What is JSONL? JSON Lines and NDJSON explained

A JSONL file puts one complete JSON record on each line instead of wrapping all records in one array with commas between them. Each line parses as JSON on its own, but the file as a whole is not one JSON document. Here is a three-line file:

{"id": 1, "user": {"name": "Ada", "city": "London"}, "tags": ["admin", "staff"]}
{"id": 2, "user": {"name": "Linus", "city": "Helsinki"}, "tags": ["dev"]}
{"id": 3, "user": {"name": "Grace", "city": "New York"}, "tags": []}

The specification at jsonlines.org sets three requirements plus a few conventions:

• The encoding is UTF-8, without a byte order mark.
• Each line is a valid JSON value. Objects and arrays are the most common, but null or a number is allowed too; a blank line is not.
• Lines end with \n. \r\n works as well, because whitespace around a value is ignored, and a newline after the last value is recommended but not required.
• The suggested extension is .jsonl, and compressed files become .jsonl.gz or .jsonl.bz2.

NDJSON (newline-delimited JSON) is the same idea under another name, with the .ndjson extension and the media type application/x-ndjson. jsonlines.org suggests application/jsonl, which is not standardized yet. rows.page reads .jsonl, .ndjson and .ldjson the same way.

JSONL shows up wherever software writes or moves records one at a time:

• Application and server logs, one event per line as it happens.
• Machine learning datasets, such as files hosted on Hugging Face.
• OpenAI's Batch API and fine-tuning, which both take JSONL files as input.
• BigQuery, which exports and loads JSON as newline-delimited JSON.

To add a record to a JSON array on disk, you have to remove the closing ], write a comma and the new object, and close the array again. With JSONL you append one line. A reader can also parse one line at a time and let it go, instead of tracking brackets across the whole file.

Valid JSON escapes every newline inside a string as \n, so a raw newline only ever separates records. That makes line-based shell tools safe: count records, take the first 100, or split a big file into chunks of 100,000 lines.

wc -l data.jsonl
head -n 100 data.jsonl > sample.jsonl
split -l 100000 data.jsonl part_

wc -l counts newline characters, so it comes up one short when the last record has no trailing newline.

• Pretty-printed records: an object spread over several indented lines breaks the one-value-per-line rule. Write each record compactly, for example with jq -c.
• Trailing commas: deleting the outer [ and ] from a JSON array leaves lines like {"id": 1},, which are not valid JSON.
• Blank lines: an empty line is not a JSON value, so strict parsers stop there. A single newline after the last record is fine.
• Wrong encoding: the spec requires UTF-8 without a byte order mark. A file saved as UTF-16 or Windows-1252 can fail to parse, so re-save it as UTF-8.

Drop a .jsonl, .ndjson or .ldjson file on rows.page, or click Paste data and paste the lines. DuckDB-WASM reads it in your browser without uploading it, up to 1.5 GB per tab; a bigger file loads its first rows with the note "Loaded the first N rows: the file is X, more than the browser can hold." Rows load in pages of 100 as you scroll, and any row opens in a side drawer with a JSON/Fields toggle.

Nested objects become dot-path columns such as user.city, up to 8 levels deep and 200 child fields per object, and arrays stay list columns. Filter with country:NL or price:>20, or type SQL against the table data, with double quotes around column names that contain dots:

SELECT "user.city", count(*) AS users
FROM data
WHERE "user.name" IS NOT NULL
GROUP BY "user.city"
ORDER BY users DESC

Export saves the current view as CSV, Parquet (zstd), a JSON array, JSONL or a Markdown table of the first 1,000 rows. rows.page has no whole-file tree view and can't edit cells, so fix broken lines in a text editor or with jq first.