Skip to content
senn-techsenn-tech
Back to blog
Infrastructure
Servers, Proxmox and network2026-11-16· By Franz Senn

Kiri starts containers fast, the eight-times number only holds on a Mac

Kiri, a container runtime written in Rust that works without Docker Desktop, runs run --rm alpine true 8.1 times faster than Docker Desktop on a Mac (22 instead of 178 milliseconds); on Linux the vendor's own measurement files show 3.7 times. Better Stack presented Kiri in a video on 14 November 2026 that quotes the eight-times figure. On 16 November 2026 we compared the vendor's measurement files and recomputed what of that arrives in server operation on our Linux hosts; the files were read again on 19 November 2026. The number is documented cleanly and describes a Mac machine, on which nothing runs at our place.

Kiri is a container runtime written in Rust that reproduces the Docker Engine interface. It works without Docker Desktop, dockerd, containerd and runc: on the Mac it runs inside a purpose-built small Linux VM, on Windows in WSL 2, on Linux directly on the host. The Docker CLI, Compose and Buildx talk to it through a dedicated Docker context. The licence is Apache-2.0, with the name excluded.

Two measurement files, two different results

The project handles benchmarks openly: scripts in the repository, raw values, one PDF per run. There are two files with the same metrics, one in the repository under bench/results.json, one delivered under runkiri.com/bench/results.json. Their values differ, and the video quotes the second one.

Measurement path, p50 in millisecondsKiriDocker DesktopOrbStackColimaMachine and date
run --rm alpine true through the CLI, bridge network22.0178.0279.0152.0Apple M3 Max, macOS 27.0.1, 2 CPUs, 8 GB, 2 November 2026
same path, repository file14.15243.3442.2no rowApple Silicon, macOS 27.0.1, no limit documented, 24 October 2026
Docker API, --network none3.1163.0238.088.3as row 1

All values are vendor figures from the two files; we did not re-measure them. 178 divided by 22 is 8.1. With the repository file the same readings give 17.2, and at the API level 83.7, where the README puts 2.55 milliseconds against 213 for Docker Desktop and speaks of roughly 80 times faster. Depending on the file, the same measurement yields a factor between 8 and 84.

The files sharpen the picture themselves: the live file states that most of those 22 milliseconds are the starting Docker CLI and not the runtime. The repository file notes that OrbStack was holding 86 containers and a Kubernetes cluster during the run, so its values came out too poor. Better Stack writes in its guide of 16 November 2026 that the figure applies to the sequence start, true, exit, remove, and says nothing about work inside a running container.

On Linux the vendor measures 3.7 times itself

On Linux a container starts in 188 instead of 697 milliseconds with Kiri, according to the vendor file of 26 October 2026. For server operation this third file counts, bench/ubuntu-2026-10-26.md. It was measured on a 16-core machine with Ubuntu 24.04.4 as a WSL 2 distribution, kernel 6.6.87, 32 GB of memory, medians from hyperfine. According to the file this is not bare-metal Linux; the scripts rerun unchanged on a native Ubuntu host. run --rm alpine true costs 188 milliseconds with Kiri, against 697 for the Docker Engine, 631 for nerdctl, 635 for Colima, 714 for Docker Desktop and 1105 for Podman. That is a lead of 3.7 times, not eight. On Windows 11 with WSL 2 and Docker Desktop the vendor measures 78 against 491 milliseconds, which is 6.3 times.

Kiri starts 3.7 to 8.1 times faster, depending on platformMac vs Docker Desktop8.1 · 22 vs 178 msWindows vs Docker Desktop6.3 · 78 vs 491 msLinux (WSL 2) vs Engine3.7 · 188 vs 697 ms09
Factor = time of the comparison runtime divided by Kiri’s time, vendor medians, division by us. Machine and comparison runtime differ per row. (source: Kiri, bench folder (vendor data))
Linux (WSL 2), medianrun --rm alpine true in msexec in msPort throughput in Gbit/s200 MB into the container in MB/s
Kiri, state of 26 October 202618818710.3invalid
Kiri, after two fixes the same day214 to 230191 to 19955.8 to 58.0200 to 207
Docker Engine 29.9.2697224100.7129
nerdctl 2.4.16315752.4131
Podman 4.9.3110527451.9121

