How to convert JSON to Parquet in your browser

rows.page opens .json, .jsonl and .ndjson files. DuckDB-WASM reads the file inside your browser tab, so nothing is uploaded unless you press Share.

1. Drop the file on rows.page, choose it from disk, or click Paste data and paste the JSON.
2. Check the grid: nested objects show up as dot-path columns like user.address.city, and arrays stay list columns that show the first 6 items and "+N".
3. Filter rows, hide columns or run SQL in the search box above the grid if you only need part of the data.
4. Open the Export menu and choose Parquet. rows.page writes the file with zstd compression.

The export follows your current view: hidden columns are left out and a filtered view writes every matching row. A SQL result exports up to 1,000,000 rows.

The Parquet file from rows.page is flat. Each nested field becomes its own column named by its path, dots included, and the export writes those columns as they are:

• {"user": {"address": {"city": "Utrecht"}}} becomes one column named user.address.city.
• "tags": ["sql", "json"] stays a single list column and exports as a Parquet list, objects inside the list included.
• Nesting below 8 levels, or an object with more than 200 fields, stays whole as one struct column.
• A text column holding JSON objects is expanded into dot-path columns when 90% or more of its first 1,000 non-empty values are valid JSON objects.
• A file that is one object wrapping a list, such as {"data": [...]}, opens as the rows of its biggest list (searched up to 4 levels deep), and a note says the other top-level fields were skipped.

Type a statement that starts with SELECT, WITH or FROM into the search box; the table is called data. Put dot-path names in double quotes, because DuckDB reads an unquoted user.address.city as a table called user and stops with an error.

SELECT
  id,
  "user.address.city" AS city,
  CAST("order.total" AS DOUBLE) AS total,
  unnest(tags) AS tag
FROM data
WHERE "order.total" IS NOT NULL
ORDER BY __rid

unnest(tags) gives one row per array item, so a record with an empty tags array drops out of the result. ORDER BY __rid uses the hidden row-number column to keep the original file order. When the result looks right, export it as Parquet.

Pick rows.page when a flat table is what you want. Use DuckDB or pandas when the Parquet file must keep its nested structure, or when the file is bigger than 1.5 GB: rows.page refuses a larger .json file and opens only the first rows of a larger JSONL file.

DuckDB

read_json reads both JSON arrays and JSONL, and COPY writes the Parquet file. Nested objects stay STRUCT columns and arrays stay LIST columns.

COPY (SELECT * FROM read_json('in.json')) TO 'out.parquet' (FORMAT parquet, COMPRESSION zstd);

A glob such as read_json('logs/*.jsonl') combines many files into one Parquet file.

pandas

pandas writes Parquet through pyarrow, so install both. Pass lines=True for JSONL and leave it out for a JSON array; nested objects become STRUCT columns here too.

import pandas as pd

pd.read_json('in.jsonl', lines=True).to_parquet('out.parquet')

Drop out.parquet back onto rows.page to check it. Parquet is not held to the 1.5 GB check because DuckDB reads it in place, and rows load in pages of 100 as you scroll.

The Columns panel shows each column's type, number of distinct values and share of nulls, plus a histogram with the value range for numbers and dates. For averages and medians, run FROM (SUMMARIZE data) in the search box. In DuckDB, DESCRIBE SELECT * FROM 'out.parquet' lists the columns and their types.