Internal Developer Platform  ·  View 17 of 21  ·  Operations

Shared Pipeline and Environments

One artefact, promoted unchanged, with recovery that does not depend on the pipeline.

Editable source SVG draw.io All views
Source Pull Request Pipeline Reference consumed, not copied Build Build ephemeral runner Sign and Attest provenance · SBOM Gate Advisory Policy Tests and Scans Dev and stage Dev auto on merge Stage same artefact Production Admission Check binding Canary automated analysis Full Rollout Recover Rollback ≤ 5 min · pipeline-free Deploy Record 400 days signal breach digest verified Internal Developer Platform — Shared Pipeline and Environments External / third party Application we own Security / platform Decision point Data store failure / alternate synchronous One artefact is promoted through every environment. A rebuild per environment would make the thing tested and the thing running two different objects. v 1.0 · owner Platform Architecture · date 2026-09

Decisions

  • The pipeline is consumed by reference, not copied into 1,400 repositories, which is what makes a fix shippable once (ADR-03).
  • The same signed artefact is promoted through every environment: a rebuild per environment makes the thing tested and the thing running two different objects (ADR-16).
  • Rollback runs outside the pipeline, because the failure may be the pipeline (ADR-13).

Targets

  • Commit to production ≤ 25 min p95 excluding deliberate soak; rollback to serving the previous version ≤ 5 min p99 (stated assumptions).
  • 5,200 deploys per week, peaking at 180 per hour midweek; 12,000 CI jobs per day.

Omitted

  • Per-pull-request ephemeral environments, deferred to Phase 3 in the requirement — they multiply provisioning load and need a data story first.
  • Manual approval gates, which are absent by design: if a check is worth blocking on it is worth automating.