The file also says where Kiri lost: an exec into a running container at 187 milliseconds against 57 for nerdctl, 10.3 Gbit/s through a published port against 100.7 for the Docker Engine, and input from outside only up to 4 MiB, because the buffer MAX_PENDING_INPUT applied inside the guest agent; 200 MB arrived as 1.2 MB. The same document states under "After the two fixes" that flow control for input and kernel-level forwarding of published ports closed those two gaps, with the start taking 214 to 230 instead of 188 milliseconds. That puts the lead over the Docker Engine at 3.0 to 3.3. The input flow control now exists in the source of main (INPUT_WINDOW in kiri-proto). Our own re-measurement on a real Linux host is still open.

An independent measurement by ComputingForGeeks of 16 November 2026 ran on four Google Cloud VMs with 2 vCPUs and Ubuntu 26.04.1 each, one engine per VM. Kiri 0.2.1 needs a mean of 69.5 milliseconds for run --rm alpine:3.22 true, Docker CE 29.9.3 needs 349.2, a factor of 5.0. Starting nginx and Redis with Compose drops from 708 to 175 milliseconds, while an uncached build takes 2,035 milliseconds with Kiri against 1,953. The test also names rough edges: the service does not start by itself after installation, containers run unconfined under SELinux, and on a host that also runs Docker CE, Kiri traffic is lost until two iptables rules are added. This is a single measurement by a third party, and we did not repeat it.

Where the lead on the Mac comes from

The primitive table in the live file, measured inside a Linux VM, explains the number better than any headline: fork plus exec 0.54 milliseconds, four namespaces plus exec 0.78, with a network namespace 1.15, creating, limiting and removing a cgroup v2 1.23, overlay mount and unmount 0.36. The most expensive step of a --rm run is deleting the veth pair at 18.9 milliseconds, which Docker itself performs before returning from --rm. There is nothing to win in the kernel work.

The path of one Docker command in Kiridocker CLI, Composespeaks the Engine APIUnix socket~/.kiri/run/docker.sockkiri daemonowns the VMvsockMac to guestGuest agent, PID 1clone3 into namespaces
Neither runc nor a shim sits between agent and container; the agent starts it with a single clone3 into namespaces and a cgroup. (Quelle: Kiri, docs/development.md (architecture))

What is winnable on the Mac sits above it: Docker Desktop and OrbStack pay for building a VM there, plus dockerd and containerd and one shim process per container. The live values name 216 MB for dockerd and containerd with OrbStack, 7.9 MB per shim, 1,133 MB at 143 shims, and a CLI floor of 19 milliseconds. Kiri arrives with 9.7 MB, of which 9.0 on the Mac and 0.69 in the guest, and no separate process per container. Its own VM is built small: 810 milliseconds with the standard arm64 kernel, 112 after trimming the kernel (allnoconfig, faster RCU, a patch against the reset polling of virtio-pci), of which 95 sit in Apple's Virtualization.framework; the guest kernel needs 8 milliseconds until PID 1.

On Linux the saving falls away, because no VM sits in between: Kiri works with clone3 directly in namespaces and cgroups v2, with the same mount and network work as the Docker Engine. Hence 3.7 and not 8.

What a user measures on the Mac under real work

In the project's issue tracker a user measured Kiri 0.2.1 against OrbStack 2.2.5 on 16 November 2026, on a Mac mini with an M4 Pro and macOS 27.0.1 (issue 362). Kiri ran with two vCPUs and 8 GB, OrbStack with 14 vCPUs, 16 GB and 16 other running containers; the machine's load average was 4 to 5. The maintainer closed the issue on 17 November 2026 as "not planned", saying benchmark results are not tracked as issues. The numbers are a single measurement by a third party, and we did not repeat them.

