Skip to main content
Pitfall

A dead DNS entry inside a prebuilt container, and why the fix isn't 'just repoint everything'

Symptom​

Building a custom Docker image on top of dustynv/faster-whisper (a prebuilt jetson-containers GPU image) fails during the pip install step for your own added dependencies, with a DNS resolution failure.

Environment​

  • Jetson Orin NX, JetPack 6.2
  • Base image: dustynv/faster-whisper (jetson-containers)
  • Adding a thin FastAPI wrapper (fastapi, uvicorn, python-multipart, soundfile) on top of the base image's existing GPU-enabled stack (faster-whisper, ctranslate2, torch)

Root Cause​

The base image has pypi.jetson-ai-lab.dev baked in as its configured pip index — a NVIDIA/jetson-ai-lab mirror used specifically to serve pre-built ARM64 CUDA wheels for packages like ctranslate2 and torch that don't have official aarch64+CUDA wheels on the normal PyPI. That mirror was returning NXDOMAIN — confirmed independently via public DNS, not a local network issue.

The trap in the obvious fix​

The instinctive fix is to just repoint the whole image's pip index at regular pypi.org and move on. Don't do that. The reason the image points at that specific mirror in the first place is that ctranslate2 and torch don't have working CUDA-enabled aarch64 wheels on ordinary PyPI — that's the entire reason a custom mirror exists. Blanket-repointing the index would "fix" the DNS error while silently swapping your GPU-accelerated STT backend for a CPU-only (or simply broken) one, with no obvious error to tell you that happened.

Fix​

Repoint only your own added, plain dependencies at regular PyPI, and leave the base image's existing CUDA-critical packages completely untouched:

# Base image already has faster-whisper, ctranslate2, torch installed
# correctly against the jetson-ai-lab mirror — do not touch those.

# Only for wrapper-specific, plain dependencies:
RUN pip install --index-url https://pypi.org/simple \
    fastapi uvicorn python-multipart soundfile

This keeps the GPU-critical install path exactly as the base image maintainers built it, and only routes your own additions through a working index.

Why this is worth documenting explicitly​

This is the same underlying category of problem as CycloneDDS version pinning (see that entry) — a dependency chain where the "obvious" fix (point everything at the working source) actively breaks the thing that made the setup work in the first place. Any time a prebuilt jetson-containers image has a custom pip index configured, assume it's there for a load-bearing reason before repointing it wholesale.

Time cost: Under an hour once diagnosed as a DNS issue rather than a package problem — the DNS confirmation itself (ruling out local network causes) was the main time sink.