Differences from pandas

pandasx is tested against pandas 3 with a parity suite that runs the same operations in both and compares the results. The behaviours below differ on purpose, or because JavaScript works differently. For the syntax mapping (operators, [], keyword arguments) see Coming from pandas.

Language-level

• Methods instead of operators and []. s.add(t), df.get('a'), df.set('c', v), df.loc(...), df.setLoc(...): JavaScript has no operator overloading or item assignment.
• Asynchronous binary I/O. Excel, Parquet, Feather, HDF5 and SQL return a Promise, because the libraries underneath are asynchronous. CSV, JSON, HTML and XML are synchronous, with …Async variants for URLs.
• Row-wise apply receives a plain object, not a Series.
• Missing values. Numbers use NaN; strings, dates, objects and the nullable dtypes (Int64, Float64, boolean) read back as null, which is what pd.NA is. There is no separate NaT object: a missing datetime is null.
• Datetime resolution is milliseconds, the resolution of JavaScript Date. Nanosecond timestamps from files are truncated to milliseconds, and pandas reads datetimes written by pandasx as datetime64[ms].
• Booleans print as true / false in console.log. Exports (CSV, HTML, XML) write True / False like pandas.

Data model

• Column labels are unique strings. Duplicate column names throw instead of being allowed. A MultiIndex column is a tuple of strings: df.columnIndex holds the tuples, df.columns the parts joined by \u001f, and flattenColumns() turns them into names like price_min.
• Row labels can be strings, numbers, booleans, dates, null or tuples (a MultiIndex).

Plotting

• SVG instead of Matplotlib. Charts are SVG strings with a fixed, colour-blind-safe visual system, not a Matplotlib figure.
• No secondary y axis. Two scales on one chart mislead readers; use subplots: true.
• At most eight series per chart. More would need repeated colours, so the call throws; select fewer columns or use subplots. Pie charts fold extra slices into "Other".

Numerics

• Weighted rolling std / var (rolling(n, { winType }).std()) slide the weights with the window, consistent with the weighted mean. pandas assigns them by absolute position.
• interpolate({ method: 'pad' }) is still accepted as an alias for ffill; pandas 3 removed it.

Files

• HDF5: writing uses the table format and does not support MultiIndex rows yet; reading supports both fixed and table.
• Parquet: categorical columns read back in pandas as strings.

Something else behaves differently? It is probably a bug: please open an issue with the pandas code and the pandasx code side by side.