Fix Up Command Foreground Interactions when sending signals#127
Merged
Conversation
Foreground `up` (without --detach) parked in waitForever() with no signal
handling and never noticed when its containers stopped. Combined with the
ContainerCommands signal machinery leaving SIGINT/SIGTERM neutered (SIG_IGN)
after image-pull/network/build calls return, Ctrl-C did nothing and the
process hung even after `container compose down` from another shell.
Replace waitForever() with runForegroundUntilStopped(), matching docker compose:
- SIGINT/SIGTERM (observed via a DispatchSource stream, so they fire despite
the SIG_IGN disposition) gracefully stop the project's containers, then exit
- a second signal force-kills the containers with SIGKILL, then exits
("press Ctrl+C again to force")
- a monitor task exits `up` once all services have run and then stopped —
whether they exit on their own or are stopped via `container compose down`
from another shell (it waits until each container is seen running first, so
it never returns before startup completes)
Also retarget the Mcrich23#27 busy-loop regression test at the new wait function and
make it reliable in the full parallel suite: getrusage(RUSAGE_SELF) is
process-wide, so a 1s window (vs 200ms) lets a real busy-loop's full-core
~1,000,000 µs dominate the transient CPU noise from other parallel suites.
Cyb3rDudu
requested changes
Jul 4, 2026
Cyb3rDudu
left a comment
Cyb3rDudu
left a comment
Collaborator
There was a problem hiding this comment.
nice one, the dispatchsource signal handling is the right way around the
SIG_IGN that containerapiservice leaves set, and the two-tap graceful→force
- auto-exit-when-stopped monitor are great.
one change before merging: runForegroundUntilStopped rebuilds names as
<project>-<service>, so it misses explicit container_name and the dotted
<service>.<dnsDomain> names from #97. for those, stopContainers /
killContainers / the monitor do client.get(id: "<project>-<service>") →
nil → skip, so ctrl-c exits container-compose but leaves the containers
running. pass the resolved names instead of service keys —
services.map { containerName(for: $0.serviceName) } — or call
containerName(for:) inside.
minor (non-blocking): the monitor needs to see each container running once
before it'll return, so a fast one-shot that exits before the monitor starts
never satisfies it. the signal path still exits fine.
Collaborator
…to fix/compose-up-foreground # Conflicts: # Tests/Container-Compose-StaticTests/WaitForeverCpuTests.swift
…ice> runForegroundUntilStopped rebuilt names as <project>-<service>, so containers named via explicit container_name or the dotted <service>.<dnsDomain> DNS convention (Mcrich23#97) were never found by the stop/kill/monitor paths — ctrl-c exited container-compose but left those containers running. Pass names resolved through containerName(for:) instead of rebuilding them from service keys. Also port the Mcrich23#126 idle-baseline technique into ForegroundWaitCpuTests (successor of the deleted WaitForeverCpuTests) so the process-wide getrusage flake fixed there isn't reintroduced: assert on incremental CPU over an immediately-preceding idle window rather than an absolute threshold.
Contributor
Author
Cyb3rDudu
approved these changes
Jul 8, 2026
This was referenced Jul 12, 2026
Cyb3rDudu
added a commit
that referenced
this pull request
Jul 12, 2026
) `container-compose up` without `-d` started the first container, then exited 1 with "Error: no runtime client exists: container is stopped", leaving nothing running. Detached `up -d` was unaffected. Root cause: in foreground mode the attached `container run` subprocess runs in a background Task while `waitUntilServiceStarted` polls concurrently. Apple Container reports a container as `.stopped` throughout the image/kernel fetch and only flips to `.running` once init starts, so an early poll caught that pre-running `.stopped`, treated it as a completed one-shot, and threw via `containerWait` (which needs the runtime client, gone once stopped). Fixes: - Track the attached `container run` subprocess with a `ForegroundRunHandle`. In `waitUntilServiceStarted`, only treat `.stopped` as terminal once that subprocess has actually exited; while it is alive, keep polling for `.running`. Detached mode (handle is nil) keeps its existing behaviour. - Seed `waitUntilAllContainersStopped`'s `seenRunning` with all container names. It runs only after `configService` has already confirmed every container started, so a `.stopped` container there is one that exited, not one that hasn't started. Without this, a fast one-shot that exits before the first poll would never be seen `.running` and foreground `up` would hang. This also makes #127's signal handling reachable: previously `configService` crashed before `runForegroundUntilStopped` could install the SIGINT/SIGTERM handler, so Ctrl-C never gracefully stopped the project. Fixes #131
Cyb3rDudu
pushed a commit
that referenced
this pull request
Jul 22, 2026
