Pandas 3 0 3

pandas 3.0.3

Overview

Pandas is the foundational Python library for tabular data manipulation and analysis. Version 3.0.3 introduces major semantic shifts from earlier versions: dedicated string dtype by default, consistent Copy-on-Write behavior (no more SettingWithCopyWarning), pd.col() expression syntax, Arrow PyCapsule Interface support, and anti-joins in merge operations.

Core Objects

• DataFrame — 2D labeled table with heterogeneous columns
• Series — 1D labeled array (single column of a DataFrame)
• Index — immutable label axis (includes DatetimeIndex, MultiIndex, CategoricalIndex, etc.)

Key v3.0 Changes

1. String dtype by default — pd.Series(["a", "b"]) now yields dtype: str instead of object
2. Copy-on-Write — all indexing returns views that behave as copies; chained assignment no longer works; SettingWithCopyWarning removed
3. pd.col() expressions — column references without lambdas in assign, loc, etc.
4. Arrow PyCapsule Interface — zero-copy data exchange with Arrow-compatible libraries via from_arrow() and __arrow_c_stream__
5. Anti-joins — merge(how="left_anti") and merge(how="right_anti") now supported

Dependencies

• Required: numpy>=1.26.0, python-dateutil>=2.8.2, tzdata (Windows/Pyodide)
• Python versions: 3.11, 3.12, 3.13, 3.14
• Optional: pyarrow>=13.0.0 (Parquet, Feather, ORC, Arrow-backed dtypes), bottleneck, numexpr, scipy, xarray, fsspec, s3fs, gcsfs

Usage

import pandas as pd
import numpy as np

# Create a DataFrame
df = pd.DataFrame({
    "name": ["Alice", "Bob", "Charlie"],
    "age": [30, 25, 35],
    "score": [85.5, 92.0, 78.3]
})

# String dtype is now default (v3.0)
df["name"].dtype  # str (not object)

# Read data
df = pd.read_csv("data.csv")
df = pd.read_parquet("data.parquet")

# Indexing — loc (label-based), iloc (position-based)
row = df.loc[0]
first_two = df.iloc[:2]

# Boolean filtering
filtered = df[df["age"] > 25]

# Groupby aggregation
result = df.groupby("category").agg(
    avg_score=("score", "mean"),
    total=("score", "sum")
)

# Merge / join
merged = pd.merge(df1, df2, on="key", how="left")
anti = pd.merge(df1, df2, on="key", how="left_anti")  # v3.0 new

# Reshaping
pivoted = df.pivot_table(index="date", columns="category", values="value", aggfunc="sum")
melted = pd.melt(df, id_vars=["id"], value_vars=["a", "b", "c"])

# Time series
ts = pd.read_csv("data.csv", parse_dates=["date"])
ts = ts.set_index("date").sort_index()
resampled = ts.resample("M").sum()

# Rolling windows
rolling_avg = df["value"].rolling(window=7).mean()

# Export
df.to_csv("output.csv", index=False)
df.to_parquet("output.parquet")

pd.col() Expression Syntax (v3.0+)

Use pd.col() instead of lambdas for column references:

# Instead of:
df.assign(total=lambda d: d["a"] + d["b"])

# Write:
from pandas import col
df.assign(total=col("a") + col("b"))

# In loc (filtering):
df.loc[col("age") > 25]

# Chained expressions:
df.assign(
    ratio=col("score") / col("max_score"),
    label=(col("ratio") > 0.8).map({True: "high", False: "low"})
)

Copy-on-Write (v3.0+)

Every indexing operation behaves as a copy. Modify the original object directly:

# v3.0 correct — modify original directly
df.loc[mask, "col"] = value

# Chained assignment does NOT work in v3.0 (no warning, just silently fails to update)
df[df["age"] > 25]["score"] = 100  # WRONG — this modifies a copy

# Use .copy() explicitly when you need an independent object
subset = df.loc[mask].copy()

Gotchas

• String dtype is default in v3.0 — code checking for dtype == "object" to find strings will break. Use is_string_dtype() or check for "str" dtype instead.
• Chained assignment silently fails — df[df["a"] > 0]["b"] = 1 modifies a copy, not the original. Always use .loc[row_mask, "col"] = value.
• No SettingWithCopyWarning — it was removed in v3.0 since Copy-on-Write makes the behavior consistent.
• pd.NA vs np.nan — the universal missing value sentinel is pd.NA for nullable dtypes. String dtype uses np.nan as its NA sentinel by default.
• mode.copy_on_write option deprecated — Copy-on-Write is always on in v3.0. The option will be removed in v4.0.
• pytz replaced by zoneinfo — time zones now use the standard library's zoneinfo module. Use pd.Timestamp("2024-01-01", tz="America/New_York") instead of pytz.timezone().
• Arrow-backed string dtype — if PyArrow is installed, string columns are Arrow-backed by default. This affects .dtype representation and some operations.
• read_csv engine defaults — the C engine is still default, but specifying engine="pyarrow" gives Arrow-native parsing with different behavior for some edge cases.
• merge anti-joins — how="left_anti" keeps rows from left that have no match in right. Useful for finding missing records.
• fillna(value=None) now works — passes the dtype-appropriate NA value instead of raising an error.

References

• [01-core-structures](references/01-core-structures.md) — DataFrame, Series, Index constructors and core attributes
• [02-data-types](references/02-data-types.md) — dtypes, nullable types, ArrowDtype, string dtype, type conversion
• [03-io-operations](references/03-io-operations.md) — CSV, Excel, Parquet, SQL, JSON, HTML, XML, pickle, clipboard
• [04-indexing-selection](references/04-indexing-selection.md) — loc, iloc, at, iat, boolean indexing, slicing, IndexSlice
• [05-groupby-aggregation](references/05-groupby-aggregation.md) — GroupBy, agg, transform, filter, apply, NamedAgg
• [06-reshaping-merging](references/06-reshaping-merging.md) — merge, concat, pivottable, melt, stack/unstack, getdummies
• [07-timeseries](references/07-timeseries.md) — Timestamp, Timedelta, date_range, resample, offset, holiday calendars
• [08-window-rolling](references/08-window-rolling.md) — Rolling, Expanding, EWM windows; first/last/nunique (v3.0)
• [09-accessors](references/09-accessors.md) — .str accessor, .dt accessor, .cat accessor, .array property
• [10-copy-on-write](references/10-copy-on-write.md) — Copy-on-Write semantics, migration from v2.x, common pitfalls
• [11-v3-changes](references/11-v3-changes.md) — Full v3.0 breaking changes, deprecations, new features, migration guide

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tangledgroup
• Source: tangledgroup/tangled-skills
• License: MIT

Install and usage instructions live in the source repository linked above.