_webpage on {gcs_base_url}/{path}.

Step 3 — Review Test Results

Check test_counts and tests in the manifest:

• tests.failed: each entry has name, duration, and message
• tests.passed: list of test names that succeeded
• tests.skipped: list of test names that were skipped

If test_counts.failed is 0 but the job failed, the issue is in cluster setup or the test harness, not in any individual test.

Step 4 — Trace the Root Cause

For each failed test, trace through the build log to understand why it failed — don't stop at the first symptom. Distinguish between what failed (the symptom) and why it failed (the root cause). Every claim must be backed by a specific log line, timestamp, or artifact.

1. Find the test's log section: search build-log.txt for the test name. Ginkgo tests are delimited by ------------------------------ markers.
2. Read the full test output: look at pod events, error messages, and the Ginkgo failure block.
3. Build a timeline: correlate timestamps across build-log.txt and other log sources (controller logs, node logs) to establish the sequence of events. Key questions: what was the cluster state before the test? Did anything change during the test? When exactly did the failure occur relative to other events?
4. Trace the causal chain: for each event in the timeline, identify what triggered it — was it a test action (kubectl command), a controller reconciliation, or something else? Cite the specific log line as evidence. Follow the chain until you reach the originating cause: test code → cluster operation → controller decision → etc.
5. Verify assumptions before concluding: if you believe "X caused Y", check whether the evidence supports it. Look for counter-evidence: did X happen before or after Y? Did X happen in other passing runs too? Could something else explain Y? Compare against a passing run if one is available.
6. Cross-reference external repos: if job_summary.extra_refs shows repos like windows-testing or cluster-api-provider-azure, the cluster template or setup script may live there. Use fetch_webpage to read the relevant files from those repos.
7. Check bootstrap controller logs: clusters/bootstrap/controllers/ contains logs from cluster management controllers (e.g. CAPZ, CAPI) that can reveal background operations like node replacements, rolling updates, or health check remediations affecting the test environment. Download with --fetch "*/controllers/*/manager.log".

Step 5 — Fetch Additional Artifacts (optional deep-dive)

Use --fetch to download any artifact listed in available_artifacts by glob pattern:

python3 <SKILL_DIR>/scripts/fetch_artifacts.py "<PIPELINE_URL>" --fetch "*/controllers/*/manager.log"
python3 <SKILL_DIR>/scripts/fetch_artifacts.py "<PIPELINE_URL>" --fetch "*/MachinePool/*.yaml" "*/MachineHealthCheck/*.yaml"

If the build was already fetched, --fetch reuses the existing manifest and only downloads the matched artifacts.

For node-level issues (kubelet errors, credential provider problems), use --fetch with a pattern like "*/machines/*/kubelet.log" or fetch individual files via fetch_webpage on {gcs_base_url}/{path} using paths from available_artifacts.

Step 6 — Return Findings

Write findings to findings.md in the output directory, then return them to the main agent. Use this structure:

1. Output directory: the absolute path where artifacts were downloaded
2. Job overview: name, build ID, status, duration, cluster topology
3. Failed tests: which tests failed and their error messages
4. Root cause analysis: the full causal chain explaining why the test failed, not just what failed. Each link in the chain must cite a specific file path and line number as evidence. Clearly separate symptoms (the error the test hit) from the root cause (what created the conditions for that error).
5. Hypotheses considered: each with supporting evidence, contradicting evidence, missing evidence, and confidence level (Confirmed / Likely / Uncertain)
6. Relevant log snippets: key error lines with file path, line number, and surrounding context
7. Suggested next steps: what to investigate or fix. Do not propose a fix until the root cause is established with evidence.

Phase 2 — Review and Present (main agent)

Review the sub-agent's findings before presenting to the user:

1. Spot-check key citations: for the primary causal chain and any Confirmed findings, verify the cited log lines and timestamps by reading them directly. Do not blindly trust sub-agent citations.
2. Check confidence labels: are they appropriate given the evidence? Downgrade if the sub-agent overclaimed.
3. Look for gaps: did the sub-agent miss an obvious log source or artifact that could confirm or refute the hypothesis?
4. Cross-compare jobs: when multiple sub-agents ran (one per job), compare their findings.md files. Classify failures into shared root causes — identify which jobs hit the same underlying issue so you only explain each root cause once. Look for what varies across passing and failing jobs (e.g. K8s version, cluster template, worker topology, region, timing) to narrow down what correlates with failure.

Then present the reviewed findings to the user, including confidence labels on each root cause claim.

Comparing Jobs

When debugging, it's often useful to compare a failing job against a passing run of the same or a similar job. If the root cause is not clear from the failing job alone, ask the user for a passing run URL to compare against. The workflow:

1. Run the script twice — once per job URL
2. Compare job_summary.command and job_summary.args — different entrypoint scripts or post-commands?
3. Compare job_summary.env — different environment variables?
4. Compare test_counts — did one job run more/fewer tests?
5. Compare tests.failed — same failures or different?
6. Check if one job has no junit_files at all — tests never started
7. Compare build_log_summary.node_info — different OS versions or node counts?

Bundled Resources

• scripts/fetch_artifacts.py — download, parse, and discover GCS artifacts