* Add shared ComposeProjectOptions for compose-file/service resolution Introduce a reusable ComposeProjectOptions (compose-file location + decode, project-name derivation, dependency-ordered service resolution, container-name mapping) and a resolved ComposeProject value. New subcommands adopt this instead of re-duplicating the boilerplate that up/down/build/logs each grew. * test(project): cover ComposeProjectOptions file discovery, naming, and ordering Covers composePath (--file precedence, discovery order, default), projectName fallback to the derived cwd name, and orderedServices topological sort/filtering. * feat(project): resolve services to candidate container names, not one guess ComposeProject mapped each service to a single container name (container_name ?? <project>-<service>), which misses the dotted <service>.<dnsDomain> names 'up' uses when the project's DNS domain is registered (#97) — the same bug class fixed for ctrl-c in #127. Replace ServiceTarget.containerName with candidateContainerNames (explicit name first, then both conventions, mirroring ComposeDown.stopOldStuff) and add existingContainer(using:) to resolve which candidate actually exists, so every subcommand built on ComposeProject shares one correct lookup instead of re-deriving names. * feat(project): delegate service selection to Service.selectServices orderedServices hand-rolled its own name filter, which ignored Compose profiles gating and didn't expand depends_on — diverging from the shared selection algorithm up/build/down adopted in #126. Route it through Service.selectServices with the ComposeFileOptions active profiles so commands built on ComposeProject select identically to the existing ones: profile-eligible services plus dependencies by default, requested services plus transitive dependencies (bypassing the profile gate) when named explicitly. * refactor(down): resolve project through shared ComposeProjectOptions First consumer of ComposeProject: drop ComposeDown's own copies of compose-file discovery, project-name derivation, topo-sort/selection, and candidate-container-name construction in favor of the shared API. Behavior is preserved (profile skip note, stop-all-candidates loop for mixed naming-mode leftovers, extra_hosts cleanup) with one refinement: explicit 'down <service>' now expands dependencies via Service.selectServices instead of the topo-sort dependedBy side effect, which only recorded each service's first visitor and so stopped an explicit request's dependencies only when visit order happened to align. * refactor(build): resolve project through shared ComposeProjectOptions Drop ComposeBuild's copies of compose-file discovery, project-name derivation, and service selection in favor of ComposeProjectOptions. Selection semantics are unchanged (Service.selectServices with active profiles, then keep services with a build config); ordering improves from dictionary order to topological, so dependencies build before dependents. * refactor(up): resolve project through shared ComposeProjectOptions Drop ComposeUp's copies of compose-file discovery/decoding, project-name derivation, topo-sort/selection, and the hand-built stop-candidates list in favor of ComposeProjectOptions and ComposeProject.candidateContainerNames. Behavior is unchanged: 'up' keeps deciding the single creation-time name itself (explicit container_name, dotted DNS name when the domain is registered, else <project>-<service>) since that depends on runtime DNS availability, and the pre-start cleanup still stops every name shape a previous run might have used. * refactor(project): simplify shared resolution layer and its consumers Cleanups from a reuse/simplification/efficiency/altitude review of the branch: - down: resolve once instead of loadCompose()+resolve() (the compose file was read and YAML-decoded twice per run); drop the stop loop's remove parameter, which was always false - up: consume resolve() instead of re-composing loadCompose + projectName + orderedServices + candidate names piecewise; this also gives up the same 'No such service' warning down/build already had - move the 'No such service' warning from resolve() into orderedServices() so every selection path warns, not just resolve() - hoist the envFilePath derivation duplicated verbatim in up and build onto ComposeProjectOptions - move sanitizeDnsDomain from ComposeUp to ComposeProject: it is project naming policy, and the shared layer should not reach into a command for it - delete dead code: ServiceTarget.existingContainer (no callers, and its first-match semantics is the exact behavior down's stop-every- candidate loop exists to avoid), ComposeProject.cwd (never read), ComposeError.invalidProjectName (no throwers left), and ComposeUp's vacuous projectName optionality (guards that could never fire once resolution guarantees a name) - narrow supportedComposeFilenames to private * fix(up): resolve healthcheck container name instead of assuming <project>-<service> waitUntilServiceIsHealthy rebuilt the container name by hand, so healthchecks for services created under an explicit container_name or the dotted <service>.<dnsDomain> DNS-mode name (#97) exec'd against a nonexistent container and burned every retry before failing. Use containerName(for:), which reads the resolved name configService records before any healthcheck runs — the same fix #127 applied to the foreground stop paths. * test: cover loadCompose missing-file throw and healthcheck name resolution Closes the two meaningful coverage gaps over this branch's diff: the composeFileNotFound path in ComposeProjectOptions.loadCompose, and a dynamic regression test proving healthchecks exec against the resolved container name — a service created under an explicit container_name previously health-checked the nonexistent <project>-<service> and failed up after exhausting retries.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
When hitting ctrl+c while running the up command in the foreground, I noticed unexpected (to me) behavior. I was expecting it to behave like docker compose.
This allows for proper signal listening. SIGINT (first ctrl+c) gracefully kills the stack. SIGKILL (second ctrl+c) force kills the stack.