Data Scientist: Hybrid-Engine Data Processing

Answer data questions through the cheapest engine and surface that can prove the answer, and decide where the computation should live before touching the data.

Execution surfaces: resident kernel first

A persistent REPL/eval kernel (many harnesses expose one for JavaScript and Python) is the default surface. Reason: each one-shot process pays roughly a second of spawn-plus-import overhead and re-scans the input file, while a resident connection amortizes both — after a one-time load, repeat queries return in milliseconds. Exploration is repeat queries, so this difference dominates the session.

1. JavaScript kernel (Bun): run scripts/ensure-js-deps.sh once; it prints the absolute import path for @duckdb/node-api. Dynamic-import it, connect once, query across cells.
2. Python kernel: the default surface for Python work. duckdb/numpy/matplotlib are typically resident; Polars and pyarrow come from scripts/ensure-py-deps.sh, which installs them once into a user cache keyed to the kernel's interpreter — sys.path.insert the printed directory and import. The interpreter itself is never mutated.
3. uv lane (uv run --with ...): isolation for a heavy or crash-prone one-shot that should not take the kernel down.
4. No kernel (plain-shell harness): the same engines as one-shots — bun -e for DuckDB-js, uv run python -c for the Python stack — batching several questions per process.

Per-surface patterns and pitfalls: read references/execution-surfaces.md before first use.

Engine selection

• DuckDB for SQL-shaped work: direct file queries, joins, aggregation, subqueries, window functions. It queries CSV/Parquet/JSON in place without loading, spills to disk past its memory limit, and reads remote files with the same syntax.
• Polars when the pipeline is DataFrame-shaped: expression-chain transforms, reshapes, streaming datasets past RAM — resident in the Python kernel via ensure-py-deps.sh. Read references/polars-lane.md — the current 1.x API differs from widely-memorized older spellings.
• numpy when numeric work goes beyond SQL/DataFrame aggregation: statistical tests, linear algebra, FFT, random sampling.
• matplotlib for every chart — read references/visualization.md first; it carries the quality bar and a mandatory visual check.

Performance folklore ("X is Nx faster at filtering") varies with data shape, cardinality, and hardware. When the engine choice materially matters, measure on the actual data instead of trusting remembered multipliers.

Placement: decide where the computation lives

Probe before you compute — one cell: file size, free RAM, and (when unclear) a row count via a direct scan. Then place the work:

• Load into memory when the working set stays within roughly a quarter of free RAM AND the session will run repeated queries: CREATE TABLE t AS SELECT ... (or a collected DataFrame) once, then iterate. One scan up front converts every later query from a file re-scan into milliseconds.
• Query in place / stream when the question is single-pass, or the data exceeds RAM: DuckDB reads files directly (FROM 'data.csv'); past RAM, cap DuckDB's memory and let it spill, or use Polars' streaming engine in the Python kernel. NEVER load a larger-than-RAM dataset fully into memory — swapping stalls the whole machine, while streaming merely takes longer.
• Query remotely, in place when the data lives elsewhere: DuckDB reads http(s)/S3 Parquet and CSV with projection and predicate pushdown, so fetch the columns and rows the question needs, never the whole file. When data sits on another machine you can execute on, ship the query to the data and return the small result. Rule: result much smaller than data — move the query; repeated local iteration planned — move a pruned copy of the data once.

Sizing heuristics and recipes: references/placement.md.

Hard rules

• NEVER use pandas. DuckDB and Polars beat it decisively on every workload this skill covers, and the environments this skill assumes do not ship it — .df() on a DuckDB result raises unless pandas is installed; convert with .pl() via Arrow instead.
• Excel files are not read directly: export to CSV or Parquet first.

Output contract

Answer the question; report row counts and timing for anything heavy; then stop — no bonus

…