diagnosing-ci-and-merge-bottlenecks

Diagnosing CI and merge bottlenecks

Engineering analytics treats a pull request like product analytics treats a user: a PR moves through a pipeline (opened → CI → review → merged → deployed) and the job is to find where it slows down. The surface is named MCP tools — you call them, you don't write SQL. Dogfooded on PostHog/posthog; the same tools serve autonomous agents (e.g. PostHog Desktop) reasoning about their own PRs. Scope is aggregate pipeline health: to take one failing test or red run to a verdict, switch to the investigating-ci-failures skill.

The tools

• `pull-requests` — the PR workhorse. Open PRs plus anything merged or closed since date_from (default

-30d), newest first. Each row carries author (nested object: handle, display_name, is_bot), repo (nested: owner, name), state, is_draft, labels, open_to_merge_seconds, ready_to_merge_seconds, and a ci rollup (runs / passing / failing / pending) from the head-SHA join. Answers most PR-level questions: which PRs have failing or pending CI, which are stuck open longest, per-author or per-repo triage, and time-to-merge stats (aggregate over the returned merged rows yourself, median and p95, never a mean; prefer ready_to_merge_seconds where non-null, it excludes draft time).

• `workflow-health` — per-workflow CI health over a window (date_from / date_to, default last 24 hours):

run_count, success_rate, p50_seconds, p95_seconds, last_failure_at. Answers "is CI getting faster or slower" and "which workflow is the slow or flaky long pole". There is no built-in trend — call it over two adjacent windows and compare. success_rate covers runs that succeeded or ended in a decisive failure (failure, timed_out, startup_failure, or stale), excluding skipped, cancelled, neutral, and action-required runs. p50_seconds / p95_seconds cover successful runs only because cancelled and failed runs end early and would bias the duration trend. Each is null when a window has no qualifying runs — guard for null before comparing two windows (a workflow can have runs in one and none in the other). run_scope=pull_request scopes to PR-attributed runs, excluding master/main (same-repo PRs only — fork runs carry no PR attribution).

• `pr-lifecycle` — a single PR's timeline: a header plus ordered events — opened, ready-for-review and

converted-to-draft transitions (when the issue-events table is synced), then a CI started/finished pair per workflow run (many on a multi-workflow repo, interleaved by time), then merged/closed. Answers "where is PR N stuck". metric_quality is partial (no review or comment events).

• `engineering-analytics-flaky-tests` — the active test-health queue from the per-test CI spans, over a

window (date_from default -7d, max 30 days). Evidence is counted per CI run, never per span or run attempt. classification is confirmed_flake only where the evidence proves nondeterminism (same_commit_recovery_run_count > 0: one commit both failed and passed the test in the same matrix job, via a "Re-run failed jobs" attempt going green or an in-job retry; a pass in a different leg, such as FOSS against EE, is not recovery); quarantined means a tolerated failure was recorded while masked; suspected_regression means only failures were recorded, which is absence of proof, not proof of a real break. A test qualifies on any same-commit recovery, a quarantined failure, any master/main failure, or failures on ≥ min_failed_prs distinct PRs (failed_pr_count). Answers "what is this failing test costing us" and picks quarantine candidates. It does not answer "which tests are flaky": this queue only sees the main Backend pytest and Frontend Jest suites, and recovery proof only arrives when someone re-runs failed jobs (or a pytest test is hand-marked @pytest.mark.flaky(reruns=N)). Counts are absolute signal, never rates: passing runs are mostly not emitted, so there is no honest denominator. A CI setup break excludes every trial of the attempt, not only its failures: a run attempt whose tests errored in 3 or more jobs or for 3 or more owning teams, or a job attempt with 100 or more distinct failed or errored tests. A real break that wide is left out too, so when a whole suite or job goes red, check workflow-health rather than reading an empty queue as healthy.

• `engineering-analytics-sources`: the team's connected GitHub sources and repos. With more than one of

either, call it first and pass the chosen entry's source_id and repo to pull-requests, workflow-health, and pr-lifecycle. Passing only source_id reads that source's default repo, not the one you picked. With a single source and repo the tools default to it.

There is no aggregate time-to-merge tool and no "counts" tool — derive those from pull-requests (the stuck/failing counts, the merge-time percentiles).

