Chris Douglas (part of the hadoop team at Y!, worked on adls gen 1 and other things at microsoft and now doing a PhD in databases at UCB) has just published some stats on experiments on write performance to iceberg.
https://cdouglas.github.io/posts/2026/03/catalog Summary: the most optimistic single-table transaction rate is low (3-4 commits/sec), dominated by the cost of reading, merging, and writing metadata in the object store. It would be interesting to consider whether single-file commits will help here -and if not, why not?
