Cloud Development Environment Platform  ·  View 17 of 21  ·  Operations

Delivering the Platform Itself

One pipeline for the host agent, the host kernel and the control plane, with a gate that cannot be waived.

Editable source SVG draw.io All views
Source Host agent Rust · Firecracker API Host kernel config Control-plane charts Build Signed artefacts digest-pinned SBOM Gates Unit & integration VM boot matrix every machine class Snapshot fidelity restore = original Canary 1% of hosts synthetic workspaces Resume SLO watch p95 · success rate Fleet Drain-and-patch waves rate-limited Live evacuation Recover Roll back a wave ≤ 10 min Halt on SLO breach regression blocks release Delivering the Platform Itself — Kernel, Host Agent, Control Plane Application we own Security / platform Data store Decision point Risk / gap failure / alternate The snapshot-fidelity gate is the one that cannot be waived: a release that restores a workspace wrongly destroys work rather than delaying it. v 1.0 · owner Platform Architecture · date 2026-10

Decisions

  • Snapshot fidelity — a restored workspace is indistinguishable from the original — is a release gate, because a release that restores wrongly destroys work rather than delaying it.
  • The canary is 1% of hosts carrying synthetic workspaces, watched on resume p95 and resume success rate rather than on error rate alone.
  • Fleet rollout is the same drain-and-evacuate mechanism the operator journey uses; there is no separate maintenance path.

Assumptions

  • A wave can be rolled back in ≤ 10 minutes; a resume-SLO regression halts the rollout automatically.
  • Every artefact is signed and digest-pinned, with an SBOM retained per release.

Risks

  • The kernel is the isolation boundary and the thing most often patched; a regression here is simultaneously a security fix and an availability risk.
  • A canary on synthetic workspaces can miss what only real workspaces do — large disks, odd nested containers, long-running dev servers.