A research paper presented at ICCCN'26 investigates how to build a metastability-tolerant execution control system (ECS) for databases. The ECS uses ticket-based concurrency pools and priority queues to protect short latency-sensitive queries while shedding excess load. Using MESSI, a discrete-event simulator written in Go, the authors discovered that two independently sensible controllers (one per queue) can create a positive feedback loop under overload — the ticket-acquisition metric rewards churn rather than real progress, causing both queues to inflate tickets and crowd the CPU. The fix: freeze the long queue's tickets at a static value and use a single controller with an ungameable signal (whether anyone is actually waiting). The paper also shows that admission control and the ECS can destructively interfere, producing sawtooth oscillations in goodput. Fixes include having the ECS export an explicit distress signal and making the outer admission control loop deliberately slower than the ECS's convergence time. Key takeaways: performance IS availability (metastability erases the boundary), and simulation before deployment is essential to surface hidden feedback loops.