Caveats you must carry into every answer

These are structural limits of today's snapshot data — state them, don't paper over them.

• `open_to_merge_seconds` is coarse. It fuses draft time and ready-for-review time into one figure. Report

it as "open to merge", never "cycle time" or "review time". Flag it when long-lived drafts inflate a number.

• `ready_to_merge_seconds` is the precise companion: merged_at minus the last observed ready-for-review

transition (only the last draft/ready switch counts), or minus created_at for a merged PR verifiably never drafted. Null means "not observed" (the PR's life isn't fully inside the synced issue-event window, or the table isn't synced), never zero, so aggregate only over non-null rows and say how many were observable.

• CI status can be stale. The CI source syncs on a watermark and does not refresh a run that completes after

newer runs land (until the workflow_run webhook ships). Treat a pending count as unsettled, not as a settled failure; lead with status, not a verdict.

• CI for a PR is the head-SHA join, nothing else. The ci rollup reflects only the latest commit's runs. There

is no other link between a PR and its checks.

• Review reads are deferred by choice. The GitHub reviews endpoint syncs review submissions with their timestamps, but reads stay deferred until a wedge tool needs them. Don't infer review behaviour from their absence.

pr-lifecycle is partial for the same reason.

• Deploy metrics live on `engineering-analytics-dora`, from the GitHub deployments tables when synced

(deploy_data_available). Its change-failure and time-to-restore fields are deploy-status proxies (no incident link) — report them under their honest field names, never as the true DORA definitions.

• Bots and drafts are present in `pull-requests` output, excluded by convention. Filter out author.is_bot

(nested under author, not a row-level field) and is_draft for throughput / merge-time questions; keep them in for bot-impact questions.

• `pull-requests` returns a capped page. 1000 rows, newest first: a fixed server-side cap, echoed in the

response as limit, that no parameter raises. When truncated is true, any percentile or count you derive covers only that newest page, not the whole window. Say so, then narrow until the real set fits: author filters to one handle, source_id / repo to one repo, and date_from shortens the window.

Choosing a tool

The high-value chain

Mirror how a human investigates: aggregate signal → confirm → concrete PR.

workflow-health  (find the slow/flaky long-pole workflow)
   → pull-requests  (confirm it's dragging merge time; list the affected PRs)
      → pr-lifecycle  (open a representative stuck PR and show the gap)

"CI median rose because e2e-playwright p95 doubled; that workflow is the long pole on PR #1234, which sat 47m in CI before merging."

Output expectations

• Lead with the verdict in one line, then the supporting numbers.
• Carry the coarse / partial / staleness caveat whenever the distinction matters.
• For multi-window or multi-workflow comparisons, a short table beats prose. Report median and p95 side by side —

never collapse them into one "average".

What NOT to do

• Don't call open_to_merge_seconds cycle time or review time — it's coarse open-to-merge;

ready_to_merge_seconds is the cycle-time figure, and only where non-null.

• Don't report a CI count as a settled failure when pending > 0 — it may be unsettled or stale.
• Don't infer reviews or approvals — review reads stay deferred until a wedge tool needs them. Don't infer per-check counts. Deploys come from

engineering-analytics-dora, not from inference.

• Don't turn per-author buckets into a leaderboard — they're for finding stuck work, not ranking people.
• Don't reach for these tools to fetch raw PR contents or diffs — they surface pipeline signal, not the PR thread.

Persisting an answer

These tools are ad-hoc reads; they cannot be saved as an insight or subscribed to. When the user wants the same numbers as a saved insight, a dashboard tile, or a scheduled email/Slack delivery, switch to the turning-engineering-analytics-into-insights skill: the underlying warehouse tables (<prefix>github_pull_requests / <prefix>github_workflow_runs, prefix from engineering-analytics-sources) are directly queryable with HogQL, and that skill carries the curated column semantics plus the insight-create / subscriptions-create workflow.

DORA environment defaults include all persistent production-marked or production-named regions. Read selected_environments for the validated exact scope. DRF validates explicit filters against real names, including transient environments. It trims and deduplicates valid names and rejects blank or unknown names with a 400 response. Omit the filter to use the default. Each regional deployment counts separately; lead time counts each PR once at its first successful deployment in any selected environment.