(5/7) Free expired slots once their workload is removed

Ashvin Sharmarequested to merge
feat/suspended-slot-ttl-free-expired into main

What does this MR do?

  • FreeExpired(key): frees an expired slot once its workload is gone. A no-op if the acquisition no longer exists.
  • A stray Release doesn't free an expired slot; only FreeExpired does.
  • An expired slot no longer keeps its instance suspended: when the last acquired slot is released, the instance stays running so cleanup can reach it.
  • New free_expired task-operation metric.

Why was this MR needed?

Part 5 of 7. Without it, expired slots stay allocated forever.

What's the best way to test this MR?

go test -race ./.... TestStrayReleaseDoesNotFreeExpiredSlot, TestFreeExpiredErrors, TestExpiredSlotDoesNotSuspendInstance.

What are the relevant issue numbers?

Part of the stack started in !117 (merged).

MR stack

Merge in order:

  1. (1/7) Give acquired slots an optional expiry (!117 - merged) • Ashvin Sharma
  2. (2/7) Let an instance mark a suspended slot exp... (!119 - merged) • Ashvin Sharma
  3. (3/7) Expire suspended slots past their expiry (!120 - merged) • Ashvin Sharma
  4. (4/7) Expose connect info for an expired slot's... (!121 - merged) • Ashvin Sharma
  5. (5/7) Free expired slots once their workload is... (!122 - merged) • Ashvin Sharma (this MR)
  6. (6/7) Remove stopped instances whose slots have... (!123 - merged) • Ashvin Sharma
  7. (7/7) Refuse to resume a suspended slot past it... (!118 - merged) • Ashvin Sharma
Edited by Ashvin Sharma

Merge request reports