polnor 0.4.0

Polnor Python SDK — query Iceberg lakehouses, manage notebooks/jobs/models, MLflow-compatible tracking.

Polnor Python SDK

MLflow-compatible tracking for jobs running on the Polnor lakehouse.

Install

pip install polnor

Or, from a local checkout (contributors / pre-release):

pip install -e ./sdk/python

Configure

Set three environment variables before importing the SDK:

Variable	Required	Notes
POLNOR_API_URL	yes	e.g. https://api.polnor.net
POLNOR_TOKEN	yes	Personal access token from /settings/tokens
POLNOR_EXPERIMENT_ID	optional	Default experiment for start_run()
POLNOR_WORKSPACE_SLUG	optional	When the token isn't workspace-scoped

Inside a Polnor job these are injected automatically.

Quickstart

from polnor import mlflow

with mlflow.start_run(run_name="train-resnet50") as run:
    mlflow.log_params({"lr": 0.01, "batch_size": 32, "optimizer": "adam"})

    for epoch in range(10):
        loss, acc = train_one_epoch()
        mlflow.log_metrics({"loss": loss, "accuracy": acc}, step=epoch)

    mlflow.set_tag("git_commit", os.environ.get("CI_COMMIT_SHA", "dev"))

The run handle is also usable without the with block:

run = mlflow.start_run()
try:
    mlflow.log_metric("score", 0.93)
finally:
    mlflow.end_run("completed")

What's supported

API	Status
start_run / end_run / active_run	✅
log_metric / log_metrics (latest value wins)	✅
log_param / log_params	✅
set_tag / set_tags	✅
autolog() context manager	✅
Per-step metric history (time series)	⏳ planned
log_artifact (presigned-URL upload)	⏳ planned
Sklearn / PyTorch / TF auto-instrumentation	⏳ planned

How it differs from upstream MLflow

• Server-side, metrics merge into a single JSONB blob — the latest value per key wins. The full (step, value, timestamp) history is in scope for a follow-up migration.
• No server-side parameter immutability check yet — calling log_param("lr", 0.02) after log_param("lr", 0.01) overwrites silently. Upstream MLflow rejects with 400.
• No mlflow.search_runs() — use the REST API (GET /api/v1/experiments/:id/runs) directly until we wrap it.

Development

pip install -e ".[test]"
pytest

Tests use responses to mock HTTP — no live API needed.