What placement forbids

What this article forbids

You cannot be concentrated for bandwidth and dispersed for failure at the same time. They are the same axis.
You cannot exceed the utilization ceiling set by shape mismatch, whatever the algorithm.
You cannot enforce a constraint the scheduler has no name for.
You cannot make a lifetime-valid placement with a point-in-time decision.

Escapes and prices

Escape	Buys	Pays
Relaxation ladders	feasibility	unpredictable performance (the same task runs at two speeds depending on where it landed)
Constraints (anti-affinity)	independence	packing density and search cost
Spot	60–90% off	lost work, and correlation you must actively diversify against
Shape menus (instance families)	less stranding	fleet complexity
Descheduling	durable constraints	preemption: the collector always costs what the mutation saved

Case, resolvedConstructed case · Lena at Brightlane

Three replicas on three machines, down together

A rack failure took down all three replicas of a service, Omar's job ran 2× slower on some days, CPUs sat at 50% with memory full, and a spot cluster lost a pipeline mid-run. Where it started: Deciding where →

Three replicas down together. The scheduler had packed all three into one rack because it was cheapest and fastest. Lena adds anti-affinity ("replicas of this service spread across racks and AZs"), accepts the small loss in packing density, and adds a descheduler that fixes drift after maintenance.
Omar's job 2× slower on some days. Its tasks sometimes landed in different zones and paid cross-zone bandwidth. The job is an AND-job, so she tells the scheduler: concentrate it (same rack if possible), not spread it. She also moves its shuffle files to local NVMe.
CPU 50%, memory full. Shape mismatch: tasks of 1 CPU and 16 GB on machines of 32 CPU and 256 GB. She adds a memory-heavy machine family for such tasks and a CPU-heavy family for the others.
Spot surprises. Spot is only used for fine-grained, idempotent tasks (tasks under 10 minutes), spread over several instance types and zones. The coordinator and long shuffles stay on regular machines. She also compares hardware in $ per job, not $ per hour, and moves part of the fleet to the newer, 2.5×-price machines.