Pandas Data Analysis taste

# Pandas Data Analysis taste of a reproducible analytics practitioner
- Keep analysis in version-controlled scripts or notebooks with outputs cleared before commit, and move reusable logic into an importable module. Confidence: 0.85
- Pin pandas, numpy and pyarrow versions in a lockfile so results do not shift silently between runs. Confidence: 0.90
- Read data with explicit `dtype`, `parse_dates` and `usecols` instead of letting pandas guess types on large files. Confidence: 0.90
- Prefer Parquet over CSV for intermediate datasets because it keeps types, compresses well and loads much faster. Confidence: 0.90
- Use the pyarrow-backed string and nullable dtypes so missing values do not quietly turn integer columns into floats. Confidence: 0.75
- Write transformations as method chains with `assign`, `pipe` and `query`, avoiding long sequences of in-place mutation. Confidence: 0.80
- Never use `inplace=True`; reassign the result so data flow stays obvious and chains remain possible. Confidence: 0.85
- Select with `.loc` and `.iloc` explicitly and avoid chained indexing that triggers `SettingWithCopyWarning`. Confidence: 0.90
- Vectorize with built-in column operations and `np.where` or `np.select`; reach for `apply` with Python functions only as a last resort. Confidence: 0.90
- Validate assumptions with assertions after each major step: row counts, unique keys, null ratios and value ranges. Confidence: 0.85
- Pass `validate="one_to_one"` or `"many_to_one"` to `merge` so silent row explosions from duplicate keys fail loudly. Confidence: 0.90
- Use `indicator=True` when joining to inspect unmatched rows before dropping them. Confidence: 0.75
- Keep timestamps timezone-aware in UTC internally and convert to local time only for presentation. Confidence: 0.85
- Use `groupby` with named aggregations so output columns get readable names without a renaming step. Confidence: 0.80
- Downcast numeric columns and convert low-cardinality strings to `category` when memory matters. Confidence: 0.75
- Switch to DuckDB or Polars once data stops fitting comfortably in memory, instead of fighting pandas with chunking hacks. Confidence: 0.70
- Set a random seed for any sampling or shuffling so results can be reproduced exactly. Confidence: 0.90
- Keep raw data immutable and write every cleaned or derived dataset to a separate, clearly named location. Confidence: 0.90
- Use snake_case column names, normalized once right after loading. Confidence: 0.80
- Chart with matplotlib or seaborn from tidy long-format frames, with labeled axes, units and a title that states the finding. Confidence: 0.75
- Write small pytest tests for cleaning functions using tiny hand-built DataFrames and `pd.testing.assert_frame_equal`. Confidence: 0.80