Measurement, median in milliseconds, pass 1KiriOrbStackFactor
run --rm alpine true472575.5
Compose with Postgres, Redis and app, up --wait111738663.5
MariaDB 11.8, empty to ready380554571.4
build --no-cache, one CPU-bound step254228231.1

Starting is several times faster, the work inside the container barely. A second issue shows the flip side: reading many small files through a bind mount, Kiri needs 4,476 milliseconds for a tar over 20,680 files on an M3 Pro with macOS 26.4.1, OrbStack 695 (issue 415, open on 19 November 2026). Anyone developing Node projects with node_modules in a bind mount spends most of the waiting time there.

Where Kiri stands next to Apple container, OrbStack and Podman

The video compares against Docker Desktop, but the Mac has more choice. Apple publishes its own runtime called container, written in Swift and built on the Containerization package.

Logo of Apple
Apple requires a Mac with Apple silicon and macOS 26 or later for container; according to the README, older versions are not supported. (Quelle: Apple / Wikimedia, public domain)
RuntimeRuns onVM modelCost for companies, as of 19 November 2026
Kiri 0.2.1macOS 14 or later (Apple silicon, Intel), Windows 10 and 11, Linuxone small VM for all containers, no VM on LinuxApache-2.0, no free threshold
Apple container 1.7.1Apple silicon only, macOS 26 or laterone lightweight VM per containerApache-2.0
Docker DesktopMac, Linux, WindowsLinux VM, on Apple silicon optionally Docker VMM (beta)free below 250 employees and 10 million US dollars revenue, above that from 9 US dollars per user and month (Pro, annual billing)
OrbStackMac (Apple silicon and Intel per its docs)one lightweight Linux VM with a shared kernelfree for personal use, 8 US dollars per user and month for business
PodmanLinux natively, Mac and Windows through podman machinea VM on Mac and Windows (Mac: libkrun or applehv)Apache-2.0

According to the Containerization README, Apple runs each container in its own VM and cites start times under one second. Kiri shares one VM and starts containers inside it in the millisecond range. When switching, Kiri creates a Docker context named kiri on first start, and kiri import brings images, volumes and containers over from OrbStack or Docker Desktop (README, as of 19 November 2026). For Podman, the security flaw is covered in Podman CVE-2026-94603: an image annotation switches the sandbox off. Which of the two designs wins on memory and isolation has been measured neither by us nor by the vendor; Kiri's measurement files do not compare against Apple container.

What our Linux fleet would gain: nothing

On 16 November 2026 we looked read-only with docker ps across our eight container hosts. There 120 containers run, none of them started on that day. The oldest start sits on 27 October 2026, the youngest on 11 November 2026. Four hosts run Ubuntu 26.04 with Docker 29.8.0 and containerd 2.3.4, the other four Alpine 3.23.8 with Docker 29.5.2 and containerd 2.2.0.

What costs money in server operation lies elsewhere: long-running processes, restarts after updates, memory at rest. A container in server operation rarely goes through the start that Kiri accelerates: on our eight hosts, all 120 containers last started three weeks ago at most (measured November 16, 2026, see above). The difference of 509 milliseconds hardly shows there. Its value sits where a stack comes up dozens of times a day: on the development machine. For the choice between containers and guest operation on Linux, our article Docker or LXC on Proxmox works through the estate.

Licence, provenance and a root group

Apache-2.0 with a NOTICE file, name and logo excluded, forks under a different name expressly welcome, so commercially usable without a free threshold; the guest kernel is Linux under GPL-2.0. For comparison, Docker's position: according to the licence page, Docker Desktop is free for companies with fewer than 250 employees and less than 10 million US dollars in annual revenue, for personal use, education and non-commercial open source projects. Above that Docker requires a subscription, with prices on the pricing page.

Docker logo
Docker ties the free use of Docker Desktop to 250 employees and 10 million US dollars in annual revenue. (Quelle: Simple Icons, file docker.svg (CC0), trademark: Docker, Inc.)

