• Sources: primary
  • Summary: Wayne walks a worker that checks a done flag, marks it and does the work, and shows that a failed step plus retry loses the work, while reordering it lets two workers do it twice. He separates machine variables from world variables, checks safety as an invariant over states and liveness over paths, and argues that every design rests on environment assumptions such as finite failures or strong consistency that break when a vendor weakens them. In Q and A he says LLMs save boilerplate when writing specs but are poor at writing properties and cannot do liveness.
  • Why it matters: Asking what happens when every retry, crash and interleaving goes the worst way catches duplicate-execution bugs in job runners and schedulers at design review.
  • Background: A safety property says something bad never happens, and a liveness property says something good eventually happens. Model checkers such as TLC for TLA+ enumerate every reachable state of a small model to check both. Wayne wrote Learn TLA+.