Skip to content- Sources: primary
- Summary: In a self-managed deployment, metadata latency rose from milliseconds to thirty seconds as SIRead locks grew to hundreds of thousands under SERIALIZABLE isolation. Materialize redesigned its compare-and-set log around a contiguity invariant, a FOR KEY SHARE lock on the predecessor row and the primary key. It then proved the design linearizable under a model of READ COMMITTED in Lean 4. The new implementation sustained three times the write throughput and degraded gracefully past database capacity.
- Why it matters: Teams running hot compare-and-set tables on PostgreSQL SERIALIZABLE can be exposed to the same SSI lock buildup, and this post shows a verified alternative.