Owl

Owl is OWID's lightweight pipeline runner: one step folder (owl_steps/<namespace>/<dataset>/vYYYYMMDD/) holds the snapshot, transform, and metadata, and writes a normal owid.catalog dataset under data/garden/. It replaces the full snapshot → meadow → garden → grapher chain for small datasets.

Read the README first

The authoritative reference is lib/owl/owl/README.md — read it before doing anything with Owl. It covers the commands, project layout, @Snapshot/@Dataset/@Action API, snapshot capture helpers, meta.yml, loading existing ETL datasets, the Grapher action, and current limitations. Do not duplicate it from memory — open it.

The framework is small and readable; for API details beyond the README, go to the source:

When to use Owl vs. classic ETL

• Owl — small/single-source datasets, prototypes, demos, agent-assisted creation.
• Classic ETL (/create-etl-steps, /update-dataset) — mature, high-traffic, or multi-stage datasets that need real DAG integration. Owl is experimental, has limited metadata validation, and no DAG integration.

If a request doesn't clearly fit Owl, say so and recommend classic ETL instead of forcing it.

Tasks

Create a new step

.venv/bin/owl new <namespace>/<dataset> [--date YYYY-MM-DD], then fill in step.py and meta.yml, then owl snapshot → owl run. Follow the README's "Creating a new step", "Snapshots", and "Datasets" sections.

Migrate a classic ETL dataset to Owl

Collapse the snapshot → meadow → garden → grapher chain into one Owl step folder:

1. Scaffold the step (owl new, or hand-create vYYYYMMDD/), using a --date matching the source's access date.
2. Move the snapshot fetch into a @Snapshot (e.g. snap.download(URL, suffix=".csv")).
3. Collapse the meadow + garden transforms into one @Dataset — load with load_snapshot(...), harmonize, tb.format([...]).
4. Port the garden .meta.yml into the Owl meta.yml: the datasets.<name> block accepts the standard catalog shape (definitions, tables → variables), so most metadata moves over verbatim. Snapshot origin goes under snapshots.<name>.origin.
5. If it was a grapher step, add an @Action(kind="grapher", default=False) that calls upsert_dataset(...).
6. Remove the classic files (snapshot .py/.dvc, meadow, garden + meta, grapher) and their active DAG entries.
7. owl snapshot + owl run, then verify the output dataset matches the old one.

Run / debug

owl run <pattern> (regex; rebuilds stale steps + upstream deps — no --force needed after edits), owl run <pattern> --grapher for Grapher actions, owl run <pattern> --force to rebuild regardless, owl viz <pattern> to render the dependency DAG. owl snapshot is the only way to fetch data — run never fetches.

Recommended step.py pattern (real datasets)

For anything beyond a toy dataset, use the catalog-Table pattern (preserves origins/metadata properly), rather than the lightweight inline-DatasetMeta scaffold:

from pathlib import Path

from owid.catalog import Table
from owl import Action, Dataset, Snapshot, SnapshotCapture
from owl.catalog import load_snapshot
from owl.grapher import upsert_dataset

from etl.data_helpers import geo

URL_DOWNLOAD = "https://example.com/data.csv"
COUNTRIES_FILE = Path(__file__).with_name("my_dataset.countries.json")


@Snapshot
def raw_data(snap: SnapshotCapture) -> None:
    snap.download(URL_DOWNLOAD, suffix=".csv")


@Dataset
def my_dataset(raw_data: Snapshot) -> Table:
    tb = load_snapshot(raw_data)                       # Table with origins from meta.yml `origin`
    tb = tb.rename(columns={"country_name": "country"})
    tb = geo.harmonize_countries(df=tb, countries_file=COUNTRIES_FILE)
    tb = tb.format(["country", "year"])
    return tb


@Action(kind="grapher", default=False)
def upsert_to_grapher(my_dataset: Dataset) -> None:
    upsert_dataset(my_dataset)

Dependencies wire by parameter name (a raw_data parameter resolves to the @Snapshot/@Dataset named raw_data in the module).

Repo conventions that still apply inside step.py

• Preserve metadata/origins in transforms: no np.where, no pd.concat/pd.to_numeric (use pr.* from owid.catalog.processing), no pd.DataFrame(tb). See the "Preserving metadata/origins" rules in CLAUDE.md.
• Run make check before committing. Don't commit, push, or open PRs unless explicitly told to.