raft / docs

Measure startup on your host

Creation can return a handle before the box accepts commands or Docker is ready. Measure the readiness your workload needs rather than treating CLI return as full startup.

Raft development images are native Debian 13 images for ARM64 and AMD64, not emulation. Images and recovery targets must match host architecture. Get images through onboarding, or use the upgrade guide to build and pin another image.

Read native startup samples

Five samples per architecture from native CI on 2026-10-06, using 4-CPU hosts and 1-CPU, 2-GiB boxes with cached images:

MedianARM64AMD64
Create returns handle0.425 s0.659 s
Create accepts a command0.866 s1.375 s
Docker ready after create2.280 s3.035 s
Resume accepts a command0.951 s1.489 s
Docker ready after resume2.626 s3.049 s
Desktop ready after its separate start request1.115 s1.592 s
Compressed image1.79 GiB1.86 GiB

Read the ARM64 samples and AMD64 samples for distributions. Hardware and SSH latency differ between runners, so these numbers do not establish an architecture advantage or compare Raft with Boat.

These samples used the noVNC desktop before Voxl was added. Their desktop timings and image sizes do not measure the Voxl image.

Measure your own host

Use a dedicated test location with a pinned image and a matching source checkout. The benchmark creates, stops, resumes and destroys its own boxes. It also starts the desktop when requested; do not target a shared or daily-use test channel.

python3 deploy/benchmark-boot.py --location lab --samples 5 --desktop > boot.json

You can request 1 to 20 samples. Each sample measures request-to-return, command readiness and Docker API readiness, then repeats after stop/resume. Desktop timing starts at its separate startup request and waits for a Voxl PNG screenshot. It does not measure remote browser rendering.

Reports include median, minimum, maximum, nearest-rank p95, SSH round trip, image size and guest boot information. With three or five samples, p95 is effectively the maximum, not a stable production percentile.

All controller-visible times include CLI and SSH overhead. They exclude image downloads and host reboots. First use after import or cache eviction can take longer while storage unpacks; hardware, load, packages and services also affect results. A stopped box restores files, not process memory.

Compare earlier measurements

The host A, host B and candidate reports retain three-sample ARM64 comparisons. Local desktop timings exclude SSH overhead. These samples show image-size and desktop-start improvements, not an overall warm Docker boot speedup.

Controller read timings measure list, capacity and doctor separately from startup. They are network-dependent exploratory samples, not guest workload benchmarks. See verification boundaries before treating any sample as a guarantee.

On this page