Kalisto/ documentation source (permission required) ↗
On this page

#Startup budget and live validation

The acceptance target is a complete functional fresh three-node cluster under sixty seconds, not merely a running VM or a stored Kubernetes Node Ready condition. Preparation is deliberately separate.

#What the clock includes

The test runs the real CLI from an empty working directory with system-only PATH. pull finishes before the timer. Inside the same clock:

  1. Create three fresh cluster VMs; precreated VMs do not qualify.
  2. Import local digest-pinned images.
  3. Reach three Ready nodes and an authenticated host Kubernetes API.
  4. Finish every rendered Cilium DaemonSet/Deployment rollout.
  5. Resolve DNS from both workers.
  6. Complete cross-node HTTP through a Kubernetes Service from both workers.

Probe images use cached platform digests and imagePullPolicy: Never. Cilium uses embedded Envoy with L7 policy support, not a separate Envoy DaemonSet. Preparation does not create cluster state/benchmark VMs. Manual k3s bootstrap, precreated cluster VMs, or partial readiness do not satisfy the acceptance scenario. First-time runtime installation/downloads/image preparation are not covered by a sub-minute claim.

#Observed results and evidence boundaries

Scenario Result Boundary
Native Apple-container CI, source 015c15f, run 4478 38.11s (38.111266833s recorded) Matching-source full API/Cilium/DNS/cross-node checks; installed Homebrew runtime reuse
Earlier README native Apple-container result 46.71s Historical completed run using existing Homebrew container 1.7.1; not the latest candidate measurement
Optional Lima, four development-Mac runs 56.07s, 53.84s, 54.26s, 54.83s Historical Lima measurements, not Apple-container results
Linux Firecracker, Fedora 44 x86_64, Firecracker 1.18.0 42.95s Fresh three-node startup and all functional probes; preparation separately 40.68s

These are measured results, not universal hardware guarantees or runtime proof for every compiled architecture. Older agent guidance recorded an Apple 52.04s run predating the service-reuse-specific proof; it is historical, not the current result. Candidate version 0.0.0-ci.4476+g015c15f4bb23 is built from 015c15f; no stable release is claimed.

Apple stop/resume preserved the ConfigMap and scale-out reached four Ready nodes; authenticated kubelet health checks passed for every resumed and added node. Historical Linux persistence/scale evidence remains platform-specific. Those lifecycle checks occur after the startup timer. Archive export/import and restored startup have no sub-minute guarantee; native restore proof is still pending.

#Authorization, storage and ownership

CI is the default execution path. Run live validation only in the authorized native job; local builds/benchmarks require explicit user authorization. Build/image artifacts, caches and large temporaries must stay on the configured mounted external build volume, never the internal SSD. Cluster VM runtime disks may use local SSD. Tests must destroy only their owned cluster, not borrowed services, shared caches or unrelated state.

The Apple CI job borrows a running Homebrew 1.7.1 service. Setup may start a stopped installed/unregistered service but never replaces a running daemon, installs a kernel into it or changes its roots. It records CLI/daemon executable identity, service roots, process identity, configuration and default system properties before flag-free pull/up with system-only PATH, comparing after preparation, stop/resume, scale and destruction. It asserts no private installer download/selection. Cleanup never stops the borrowed daemon or deletes its persistent application root.

#Apple-container live budget

From a source checkout, only in authorized native macOS CI:

KALISTO_LIVE_APPLE_CONTAINER_BOOT_BUDGET=1 \
  cargo test --release --test macos_three_node_boot_budget \
  apple_container_k3s_has_functional_three_node_cilium_under_one_minute -- --nocapture

Source: three-node budget test.

#Optional Lima release budget

Native VZ preparation supplies a sanitized compressed raw image. No manual template, hand-built qcow2 or host Helm/Crane is required.

KALISTO_LIVE_THREE_NODE_BOOT_BUDGET=1 \
  cargo test --release --test macos_three_node_boot_budget -- --nocapture

The test supplies its own empty workdir and does not use KALISTO_LIMA_TEMPLATE. Evidence under target/kalisto/boot-budget/ includes aggregate summary.json, per-run result.json, preparation/startup logs, pod/events JSON and VM boot/image-import logs captured before cleanup.

#Firecracker live test

Requires the Linux prerequisites, explicit existing local kernel/rootfs and authorization:

KALISTO_LIVE_FIRECRACKER_TESTS=1 \
KALISTO_FIRECRACKER_KERNEL=/absolute/path/to/kalisto-vmlinux \
KALISTO_FIRECRACKER_ROOTFS=/absolute/path/to/rootfs.raw \
  cargo test --release --locked --test linux_firecracker_live -- --nocapture

It runs the real CLI from an empty directory with a dedicated reusable XDG cache and fresh unique cluster state. pull is outside the strict sixty-second startup/probe clock. Persistence and scale-out are checked afterward. Evidence is retained under target/kalisto/firecracker-budget/ or the corresponding CARGO_TARGET_DIR. Only the test-owned cluster is destroyed. Source: Linux live test.

#Legacy Lima live probes

These opt-in source-checkout tests are not the product's prepared no-CNI/Cilium baseline. Contributor probes explain their Flannel configuration.

KALISTO_LIVE_LIMA_TESTS=1 \
cargo test --test macos_lima_live -- --nocapture
KALISTO_LIVE_LIMA_K3S_TESTS=1 \
cargo test --test macos_lima_k3s_live -- --nocapture

Worker-join variant:

KALISTO_LIVE_LIMA_K3S_TESTS=1 \
KALISTO_LIMA_K3S_WORKERS=1 \
cargo test --test macos_lima_k3s_live -- --nocapture

The separate archive live test has pending native restore evidence. Fake disk tests do not prove live VM restoration.

#Product benchmark versus acceptance

kalisto bench --runs 3 --workers 2 --addons cni/cilium

bench runs the real lifecycle repeatedly and prints per-run timings, but the functional budget tests above provide stronger API, DNS and cross-node traffic assertions. A lifecycle duration alone is not equivalent proof. Do not run this optional benchmark without authorization.