telemetry: _retrieve_os() returns empty string on Amazon Linux 2023 — 35 of 43 released images report no OS

[Adityaj0]

Summary

deep_learning_container.py:_retrieve_os() cannot parse Amazon Linux's /etc/os-release. It returns "" on every Amazon Linux 2023 image, so the OS segment of the aws-dlc-autogenerated-tag-do-not-delete EC2 instance tag — and the OS dimension of DLC usage telemetry — is blank for most released DLCs.

Root cause

The function matches /etc/os-release with two regexes:

if re.match(r"^ID=\w+$", line):
    name = re.search(r"^ID=(\w+)$", line).group(1)
if re.match(r'^VERSION_ID="\d+\.\d+"$', line):
    version = re.search(r'^VERSION_ID="(\d+\.\d+)"$', line).group(1)

They assume an unquoted ID and a quoted, dotted VERSION_ID. That describes Ubuntu, not Amazon Linux.

Verbatim /etc/os-release from amazonlinux:2023 (the base under nvidia/cuda:*-amzn2023, which the DLC Dockerfiles build FROM):

NAME="Amazon Linux"
VERSION="2023"
ID="amzn"
ID_LIKE="fedora"
VERSION_ID="2023"
PLATFORM_ID="platform:al2023"

• ID="amzn" is quoted → ^ID=\w+$ does not match → name stays ""
• VERSION_ID="2023" has no dot → ^VERSION_ID="\d+\.\d+"$ does not match → version stays ""

Both captures fail independently, so the function returns "".

For contrast, ubuntu:24.04 ships ID=ubuntu (bare) and VERSION_ID="24.04" (dotted) — both regexes match, which is why this was never caught.

Observed effect

Running the current _retrieve_os() against those two real files:

_retrieve_os() on amzn2023     -> ''
_retrieve_os() on ubuntu24.04  -> 'ubuntu24.04'

The tag built in tag_instance() is f"{framework}_{container_type}_{framework_version}_python{py_version}_{device}{cuda_version}_{os_version}", so on an AL2023 image it ends in a bare trailing underscore:

vllm_server_training_0.26.0_python3.12.0_gpu_cuda13.0_

instead of

vllm_server_training_0.26.0_python3.12.0_gpu_cuda13.0_amzn2023

Measured impact

Across the released image configs on main (release.release: true with a prod_image), cross-referencing each config's build.dockerfile for whether it wires in deep_learning_container.py / bash_telemetry:

	count
Unique released images	46
...that ship the telemetry script	43
...of those, running on amzn2023 → empty OS field	35 (81%)
...of those, running on Ubuntu → correct OS field	8

Affected frameworks: base, huggingface-pytorch-training, openfold3, pytorch_runtime, ray, sglang_server, sklearn, tensorflow, vllm_omni, vllm_server, xgboost.

The share is growing, not shrinking — the repo is actively migrating images to Amazon Linux 2023, and telemetry was just extended to the base images in #6499.

The repo already has the correct idiom

scripts/ci/autocurrency/detect-versions.sh:111 reads the same file the right way:

bash -c 'source /etc/os-release && echo "${ID}${VERSION_ID}"'

That yields amzn2023 / ubuntu24.04 — exactly the os_version spelling used in .github/config/image. The telemetry script is the outlier.

Reproduction

No container required:

import sys
from unittest.mock import patch
sys.path.insert(0, "scripts/docker/telemetry")
import deep_learning_container as dlc
import io

AMZN = 'NAME="Amazon Linux"\nID="amzn"\nID_LIKE="fedora"\nVERSION_ID="2023"\n'
real = open
with patch("builtins.open", lambda f, *a, **k: io.StringIO(AMZN) if f == "/etc/os-release" else real(f, *a, **k)):
    print(repr(dlc._retrieve_os()))

''

Suggested fix

Parse os-release as the key=value format it is — split on the first =, strip surrounding quotes, take ID and VERSION_ID verbatim. That is a strict superset of the current regexes, so Ubuntu output is unchanged while Amazon Linux starts reporting amzn2023.

Worth noting separately: _retrieve_os() currently calls open() with no error handling, and tag_instance() calls it before its try block — so an unreadable /etc/os-release raises and kills the tagging process before it can tag at all.

PR: #6530