Job teardown ran every step on one shared minute. When removing the job container took longer, the network removal ran on an expired context and the network stayed behind until the orphan reclaim, which only takes networks older than `workdir_cleanup_age`, 24h by default. All that time it held a subnet of the daemon's address pool.
The job and service containers are now removed in parallel, and the docker proxy sweep, the network and the volumes get a fresh budget after them. The idle reclaim takes unused networks and volumes once they are 10 minutes old, and network removal treats a network that is already gone as removed.
Reviewed-on: https://gitea.com/gitea/runner/pulls/1318
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[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]>
Make cache requests go through the runner and remove outbound traffic from the cache server. The runner forwards them to `cache.external_server` when one is set, and attaches itself to each job network so jobs can reach it.
Benefits:
1. Jobs need no http proxy rules for the cache server, which `@actions/http-client` mishandles for CIDR ranges in `no_proxy` (https://github.com/actions/toolkit/pull/2495).
2. Only runners reach a shared cache server, which needs no route to Gitea.
3. A runner container sharing the host Docker socket now caches, where jobs previously could not reach it.
Breaks:
1. Jobs that reach `external_server` but not their runner, such as a remote `DOCKER_HOST`. The built-in cache has always needed that reachability, so both modes now agree. Set `cache.host` and `cache.port` to an address those jobs reach.
2. Jobs on `container.network: bridge`, where Docker rejects the alias, so the runner cannot attach and their cache fails as it does today. Leave `container.network` empty, or name a user-defined network.
*Written by Claude.*
---------
Co-authored-by: bircni <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1229
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>
Implement speedups to job start and shutdown.
- Create the job container while services are still becoming healthy, and poll their health at a flat one second instead of a 2s to 32s doubling backoff
- Pull each service image once instead of twice, and fetch a warm action cache once instead of twice
- Report the job result before reclaiming its volumes, and reap volumes stranded by a runner that died mid-job
| Step | Scenario | Before | After |
| --- | --- | --- | --- |
| Complete job | Large workspace volume | 4.2s | 0.4s |
| Set up job | One service, 2s health interval | 7.08s | 3.26s |
| Set up job | Two cached actions from github.com | 1.81s | 1.34s |
| Set up job | Two cached actions from gitea.com | 2.42s | 1.95s |
| Set up job | Two actions, cold action cache | 6.62s | unchanged |
| Set up job | Minimal job, no services or actions | 0.62s | unchanged |
Assisted-by: Claude Code:Opus 5
Reviewed-on: https://gitea.com/gitea/runner/pulls/1218
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>
Revert #1136 and use actionslib instead.
revert chore: bump the module path to `/v3`, take the version from the VCS stamp (#1136)
gitea can not consume the runner's api by version while it's version mismatches the module version:
```
go: gitea.com/gitea/[email protected]: invalid version: module contains a go.mod file,
so module path must match major version ("gitea.com/gitea/runner/v3")
```
Fix that by bumping the module version now. The existing `v3.0.0` and `v3.0.1` tags stay unusable, so a new tag is needed after this lands.
Also drop the version `-X` linker flags, which would otherwise have to repeat the new path in both `Makefile` and `.goreleaser.yaml`, where a stale path makes injection silently no-op. Go has recorded the module version in the build info since 1.24, so `Version()` reads it from there, keeping the variable as an override for builds without a VCS stamp.
That part started as https://gitea.com/gitea/runner/pulls/1137 but belongs here: the stamp resolves against the tags that are legal for the module path, so without the `/v3` bump it would report `v1.0.9-0.<ts>-<sha>`. Since `release-nightly.yml` triggers on every push to `main`, splitting them would publish a nightly with a `v1` version.
---------
Co-authored-by: bircni <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1136
Reviewed-by: techknowlogick <[email protected]>
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1148
Reviewed-by: Zettat123 <[email protected]>
gitea can not consume the runner's api by version while it's version mismatches the module version:
```
go: gitea.com/gitea/[email protected]: invalid version: module contains a go.mod file,
so module path must match major version ("gitea.com/gitea/runner/v3")
```
Fix that by bumping the module version now. The existing `v3.0.0` and `v3.0.1` tags stay unusable, so a new tag is needed after this lands.
