Skip to content

08 — Supervisors

Status: supervisor strategies, restart rate limits (restart up_to N in DUR), and uniform-jitter backoff (backoff D1..D2) are implemented at the runtime layer as of v0.7. See docs/internals/supervisors.md.

A supervisor owns a set of child agents and decides what to do when one fails. Mighty's supervisors borrow from Erlang/OTP: strategies are named, and each child gets an explicit restart policy.

The program

supervisor SearchFlow(strategy: one_for_one) {
  child planner = spawn Planner()
  child fetcher = spawn Fetcher(net)

  on_fail(planner) { restart up_to 3 in 30s }
  on_fail(fetcher) { backoff 100ms..2s; restart }
}

What is interesting

  • supervisor SearchFlow(strategy: one_for_one) names the supervisor and selects its strategy. Per spec §15, valid strategies are one_for_one, one_for_all, rest_for_one, and escalate.
  • child planner = spawn Planner() declares a child agent. The handle planner is in scope inside on_fail clauses.
  • on_fail(planner) { restart up_to 3 in 30s } says: if planner fails, restart it up to 3 times in any 30-second window. Exceeding the budget triggers MT5014 restart_limit_exceeded and escalation per the strategy.
  • on_fail(fetcher) { backoff 100ms..2s; restart } adds uniformly distributed jittered backoff between 100ms and 2s before restarting.

Try it

mty check examples/10_supervisor.mty

Next

Continue to 09 — Arenas.