• Sources: primary, reddit
  • Summary: incident.io ran each item of a FOR UPDATE SKIP LOCKED batch in its own savepoint, which placed a MultiXact in every locked row. Each skipped row then cost a MultiXact lookup behind one lightweight lock, so adding workers made claim acquisition slower. Splitting the claim from the work gave a 3.5x throughput gain.
  • Why it matters: A queue built on SKIP LOCKED can hit a ceiling that looks like database load but comes from how the claim transaction is structured.