Convert Parquet to JSONL — Free, Private, In-Browser

Parquet to JSONL is the conversion you reach for when the next system consumes records line by line rather than a single document. Each Parquet row becomes one JSON object on its own line — newline-delimited JSON, sometimes called NDJSON — with no enclosing array and no commas between records. That framing is the whole point: you can append to the file, split it across workers, or stream it a line at a time without ever holding the full result in memory. It's the native shape for log shippers, Kafka producers, Elasticsearch bulk loads, BigQuery and Snowflake newline-delimited loaders, and any tool that reads one record, processes it, and moves on.

Unlike a flattening to CSV, JSONL keeps your Parquet's shape intact. STRUCT columns serialize to nested JSON objects and LIST columns to JSON arrays, so a column of addresses or an array of tags round-trips as real structure rather than a stringified blob. Scalar types map to their JSON equivalents: INTEGER and DOUBLE become bare numbers, BOOLEAN becomes true/false, VARCHAR becomes a string, and NULLs become JSON null. Because every line is a complete, self-contained object, each one is independently parseable — a malformed record doesn't poison the rest of the file, which is exactly what you want feeding an ingestion pipeline that commits row by row.

A few types don't have a JSON home. DATE, TIME, and TIMESTAMP have no native representation, so they serialize to strings — ISO-8601 text you'll likely need to cast back on the way in. Big integers are the classic trap: JSONL emits them as bare numbers, but many parsers (JavaScript's JSON.parse foremost) read JSON numbers as IEEE doubles and silently lose precision above 2^53, so large BIGINT IDs are safer treated as strings downstream. DECIMALs and BLOBs also need a plan, since JSON has neither. All of it runs locally in your browser via DuckDB-WASM — your Parquet file is never uploaded.