SpacetimeDB: A Short Technical Review

SpacetimeDB launched version 2.0 earlier this week with a peculiar marketing approach: a surreal, meme-y video mocking its competitors while drinking "competitor's tears," paired with benchmarks that seem too good to be true—and the author says they are, indeed, not true. The benchmarks even mock other databases, with a magnifying glass inviting viewers to zoom in on how badly the losers perform. The author finds this distasteful but acknowledges interesting ideas in the product and sets out to review it fairly.

The author's main critique is that the benchmarks are technically flawed and dishonest. They compare SpacetimeDB against databases that make very different tradeoffs, which makes the comparison appealing but unfair. He draws on his own experience at PlanetScale, where he helped build a MySQL extension for vector similarity search that was fully transactional and stored vector data on disk using MySQL's buffer pools—unlike pgvector's HNSW approach, which requires the similarity graph to fit in memory. As he notes, it was immensely alluring to take an EC2 instance with 32GB of RAM, load 64GB of vector data into his database, run the same queries against Postgres, and show PlanetScale handling tens of thousands of queries per second while pgvector struggled. That kind of benchmark would not reflect a real comparison.

The author argues that newcomers often believe winning requires "the best performance," but sustainable database offerings are built on good, honest technical work. By not providing honest benchmarks, SpacetimeDB undercuts its own credibility, even though the underlying product may have interesting ideas.