acoustic-carpenter-78188
04/22/2026, 12:25 AMflyte run python-script end-to-end against a Union Cloud tenant surfaced five independent issues. This PR fixes all five. None of the changes alter the public API surface; they are all internal CLI / task runner fixes.
Changes
1. --image <uri> was unusable
cli/_run_python_script.py routed --image through flyte.Image.from_ref_name(name), which only looks up named images registered in the config's image map. Passing a raw URI (the documented usage: "Container image URI") raised Image name '...' not found in config. Available: [].
Fix: branch on whether the value looks like a URI (contains / or :) and use Image.from_base(image_uri) in that case, falling back to from_ref_name for bare names.
2. No way to bundle anything besides the script itself
_run_python_script.py called the runner with _bundle_relative_paths=(script.name,). That's fine for self-contained scripts, but any script that imported a sibling helper or read a local config file failed in the pod (ModuleNotFoundError, FileNotFoundError) with no obvious escape hatch — copy_style="custom" was hard-coded and the user had no lever to widen the bundle.
Fix: expose the existing Environment.include mechanism on the python-script command as --include-files (repeatable, accepts paths or globs). Relative entries are anchored at the script's directory via env._declaring_file = str(script) (the default stack-walk lands on CLI internals, which would resolve globs against the wrong directory). Absolute paths pass through unchanged.
flyte run python-script train.py \
--include-files "*.py" \
--include-files "configs/settings.yaml"
Same param is available in the Python API as include_files=[...]. Default behavior is unchanged — still just the script — so this is purely additive.
3. Every script registered under the same task identity
The task name is computed as env.name + "." + func.__name__ (_task_environment.py:319). Since the env name was hardcoded "python_script" and the function name is always execute_script, every script ended up registered as python_script.execute_script. In the UI this caused the same run IDs to appear under every script-stem grouping (the UI groups runs by short_name, but pulls the run list from the shared task identity).
Fix: make env name script-specific: f"python_script_{script.stem}". Each script now has a distinct backend task, and the UI groups runs correctly per script.
4. Default resources too heavy for a "hello world"-positioned feature
Defaults were --cpu 4 --memory 16Gi, which do not fit on common smaller node shapes (e.g. t3a.xlarge) — users would see their first flyte run python-script hello.py sit in Pending indefinitely. Users who need more can still override.
Fix: --cpu 1 --memory 2Gi.
5. No live log streaming
The task used subprocess.run(stdout=tempfile, stderr=tempfile) inside the pod, capturing all output to tempfiles. Only the last 1000 bytes were returned as PythonScriptOutput.stdout after exit. Users could not observe a running script via the logs tab / kubectl logs — the script's output appeared nowhere until completion.
Fix: switch to subprocess.Popen with stdout=PIPE, stderr=STDOUT, iterate lines as they arrive, echo each to sys.stdout (which the pod emits to its stdout stream → k8s → UI logs tab), and keep a collections.deque(maxlen=80) tail for the structured output. Also pass python -u for unbuffered interpreter so print() flushes line-by-line.
flyteorg/flyte-sdkacoustic-carpenter-78188
04/22/2026, 2:26 AM