How does a Parquet viewer read columnar files without uploads?
An in-memory buffer adapter reads trailing footer bytes first to decode column definitions and row-group offsets inside a local Web Worker. Selective decompression extracts only the requested row window, letting this Parquet to CSV and Excel converter inspect binary tables in browser memory.
Unlike row-oriented formats, the Apache Parquet file format specification wraps columnar row groups between four-byte PAR1 header and footer magic markers. Our reader checks that uploaded files meet a 12 byte minimum length and verifies both 4 byte PAR1 signatures (0x50 0x41 0x52 0x31) before decoding the trailing Thrift metadata.
After validating those bytes, hyparquet reads metadata via a zero-copy slice wrapper (AsyncBuffer) inside dev.worker.ts without duplicating buffers across threads. Click Load sample Parquet for sample-telemetry.parquet (4 rows with id, service, latency_ms, and region) or inspect local SQLite databases in WebAssembly.
Summary cards show Total Rows, Schema Columns (row groups and format version), Compression codecs, and Writer metadata without network transfers.
Inspecting Parquet column schemas, INT64 types, and row groups
Every Parquet file separates nested logical presentation from compact column chunks stored on disk. The schema grid reveals five attributes per column: Column, Physical Type, Logical / Converted, Repetition, and Codec.
Physical types, logical annotations, and INT64 precision
Columns store low-level bytes as primitives like INT64, DOUBLE, or BYTE_ARRAY while attaching semantic tags such as UTF8, DECIMAL, or TIMESTAMP_MILLIS described under Apache Parquet physical and logical types. Inspecting both physical primitives and converted annotations clarifies why raw storage differs from rendered spreadsheet columns.
When 64-bit INT64 IDs decode into JavaScript bigint values, integers beyond 53 bits exceed IEEE 754 precision limits documented in the MDN Number.MAX_SAFE_INTEGER reference. Values inside -9,007,199,254,740,991 to 9,007,199,254,740,991 stay numeric, whereas larger identifiers serialize as exact decimal strings so trailing digits never corrupt.
Handling timestamps, binary blobs, and nested structs
Non-primitive columns normalize into readable inspection strings before table rendering or spreadsheet export. Native JavaScript Date objects format to standard ISO 8601 UTC timestamps (toISOString()), while Uint8Array binary columns display compact 0x hex previews covering the first 16 bytes with an ellipsis when truncated.
Nested STRUCT, LIST, and MAP fields run through a BigInt-safe JSON serializer so complex warehouse events remain readable in single spreadsheet cells. When debugging line-oriented JSON logs exported alongside Parquet tables, verify raw event lines with the JSONL Viewer & Validator offline.
Windowed row-group pagination with Start row and Page size
Loading millions of columnar rows simultaneously would exhaust browser DOM memory fast. Set Start row (0 up to totalRows - 1) and choose a Page size of 100, 250 (default), 500, or 1,000 rows, then click Load slice to materialize only that window.
| Parquet type or value | Raw storage example | Normalized viewer output | CSV / Excel / JSON behavior | Precision or length note |
|---|---|---|---|---|
INT64 within safe bounds |
4291042 (bigint) |
Numeric 4291042 |
Written as numeric cell or JSON number | Up to 53-bit IEEE 754 integers |
INT64 outside safe bounds |
9223372036854775000 (bigint) |
Decimal string "9223372036854775000" |
Exported as string to prevent rounding | Preserves full 64-bit precision |
TIMESTAMP (Date) |
Epoch millisecond column | 2026-11-15T10:30:00.000Z |
ISO 8601 UTC timestamp string | Uniform timezone formatting |
BYTE_ARRAY binary |
Uint8Array(64) |
0x50415231... hex preview |
Compact 16-byte prefix summary | Avoids unreadable binary glyphs |
Nested STRUCT / LIST |
{ region: "iad", ok: true } |
Compact JSON string | Preserves nested keys inside single cell | BigInt-safe stringification |
Decoding Snappy, Zstd, and Gzip compression codecs locally
Data warehouses write column pages with distinct block compressors depending on whether queries prioritize CPU throughput or storage footprint. Our worker reads codec tags directly from row-group column metadata and dispatches matching pure-JavaScript or WASM-free decoders.
Supported compression formats and Parquet magic checks
Using hyparquet-compressors, the viewer decodes UNCOMPRESSED, SNAPPY, GZIP, and ZSTD pages straight from local byte slices in accordance with Apache Parquet data page compression codecs. Even mixed-codec files where one column uses SNAPPY and another uses ZSTD decode transparently column by column.
Files lacking matching leading and trailing PAR1 markers fail fast with an explicit invalid Parquet signature message before any row allocation happens. Unsupported experimental codecs such as LZO or Brotli likewise raise clear diagnostic errors naming the missing codec.
Filtering rows and exporting slices to CSV, Excel, and JSON
Typing into Filter loaded rows... searches across every normalized column value in your active slice and updates Total Rows loaded counters dynamically. Matching results export straight from browser RAM via three toolbar actions:
- Export CSV: Prepends a UTF-8 byte order mark (
\uFEFF) and applies MITRE CWE-1236 formula escaping so cells starting with=,+,-, or@never trigger spreadsheet code execution. - Export Excel: Writes filtered rows into a typed
Parquet Datasheet inside a.xlsxworkbook ready for desktop Excel or LibreOffice Calc. - Export JSON: Outputs a two-space indented JSON object array of your filtered slice, which pairs cleanly with our flat CSV to structured JSON converter when harmonizing mixed pipeline formats.
When should you slice rows before converting Parquet to CSV or Excel?
Spreadsheet applications slow down dramatically past tens of thousands of unindexed rows, and browser tables stall when painting thousands of DOM cells simultaneously. Choosing targeted Start row offsets and moderate Page size windows keeps exploration snappy across 50 MB analytics partitions.
Because Filter loaded rows... filters your currently loaded window (250 rows by default, up to 1,000 in UI controls and 10,000 in library helpers) and renders at most 500 table rows on screen, slicing protects tab stability. Check row-group totals and column widths first, start at row 0 with 250 rows to inspect schema mappings, then step forward by 1,000-row pages when auditing tail partitions or data anomalies.
Tip: Export targeted slices in batches of
500or1,000rows instead of opening massive raw files in spreadsheet software. You retain 100% of column fidelity without exhausting workbook memory.
Integrating Parquet exports into spreadsheet and lakehouse checks
Data engineers and financial analysts frequently pull partitioned Parquet extracts from Spark, DuckDB, Snowflake, or Pandas object storage for fast reconciliation without configuring Python or CLI runtimes. Exporting filtered slices locally guarantees confidential revenue figures, customer identifiers, and internal telemetry stay off untrusted third-party upload sites.
Every CSV download writes a UTF-8 byte order mark alongside MITRE CWE-1236 formula escaping so cells starting with symbols open safely in Excel. Excel workbooks (Parquet Data sheet) and formatted JSON slices retain clean column typing across downstream validation scripts, audit trackers, and BI workbooks without losing large identifier precision.
When audit checks span multiple daily Parquet extracts exported to CSV, stitch downloaded column slices together cleanly using multi-file CSV joins without Python scripts inside the same private browser workspace before finalizing quarterly reconciliation reports, financial sign-offs, or engineering schema handoffs.