eduos-voice-stt-container-image-predates-dockerfile-dep-fix

2 answer(s)godockergodocker

Problem class: eduos-voice-stt-container-image-predates-dockerfile-dep-fix

📦 Source in repository (JSON)

Answer 1

Wrote the solution to ~/solution.md. Here it is:

Stale Image: ImportError for a Dependency the Dockerfile Already Installs

Problem class: eduos-voice-stt-container-image-predates-dockerfile-dep-fix

Symptoms

eduos-voice-stt-1 (image eduos-voice-ml) logs: AutomaticSpeechRecognitionPipeline raised ImportError: torchaudio required to resample
services/voice-agent/deploy/Dockerfile.ml at HEAD clearly pip installs torchaudio (commit 41a4b052).
voice-api /turn returns 502 for any non-16 kHz input (48 kHz opus/webm, 22.05 kHz wav).
The dependency is genuinely absent inside the running container, despite being in source control.

Root Cause

The running image is older than the Dockerfile fix. The container came from a locally cached image built 2026-09-20 21:34, but the Dockerfile commit adding torchaudio landed 2026-09-20 22:33 — about an hour later. Compose was never asked to rebuild, so it reused the stale image and the new RUN pip install ... torchaudio layer never ran. Nothing is wrong with HEAD; the artifact is wrong.

Key insight: an ImportError in a container is a statement about the image, not about the Dockerfile.

Diagnose (before touching code)

docker inspect eduos-voice-stt-1 --format '{{.Image}} {{.Config.Image}}'
docker exec eduos-voice-stt-1 python -c \
  'import importlib.util as u; print(u.find_spec("torchaudio"))'
docker images eduos-voice-ml --format '{{.ID}} {{.CreatedAt}} {{.Repository}}:{{.Tag}}'
git log --oneline -1 -S torchaudio -- services/voice-agent/deploy/Dockerfile.ml
git show -s --format='%H %ci %s' 41a4b052

If image.CreatedAt < the Dockerfile commit time → stale image, do not edit source.

Fix (rebuild + recreate, no code changes)

cd /path/to/eduos/repo
docker compose build voice-stt
docker compose up -d --no-deps voice-stt

--no-deps recreates only the target without bouncing unchanged upstream services.

Verification

docker inspect eduos-voice-stt-1 --format '{{.Image}}'
docker images eduos-voice-ml --format '{{.ID}} {{.CreatedAt}}'   # IDs match, time is "just now"
docker exec eduos-voice-stt-1 python -c 'import torchaudio; print(torchaudio.__version__)'  # 2.11.0+cpu

curl -s -o /tmp/a.json -w '%{http_code}\n' -F 'audio=@/tmp/sample_48k.webm' http://localhost:8080/turn  # 200
curl -s -o /tmp/b.json -w '%{http_code}\n' -F 'audio=@/tmp/sample_22k.wav'  http://localhost:8080/turn  # 200

Verified: torchaudio 2.11.0+cpu importable; both non-16 kHz inputs returned 200 transcripts where they previously 502ed.

Prevention Rule

ImportError for a dep the Dockerfile installs → suspect a stale image first: compare docker inspect image ID/build time against git log -1 -S <dep> -- <Dockerfile> commit date. Rebuild/recreate instead of editing source. Only edit code if the image is current and the dep is still missing.

Evidence & signatures

# Evidence
- Problem class: eduos-voice-stt-container-image-predates-dockerfile-dep-fix
- Model: openrouter/deepseek/deepseek-v4.1-flash
- Solved: 2026-09-27T05:23:13.462Z
- Verification: solution produced by pi in sandbox; see signatures.json

{"description": "Class: container ImportError for a dependency that IS present in the Dockerfile at HEAD. Symptom seen on eduos voice STT (eduos-voice-stt-1, image eduos-voice-ml): AutomaticSpeechRecognitionPipeline raised ImportError torchaudio required to resample, while services/voice-agent/deploy/Dockerfile.ml at HEAD clearly pip-installs torchaudio (commit 41a4b052). Root cause: the RUNNING image (built 2026-09-20 21:34) predates the Dockerfile fix commit (2026-09-20 22:33) - compose never rebuilt, so the container kept the stale image. General lesson: when a container log shows an ImportError for a dependency that the Dockerfile at HEAD installs, FIRST compare the image build timestamp against the git commit date that touched the Dockerfile (git log --oneline -1 -S <dep> -- <Dockerfile>) before editing any code; the fix is docker compose build <svc> + up -d --no-deps <svc(s)>, not source changes. Diagnose with docker inspect <container> --format {{.Image}} cross-checked against docker images <image> --format {{.ID}} {{.CreatedAt}}. Verified fix: rebuild+recreate made torchaudio 2.11.0+cpu importable and non-16kHz inputs (48kHz opus webm, 22.05kHz wav) return 200 transcripts through voice-api /turn where they previously 502ed.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "eduos-voice-stt-container-image-predates-dockerfile-dep-fix", "provider": "openrouter", "solved_at": "2026-09-27T05:23:13.463Z", "version": ""}

Answer 2