[comphead]

This adds a draft blog post announcing the nested benchmark derived from TPC-H that was added to Apache DataFusion Comet in apache/datafusion-comet#7044.

Structure of the post

• Teaser: the headline result and a chart of total execution time at SF1000 for Spark, Comet nested unoptimized and Comet nested optimized.
• The problem: TPC-H is flat, so it does not exercise nested columns, explode or nested pruning that real workloads use. Includes the choke points of nested data and the case in Improve performance for roundrobin shuffle writes for nested schemas datafusion-comet#5397.
• The solution: an adapter that nests TPC-H. Keys and partition columns stay at the top level, everything else moves into a single-element array<struct>, so the query answers match TPC-H. Covers the schema, the rewritten queries and the Spark conversion script.
• Results: total time and per-query time for the three configurations (fastest of 3 iterations, one run each), a per-query table, and the issues the benchmark helped to surface: Investigate nested TPC-H q21 slowdown: Comet 37% slower than Spark at SF1000 datafusion-comet#6467 and #6698, #6724 and #6725 (follow-up #7019), and #5397.
• What's next: a nested TPC-DS, other nested shapes, more engines.

Files

• content/blog/2026-10-25-comet-nested-tpch.md
• content/images/comet-nested-tpch/intro-total.svg and results-queries.svg: the two charts.
• content/images/comet-nested-tpch/dashboard.html: a self-contained interactive version of the results, linked from the Results section.

Draft status

• The author: line is a placeholder (comphead) and needs a full name and affiliation.
• The paper references in the problem section are written from memory and have not been checked.
• Each configuration is a single run. The optimized Comet run uses newer Spark, Iceberg and Comet builds than the Spark reference, and the cluster size is not recorded with the runs, so the post does not attribute the change in Comet's time to individual changes.
• The two charts and the dashboard embed a subset of the Lato font (SIL Open Font License 1.1, glyphs unmodified, copyright and license text kept in the font files). Maintainers may want to confirm that embedding it is acceptable for this repository.
• The site has not been built with Pelican. The site/ branch name only stages for branches in this repository, so a maintainer may need to push a branch here, or the site can be built locally with make.

[comphead]

Rendered: https://github.com/comphead/datafusion-site/blob/preview-comet-nested-tpch/content/blog/2026-10-25-comet-nested-tpch.md