CSV, JSON, Avro, Parquet — When to Use Which

Four formats, four trade-offs

Format	Layout	Schema	Best for
CSV	row, text	none	One-off exchange, smallish files, human inspection
JSON	row, text	self-desc	API payloads, nested records, log lines
Avro	row, bin	embedded	Streaming + schema evolution (Kafka)
Parquet	column, bin	embedded	Analytics — scan a few columns over many rows

Rule of thumb: OLTP / streaming → row. OLAP / analytics → column. Anything crossing trust boundaries → schema-embedded.