Three things were making the Docker test runs noisier and slower than
they needed to be:
1. -v was hardcoded in Dockerfile.test:35 CMD and in docker-compose.
test.yml's integration-test-runner command. The ci.yml change to
drop -v from the bare pytest call missed both — Docker runs use
the image's CMD, not the workflow's.
2. -n 30 was hardcoded as the xdist worker count. On a 2-vCPU CI box
that's 30 Python processes fighting over 2 cores — mostly IPC and
import-thrash overhead. -n auto adapts to the host: 2 on CI, 30
on a 30-core dev box. Same final-result throughput on the dev
box, much better on small runners.
3. pip install had --no-cache-dir and no BuildKit cache mount, so
every Docker build re-fetched ~50 packages from PyPI (~60-90s
on a cold pip cache). Adding `RUN --mount=type=cache,target=
/root/.cache/pip` (with the `# syntax=docker/dockerfile:1.7`
directive that enables it) makes subsequent builds re-use the
download cache so they only do install work, ~5s instead of
~90s. DOCKER_BUILDKIT=1 is already exported in test_docker.sh
and is the GHA default since runner image 2023, so the cache
mount is always honoured.
Verified locally: Docker build is 19s warm (was ~90s cold each
time), test run is 102s with 5287 passed / 1 skipped (the
by-design spoolbuddy importorskip) — clean output, no [gwN]
worker spam, no "created: 30/30 workers" startup line.
GHA-side per-run cold-build slowness still happens because GHA
runners are ephemeral; a follow-up using docker/build-push-action
with type=gha cache backend would persist the BuildKit cache
across CI runs but that's a bigger workflow change.
Dockerfile.test only COPYed backend/ and pyproject.toml, so the
integration test at tests/integration/test_gcode_viewer.py:63
silently pytest.skip'd in every Docker run with "gcode_viewer/
index.html not present at /app/gcode_viewer/index.html".
That was deliberate fallback behaviour for unit-test environments
where the assets are intentionally absent, but in CI it meant the
#1218 packaging regression (3D Preview returning {"detail":"Not
Found"} because the embedded PrettyGCode viewer wasn't bundled into
the prod image) had no test guarding against a recurrence — the
test that was supposed to catch it was the one being skipped.
Add COPY gcode_viewer/ ./gcode_viewer/ to the backend-test stage,
matching the path the production Dockerfile uses (static_dir.parent
/ "gcode_viewer" = /app/gcode_viewer/) so the assertion runs against
the same layout the app sees at runtime. Path-anchored comment in
the Dockerfile so a future maintainer doesn't strip the COPY as
unused.
Test Summary:
- Build tests: 3 passed (image build, backend imports, static files)
- Backend unit tests: 378 passed (9 docker tests excluded)
- Frontend unit tests: 137 passed
- Integration tests: 9 passed (health, API endpoints, persistence, WebSocket)
Changes made to fix the Docker test suite:
1. Added curl to the production Dockerfile for integration tests
2. Removed deprecated version attribute from docker-compose.test.yml
3. Added --pull flag to all build commands to ensure fresh images
4. Added explicit build step before starting integration container
5. Fixed WebSocket test to accept 200 as a valid response
6. Excluded docker-marked tests from backend unit test runs (-m "not docker")
maziggy