feat: Improve Table Previews

[camielvs]

Description

Follow-up to #2557. This PR adds extra viewer affordances for both Parquet and CSV/TSV.

1. Load more / Load max (Parquet). The base PR previews the top 100 rows and stops. Here the viewer can pull additional pages on demand via range reads — Load more fetches the next batch, Load max fills up to the preview cap. Only the newly requested range is fetched each time; rows already loaded are never re-read, and the whole file is never downloaded.

2. Preview limit is a cell budget, not a row count. The preview table renders every cell into the DOM (no virtualization), so its cost scales with rows × columns, not rows alone. A flat 1,000-row cap therefore over-protects narrow tables and under-protects wide ones. Instead, the preview is bounded by a 50,000-cell budget: the row limit is floor(50,000 / columnCount), with an absolute backstop of 10,000 rows so a very narrow table still can't flood the DOM. This adapts to table shape:

Columns	Preview row limit
≤ 5	10,000 (backstop)
10	5,000
20	2,500
100	500

Rows are atomic: the budget is floored to whole rows up front, so the rendered cell count never exceeds 50,000 and no partial rows are shown. The same limit applies to both Parquet (Load max) and CSV/TSV previews.

3. Download full dataset. When the preview limit is reached but the file still has more rows than are shown, the footer surfaces a Download full dataset link so the user has a clear escape hatch to the complete data (opens the signed URL for remote artifacts).

4. CSV/TSV parity. CSV/TSV previews now report the exact total row count and column count in the header, matching Parquet. Counting is done with a streaming parse (every row is counted, only the preview rows are retained). Remote CSVs get a Download full dataset link to the signed URL; inline CSV values trigger a direct file download.

Related Issue and Pull requests

• Stacked on feat: Improve Parquet Viewer #2557

Type of Change

• Improvement

Checklist

• I have tested this does not break current pipelines / runs functionality
• I have tested the changes on staging

Test Instructions

1. Open a run with a Parquet artifact that has more than 100 rows and open its preview.
2. Confirm a top-100 preview loads, then use Load more / Load max and confirm additional rows load without a full-file download.
3. Confirm the preview stops at the cell budget: for a wide table (e.g. 50+ columns) Load max stops well below 10,000 rows, and once the limit is reached a Download full dataset link appears.
4. Open a CSV/TSV artifact and confirm the header shows exact row/column counts, and that the download affordance works for both remote and inline CSV.

Additional Comments

[github-actions]

🎩 Preview

A preview build has been created at: 07-29-feat_parquet_improvements_2/227a763

[camielvs]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• feat: Artifact Previews on v2 Runs Output Nodes #2588
• feat: Improve Table Previews #2593 👈 (View in Graphite)
• feat: Improve Parquet Viewer #2557
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.