Also drop the version `-X` linker flags, which would otherwise have to repeat the new path in both `Makefile` and `.goreleaser.yaml`, where a stale path makes injection silently no-op. Go has recorded the module version in the build info since 1.24, so `Version()` reads it from there, keeping the variable as an override for builds without a VCS stamp.
That part started as https://gitea.com/gitea/runner/pulls/1137 but belongs here: the stamp resolves against the tags that are legal for the module path, so without the `/v3` bump it would report `v1.0.9-0.<ts>-<sha>`. Since `release-nightly.yml` triggers on every push to `main`, splitting them would publish a nightly with a `v1` version.
---------
Co-authored-by: bircni <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1136
Reviewed-by: techknowlogick <[email protected]>
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>
Leaked per-job networks each hold a subnet of the daemon's address pool and nothing reclaimed them, so a host eventually fails every job with `all predefined address pools have been fully subnetted`. This is what CI hit in https://gitea.com/gitea/runner/actions/runs/742177.
- teardown no longer loses a container, and its network with it: a failed `ContainerRemove` was reported as success, a container whose id was never learned was skipped, and the daemon's own `AutoRemove` teardown was raced
- the idle cleanup reclaims what teardown cannot: networks carry `com.gitea.runner.uuid`, so a runner only touches its own, and a cutoff keeps a job starting during the pass out of scope
- pull failures reported mid-stream were discarded, surfacing later as a confusing `No such image`; they now propagate, and fall back to a local copy instead of failing the job
- `NetworkCreate` retries pool exhaustion briefly, then says which knobs to turn
- digest-pinned images are not re-pulled, and removal kills first so it never waits out Podman's stop timeout (measured 10.1s → 0.09s per container)
---------
Co-authored-by: bircni <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1124
Reviewed-by: bircni <[email protected]>
Fixes#1023.
## Problem
In Windows host mode, a single stalled delete syscall (AV/EDR filter driver, unresponsive mount, dying disk) wedged the job forever at `Cleaning up container`. `HostEnvironment.Remove()` bounds every teardown phase (`terminateRunningProcesses`, both `removePathWithRetry` calls) except the `CleanUp` callback — an unbounded `os.RemoveAll(miscpath)` assigned in `startHostEnvironment`. The runner then held its capacity slot indefinitely, the task was reaped as a zombie, and there were no diagnostics.
## Fix
- **Bound the cleanup (availability):** `Remove()` now runs `CleanUp` under `hostCleanupTimeout` (30s) via `runWithTimeout`; on timeout it logs a warning and continues job completion. The stuck goroutine is left to finish (a delete syscall can't be interrupted). Added debug logs around the phase.
- **Reclaim the leak (disk hygiene):** a timed-out cleanup can leave a scratch dir behind, so the existing idle stale-dir sweep is extended to also remove orphaned host-mode scratch dirs (16-hex names) under `Host.WorkdirParent`, leaving the shared `tool_cache` and operator data untouched. The `bind_workdir` gate is dropped from `shouldRunIdleCleanup` so host-mode runners run the sweep.
Reviewed-on: https://gitea.com/gitea/runner/pulls/1024
Reviewed-by: Lunny Xiao <[email protected]>
- Add idle-time cleanup for stale bind-workdir task directories instead of cleaning them on the task execution path.
- Make cleanup behavior configurable with `runner.startup_cleanup_age` as the stale-age threshold (default: `24h`) and `runner.idle_cleanup_interval` as the idle cleanup cadence (default: `10m`).
- Restrict cleanup scope to numeric task directory names only, to avoid touching operator-managed folders.
- Document the cleanup settings in `config.example.yaml` and `README.md`.
- Add tests for stale-directory cleanup, idle cleanup throttling, and config default/override parsing.
## Why
When a runner or host crashes, normal per-task cleanup may not run, leaving stale task directories under the bind-workdir root. Running this cleanup only while the runner is idle recovers that disk space without adding overhead to active job execution.
If you want, I can also tighten the wording around `startup_cleanup_age`, since the key name now reads a bit misleadingly relative to the actual behavior.
---------
Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/870
Reviewed-by: silverwind <[email protected]>
silverwind
Lunny Xiaoandbircni
