Parquet Reader

Open .parquet files in VS Code and read them as a table — column types, per-column statistics, and full nested values, without leaving the editor and without loading the whole file into memory.

Features

Remote files over HTTPS, S3 and Google Cloud Storage

Run Parquet: Open Parquet from URL… from the Command Palette:

Then paste an https://, s3:// or gs:// URL:

• https:// — the server must support HTTP range requests. If it does not, the viewer says so instead of silently downloading the entire file.
• s3://bucket/key.parquet — signed by shelling out to aws s3 presign, so your existing AWS profile or SSO session does the signing and no credential ever passes through this extension. Presigned https:// URLs work directly, with no AWS CLI required.
• gs://bucket/key.parquet — read with an access token from gcloud auth print-access-token, so whatever account gcloud is logged in with — a Google account or a service account — does the reading. The token is held in memory only, and is renewed automatically when it expires while a file is open.

Remote files are read the same way as local ones: only the footer and the data pages you are actually looking at cross the network.

Opens from the Explorer

Double-click any .parquet or .pq file. No command to run, no conversion step. The viewer is strictly read-only — it never writes back to your file.

The header reports what the file actually contains: row count, column count, file size, the compression codecs in use, and the writer that produced it.

Data tab

• Rows per page is picked in the footer — 25, 50, 100, 250, 500 or 1000 — with row numbers and a sticky two-line header showing each column name and its type
• Numeric columns are right-aligned; NULL is rendered distinctly from an empty string
• Nested columns (STRUCT, LIST, MAP, VARIANT) stay in a single column, rendered as JSON
• Long values are shortened to 200 characters in the grid, so a page stays fast to render

Search within the page

Type in the box above the table to keep only the rows that match and highlight the matches. It is case-insensitive and looks at every column.

This searches the page you are looking at, not the whole file — the count beside the box says so. It runs entirely inside the view: nothing is sent to the extension, and no byte is read.

Column sorting

Click a column header to cycle ascending → descending → back to file order.

Sorting compares the values themselves rather than their printed form, so 9 sorts before 10 and timestamps sort chronologically. NULL sorts last in both directions, because it is the absence of a value rather than the smallest one.

Nested columns (STRUCT, LIST, MAP, VARIANT) and repeated columns hold more than one value per row, so there is nothing single to order by. Their headers show a not-allowed cursor, and clicking one explains why in the toolbar rather than silently doing nothing.

Sorting reads the whole file into memory, so it is capped by parquetReader.sortCellBudget (500,000 cells by default, counted as rows × columns). Beyond that limit the headers stop responding and explain the limit, rather than freezing the window on a file that will not fit.

Cell detail

Click any cell to see its untruncated value in the pane below the table. Nested values are pretty-printed, and Copy puts the full value on your clipboard. Press Esc to close.

The full value is fetched on demand, one cell at a time — it is never included in the page payload, which is what keeps wide, deeply nested files responsive.

Schema tab

One row per leaf column, using dotted paths for nested fields. For each column: physical type, logical type, nullability, null count, min, max, and compressed size summed across all row groups. Columns whose writer omitted statistics show — rather than a misleading zero.

Follows your theme

Light, dark, and high contrast themes are all supported — the viewer uses VS Code's own colour tokens rather than its own palette.

Performance

The viewer never reads a whole file. It reads the footer for metadata, then only the data pages needed for the rows on screen, using the file's offset index when one is present.

A 180 MB file with 1,244,896 rows stored in a single row group on S3 opens in 1–2 seconds, and page navigation stays under a second. Without offset-index support the same file took about 30 seconds to display its first page, because the first read had to pull every column chunk in the row group.

Files written without an offset index (for example, PyArrow without write_page_index=True) fall back to reading whole column chunks, which is correct but slower to open.

Requirements

• VS Code 1.90 or newer
• The AWS CLI, only if you want to open s3:// URLs.
• The Google Cloud SDK (gcloud), logged in, only if you want to open gs:// URLs. It must be on the PATH VS Code sees. VS Code reads PATH from your shell once, when it starts — if you installed the SDK while VS Code was running, quit VS Code entirely and start it again.

Everything else works out of the box.

Extension settings

Page size does not change how much of the file is read: the smallest unit Parquet can read is a data page, which is already larger than any page of rows, so 1000 rows costs the same bytes as 100. What it does change is rendering, and that grows with rows × columns rather than with rows alone. So the effective page size is capped at roughly 20,000 cells: a 4-column file gets the full 1000 rows, a 60-column file gets 333. The footer says when the value was lowered.

The Rows per page picker in the footer writes this same setting, so a choice made there sticks for the next file you open. Both settings take effect immediately in open tabs — changing the page size keeps whichever row was at the top of the page in view.

Known limitations

• Read-only. The viewer never writes back to your file.
• Search covers the current page, not the whole file. Whole-file search is not implemented yet.
• Sorting is limited to files under parquetReader.sortCellBudget, and to one column at a time. There is no filtering or query support yet.
• A BYTE_ARRAY column carrying no logical type is decoded as UTF-8 text by the underlying reader, so genuinely binary columns (hashes, blobs) display as unreadable text.
• Presigned S3 URLs expire after one hour. Reopening re-signs automatically, because the original s3:// URL is what gets stored, not the signature.

License

MIT — see LICENSE.