Adds `CopyTarStream` to the job environments. It extracts directories and files into the job as they are written, kept below the destination and owned by the job user along with any directories it creates, so built-in actions can put large content in place without buffering it. The host's `GetContainerArchive` now streams instead of building the archive in memory.
---------
Co-authored-by: silverwind <[email protected]>
Co-authored-by: silverwind <[email protected]>
Co-authored-by: Lunny Xiao <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1340
Reviewed-by: Lunny Xiao <[email protected]>
Reviewed-by: silverwind <[email protected]>
Co-authored-by: bircni <[email protected]>
The Kubernetes backend copies actions and workspaces into job containers with `tar` over `pods/exec`, so every byte goes through the API server. On managed clusters that often sits in other zones than the nodes, so each byte can cross zones twice, and the copies load the API server.
With `kubernetes.transfer.mode: auto` or `direct`, an init container copies `gitea-runner` into the job pod, and the job container fetches directories and copies of 64 KiB or more gzipped from the runner over the pod network. `auto` falls back to the API server for a job pod that cannot reach the runner, `direct` fails the job.
Copies can hold secrets, so each is served once, over TLS 1.3 pinned to a certificate the runner generates, only to the job pod's IPs.
Copying setup-go@v5, cache@v4, setup-java@v4 and build-push-action@v6 sends 11.6 MB through the API server on main and 532 B with this PR.
Co-authored-by: silverwind <[email protected]>
Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1354
Reviewed-by: silverwind <[email protected]>
Co-authored-by: Lunny Xiao <[email protected]>
The Kubernetes backend copies actions and other files into the job container with `tar` over `pods/exec`, so the data goes through the API server, which on managed clusters often sits in other zones than the nodes (each copied byte then crosses zones twice). Actions are megabytes of JavaScript bundles that gzip about fivefold:
| Action | tar | gzip |
|---|---|---|
| actions/setup-go@v5 | 8.0 MB | 1.5 MB |
| actions/cache@v4 | 11.6 MB | 2.0 MB |
| actions/setup-java@v4 | 15.3 MB | 2.5 MB |
| docker/build-push-action@v6 | 12.9 MB | 3.0 MB |
The startup probe now also reports whether the job container has `gzip`. If it does, `CopyDir` and copies of 64 KiB or more are gzipped at `BestSpeed` and extracted with `gzip -dc | tar -xf -`. Smaller copies (run scripts, file command files) stay uncompressed, where gzip would save next to nothing and cost a process in the pod. Images without `gzip` keep the plain tar stream, so the image requirements do not change.
This is independent of #1354 and gets most of its traffic saving with no new moving parts.
---
*Written by Codet (AI coding agent) on behalf of @lunny.*
---------
Co-authored-by: Zettat123 <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1355
Reviewed-by: Zettat123 <[email protected]>
Contains a number of small bug fixes where jobs behaved differently from GitHub or failed silently.
1. Hooks fail on a refused `::set-env::` or `::add-path::` and no longer inherit the last step's identity
1. Docker build, exec and wait errors are returned instead of swallowed, token-only registry credentials work, logs of `--tty` services decode
1. Env files split lines like GitHub on each OS, so a carriage return is kept on Linux and CRLF ends a line on Windows, empty heredoc names or delimiters are rejected
1. Post steps run after their pre step even when main was skipped, composites see undeclared inputs, `runs.args` wins without `with.args` being parsed
1. Docker actions skip evaluating defaults of inputs given in `with` and read `entryPoint` and `args` in any case, a lowercase `dockerfile` is found, undeclared inputs get GitHub's warning
1. A user-set `CI` is kept, `GITHUB_REPOSITORY_ID` and the other missing `GITHUB_*` variables are exported
1. `timeout-minutes` can read the step's own `env`, a step-level `ACTIONS_ALLOW_UNSECURE_COMMANDS: false` beats a job-level `true` while the runner's own opt-in still wins, `pwsh` quotes the script path
1. A step `env:` given as one expression such as `${{ fromJSON(matrix.env) }}` is applied, `hashFiles` fails with GitHub's messages on script errors and after 120 seconds, a job image without node still gets an empty hash
1. A step whose command files cannot be reset fails instead of importing stale ones, a timeout between steps fails the job instead of reporting success
1. One named volume can be mounted at several paths, host-mode PTY steps keep their last output when they fail, host scratch directories are removed when job setup fails
1. Cancelling a step right after it starts kills its whole process tree
1. Cache prefix restores pick the newest entry created within the same second, also on S3, cache commits with a wrong size are rejected
1. Artifact chunks are written at their offset, download URLs are escaped, listing a run without artifacts no longer crashes
1. Poll backoff reaches `fetch_interval_max`, proxy passwords are masked in more forms, secrets are also masked in step names and in their quoted form in the runner log, a failed final report is logged
1. On Windows hosts a differently cased unsecure-commands opt-in or proxy variable counts, and an early cancel no longer leaks a job object handle
1. Interactive `register` prompts for a missing token and stops on SIGTERM, `exec --list --detect-event` no longer panics without events, `config set` works on `cache.s3` keys
Reviewed-on: https://gitea.com/gitea/runner/pulls/1338
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>
A new node can run pods before its kubelet has a serving certificate, so the first exec into the job container fails with `error dialing backend: remote error: tls: internal error` and the job fails during setup.
