Browser Parquet Inspector & Size Calculator
Inspect local .parquet schemas, row groups, compression codecs, and preview rows right in your browser—or estimate Parquet storage savings versus raw CSV and NDJSON.
Drop a local .parquet file here, or click to browse
Supports Uncompressed, Snappy, ZSTD, Gzip, Brotli, and LZ4 codecs via hyparquet.
How Apache Parquet Binary Structure & Compression Work
Apache Parquet is a columnar binary storage format engineered for analytical workloads across DuckDB, Spark, Snowflake, BigQuery, and AWS Athena. Unlike row-oriented formats such as CSV or NDJSON—where every record stores all fields sequentially—Parquet groups values by column into Row Groups and Data Pages. Because values within a single column share the same physical data type and often have low cardinality or monotonic ordering, encoders like Dictionary Encoding, Run-Length Encoding (RLE), and Delta Encoding dramatically shrink data even before block compression is applied.
When you drop a local .parquet file into DevPayload, the browser reads the file via File.arrayBuffer() and uses hyparquet to jump directly to the trailing 8 bytes of the file. Those final bytes contain the 4-byte footer length (little-endian int32) and the 4-byte magic string "PAR1". Reading the Thrift-encoded FileMetaData footer allows instant inspection of schema types, row counts, null statistics, and per-column compressed versus uncompressed byte sizes without scanning the entire dataset.
When decompressing preview rows, hyparquet-compressors decodes Snappy, ZSTD, Gzip, Brotli, and LZ4 pages strictly inside client-side memory. Use the Storage & Compression Estimator tab to model projected data lake storage footprints before provisioning S3 buckets or partitioning tables.
Parquet Binary Footer Layout & Compression Ratio
Every valid Apache Parquet file starts and ends with the 4-byte magic header PAR1. Readers seek to the end of the file to read the Thrift-encoded footer metadata in O(1) I/O.
[4B "PAR1"] [Row Group 0: Column Chunks & Data Pages] ... [FileMetaData (Thrift)] [4B FooterLen (LE)] [4B "PAR1"]
Compression Ratio = Total Uncompressed Column Bytes / Total Compressed Column Bytes
Storage Savings % = ((Raw_CSV_Bytes - Parquet_Bytes) / Raw_CSV_Bytes) * 100 Code Equivalents (Python & JavaScript / Node.js)
import pyarrow.parquet as pq
pf = pq.ParquetFile("sample_api_telemetry.parquet")
meta = pf.metadata
print(f"Rows: {meta.num_rows}, Row Groups: {meta.num_row_groups}")
for i in range(meta.num_columns):
col = meta.row_group(0).column(i)
ratio = col.total_uncompressed_size / max(col.total_compressed_size, 1)
print(f"{col.path_in_schema}: {col.compression} ({ratio:.2f}x)")
# Preview first 100 rows as Pandas DataFrame
df = pf.read_row_group(0).to_pandas().head(100)
print(df.head()) import { parquetMetadata, parquetReadObjects } from 'hyparquet';
import { compressors } from 'hyparquet-compressors';
import fs from 'node:fs/promises';
const buf = (await fs.readFile('sample_api_telemetry.parquet')).buffer;
const meta = parquetMetadata(buf);
console.log('Total rows:', Number(meta.num_rows));
console.log('Row groups:', meta.row_groups.length);
const rows = await parquetReadObjects({
file: buf,
compressors,
rowStart: 0,
rowEnd: 100,
});
console.table(rows.slice(0, 5)); Frequently Asked Questions
Does this Parquet viewer upload my file to a server? +
No. Your .parquet file is read locally via the browser File API (file.arrayBuffer()) and parsed entirely in JavaScript using hyparquet and hyparquet-compressors. Zero bytes leave your machine.
Which Parquet compression codecs are supported in the browser? +
DevPayload supports Uncompressed, Snappy, ZSTD, Gzip, Brotli, and LZ4/LZ4_RAW compressed Parquet files directly in the browser.
Why is Parquet typically 60%–85% smaller than CSV or JSON? +
Parquet stores homogeneous column values contiguously, enabling dictionary encoding, bit-packing, and run-length encoding before applying ZSTD or Snappy compression—eliminating repeated JSON field keys and verbose text formatting.
Why does the data preview show the first 100 rows? +
Limiting the interactive table preview to the first 100 rows keeps DOM rendering instantaneous even on multi-hundred-megabyte Parquet files, while still letting you export that preview to CSV or JSON.