Intro

# ===== The values =====
# NumPy: typed n-dimensional arrays + vectorised math
# Pandas: labelled tables (DataFrame) + time series + groupby + I/O

import numpy as np
import pandas as pd

# ===== NumPy basics =====
a = np.array([1, 2, 3, 4])
b = np.arange(0, 10, 2)
m = np.zeros((3, 4))
print(a.shape, a.dtype)
print(a + 10)        # vectorised
print(a.mean(), a.std())

# Broadcasting:
x = np.array([[1, 2, 3], [4, 5, 6]])
y = np.array([10, 20, 30])
print(x + y)

# ===== Pandas basics =====
df = pd.DataFrame({
    'name': ['Alex', 'Sam', 'Lee'],
    'age':  [30, 25, 40],
    'city': ['Sydney', 'Sydney', 'Melbourne'],
})
print(df)

print(df[df.age > 28])
print(df.groupby('city').age.mean())

df['decade'] = (df.age // 10) * 10
print(df.sort_values('age', ascending=False))

# ===== I/O =====
df = pd.read_csv('data.csv', parse_dates=['date'])
df.to_parquet('data.parquet')

# ===== When they win =====
# - Any tabular or numeric data work in Python
# - Foundation for scikit-learn, statsmodels, matplotlib, seaborn
# - Notebooks + dataframes are the lingua franca of analysis

# ===== When they hurt =====
# - Multi-GB data on one machine (use Polars, Dask, DuckDB, Spark)
# - Heavy SQL-style work (DuckDB on parquet often beats Pandas now)

# ===== Patterns to internalise =====
# - Vectorise: no Python for loops over rows
# - Specify dtypes on read; lets Pandas skip inference
# - Chain operations; avoid intermediate variables that hide intent
# - Use parquet over CSV for repeated reads

# ===== Pitfalls =====
# - SettingWithCopyWarning: chain into .loc or use copy()
# - apply with a Python function on a hot column -> 10-100x slower than vectorised
# - Memory blowups on large groupby + agg; check chunk size
# - Inconsistent NaN handling (np.nan, pd.NA, None) across versions

# NumPy: fast n-dimensional arrays + linear algebra.
# Pandas: tabular data (Series + DataFrame) built on NumPy.