That exec only reads the arch and PATH, so it now retries for up to a minute unless the API server rejects it (4xx) or the command ran.
Reviewed-on: https://gitea.com/gitea/runner/pulls/1337
Reviewed-by: bircni <[email protected]>
Reviewed-by: Lunny Xiao <[email protected]>
Co-authored-by: Zettat123 <[email protected]>
On Kubernetes a service is reached by its id made a DNS label, so some service ids might become unreachable. For example, the `devstoreaccount1.azurite.local` service of Gitea's [MSSQL tests](https://github.com/go-gitea/gitea/blob/2fe74d3cdd5184a7e22e32e47772caa78d077489/.github/workflows/pull-db-tests.yml#L229) is only in the pod's hosts as `devstoreaccount1-azurite-local`, while the tests connect to the id as written. Container names must be DNS labels, but `hostAliases` accepts DNS subdomains, so an id that is one is now added there too, and to `no_proxy`, as on Docker where a service is reached by its id.
The warning for a service not reachable by its id now also covers ids that differ only in case, as Kubernetes hostnames are lowercase.
---------
Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1320
Reviewed-by: Lunny Xiao <[email protected]>
Reviewed-by: bircni <[email protected]>
Reviewed-by: silverwind <[email protected]>
Co-authored-by: Zettat123 <[email protected]>
The Kubernetes backend extracted CopyDir archives at "/" with entries like
"var/run/act/actions/<sha>/...". Ubuntu's patched tar refuses to extract
through an absolute symlink, failing with "Invalid cross-device link" on
images that link /var/run to /run, so actions never reached the job
container. Extract at the destination with relative entries instead, like
Copy already does.
This is a similar issue to #1128, but for k8s instead of Docker.
I built the image and deployed it to my k8s cluster to verify the fix.
---------
Co-authored-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1286
Reviewed-by: silverwind <[email protected]>
Co-authored-by: Drew Inglis <[email protected]>
Adds a Kubernetes backend. A `name:kubernetes://image` label runs the job as one pod through the Kubernetes API, with services as containers of the same pod. Step env is exported by a script sent over exec stdin, keeping it out of the pod spec and process arguments, and service env goes into a per-job Secret. Docker actions and container options are rejected. See `docs/kubernetes.md`.
Operators shape job pods with `kubernetes.pod_template`, with per-label `pod_templates` merged in `runs-on` order and a `$services` entry for service containers. The idle cleanup removes pods and Secrets a runner left behind.
The client calls the Kubernetes API directly and runs step commands over `coder/websocket` (same dependency used in gitea) instead of pulling in the heavyweight `kubernetes/client-go`. Includes an e2e test against k3s.
Replaces: https://gitea.com/gitea/runner/pulls/1000
Fixes: https://gitea.com/gitea/runner/issues/31
Assisted-by: Claude Code:claude-opus-5-5
Co-authored-by: Zettat123 <[email protected]>
Co-authored-by: bircni <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1260
Reviewed-by: Lunny Xiao <[email protected]>
Reviewed-by: Zettat123 <[email protected]>
Co-authored-by: silverwind <[email protected]>
bircni
Zettat123