The maintainer density is where we see the fragility. The repository has existed since 2 November 2026, there are six versions with v0.2.1 of 15 November 2026, 147 Rust source files, 98 open issues and 6,257 stars (as of 16 November 2026). 88 of 104 author lines in the history belong to one account, which is 84.6 percent. On 19 November 2026 GitHub shows 10,956 stars, 124 open issues and 85 open pull requests; the main account has 119 commits, the next nine accounts 1 to 5 each. A project of this density can disappear in a quarter, the way it appeared.

On Linux one more line belongs on the table before a first installation. The daemon runs as root and listens on /run/kiri/docker.sock, the group kiri may write to the socket, and the vendor calls membership root-equivalent in its security docs, like the Docker group. kiri-lite delivers the engine part alone, BuildKit runs over the docker-container driver of Buildx, a built-in builder stands on the roadmap. Behind that sits the gap we would notice first with Compose: --add-host with host-gateway reports "host-gateway is not supported by the docker-container driver", as the known issues say.

On Mac security, the release notes for 0.2.2 have been in the repository since 19 November 2026; the release itself was not yet published that morning. They list several mostly privately reported flaws, among them: an image could reach files on the Mac that the user had not shared, and the VM will now get /Users, /Volumes and /private instead of the whole disk. The repository's security page lists no published advisories on 19 November 2026.

The sharing feature sends data outward

Usable in a company environment? For development on a laptop yes, with open eyes. For anything attached to real data only without the sharing feature: Kiri Share makes a local container reachable from outside under an address on *.runkiri.cloud. The vendor docs and the Share page describe it this way: sign-in through GitHub, one outbound encrypted connection to the edge server edge.runkiri.com through which every visit returns to the local port, and method, path, status and time per request shown in the interface. Whether the vendor stores those details the docs do not say. After an abort the connection rebuilds itself. Version 0.1.17 repairs an abort report that sent the GitHub account along.

The pages that would have to regulate privacy and the legal notice cannot be found on runkiri.com: privacy, terms, imprint, legal, about, security and docs answer on 19 November 2026 with exactly the same 80,373 bytes as the start page, an invented path variant does too, and the start page links to none of them. The control against share/ returns different byte counts, so the measurement is not running into emptiness. The vendor's security docs sit on the subdomain docs.runkiri.com.

What we checked ourselves

Every figure comes from vendor files of 16 and 19 November 2026; the estate was measured read-only on 16 November 2026. We did not install or start Kiri, so the Mac figures stay vendor values, and it is open whether the port and exec deficits exist on macOS as well. The values from issues 362 and 415 come from third parties.

Checking it is small in scale: set up a Compose track with the real services, measure start, exec, bind-mount reads and port throughput against the runtime in use, then decide about the sharing feature. For developers on a Mac the test is worth it; for Linux servers the vendor figures give no reason to switch. More operations reports on containers and virtualisation are in the hub Servers, Proxmox and networking, and our IT services for companies cover a look at your own estate.

Further Reading

Questions?
Is Kiri faster than Docker Desktop?+

At startup yes, for work inside the container barely. On the Mac the vendor measures 22 instead of 178 milliseconds for `run --rm alpine true` (8.1 times, measurement file of November 2, 2026). On Linux it is 3.7 times (188 against 697 milliseconds, WSL 2, October 26, 2026). A user measures only 1.1 times against OrbStack for builds.

Does Kiri need a licence on the Mac the way Docker Desktop does?+

No. Kiri sits under Apache-2.0, with only the name and the logo excluded. According to Docker's licence page, Docker Desktop is free for companies with fewer than 250 employees and less than 10 million US dollars in annual revenue, and needs a paid subscription above that.

Is Kiri available for Linux and Windows?+

Yes, both. On Linux the daemon runs straight on the host without a VM, on Windows inside WSL 2. The vendor's Linux measurement ran in a WSL 2 distribution on a Windows PC, so it is not bare-metal Linux; a measurement of our own on a real Linux host is still missing.

Can Kiri be used in a company environment?+

For development on a laptop, yes. The sharing feature should stay switched off there, because requests run through vendor servers and a GitHub account is required.