# Reading and Writing JSON, CSV and Parquet in Python

> Which file format to use where, how to read each one without surprises, and why Parquet is the default for analytics.

- Source: https://akshaythakare.com/notes/python/files-json-csv-parquet
- Topic: Python
- Level: beginner
- Reading time: 2 min
- Tags: python, file-formats, parquet, interview
- Updated: 2026-10-11

## Three formats, three jobs

| Format | Shape | Schema | Best for |
| --- | --- | --- | --- |
| CSV | Rows of text | None — everything is a string | Exchange with spreadsheets and legacy systems |
| JSON / JSONL | Nested documents | Implicit, per record | APIs, logs, raw landing zones |
| Parquet | Columnar, compressed | Embedded and typed | Analytics, data lakes, warehouses |

## CSV: always state the types

```python
import csv
from decimal import Decimal

with open("orders.csv", newline="", encoding="utf-8") as f:
    for row in csv.DictReader(f):
        amount = Decimal(row["amount"])   # never float for money
        qty = int(row["qty"])
```

CSV has no types, no standard for nulls and many quoting dialects. Always parse explicitly and decide what an empty string means.

## JSON Lines: one record per line

```python
import json

with open("events.jsonl") as f:
    for line in f:                       # streams: memory stays flat
        event = json.loads(line)
```

JSONL appends and streams well, so it suits raw landing zones. A single large JSON array must be loaded whole.

## Parquet: columns, compression, types

```python
import pyarrow.parquet as pq

table = pq.read_table("sales.parquet", columns=["store_id", "net_sales"])  # reads 2 columns only
```

Parquet stores each column together. A query that needs 2 of 40 columns reads about 5% of the bytes, and min/max statistics let engines skip whole row groups.

> **Tip: Interview framing**
>
> "Raw lands as JSONL because it is append-friendly and keeps the source shape; everything after cleaning is Parquet because analytics reads few columns from many rows."

## Common mistakes

- Using `float` for money. Use `Decimal` or integer cents.
- Opening CSVs without `newline=""`, which mangles embedded line breaks.
- Writing thousands of tiny Parquet files. Aim for roughly 128 MB–1 GB per file.