Noisy Neighbour

The failure mode that shared infrastructure exists to create. It appears at every layer: a tenant running an expensive query saturating a shared database, a container consuming disk I/O on a shared host, one customer's batch job filling a shared queue, a hot partition slowing every request routed to that node.

Its distinctive quality is that the victim has done nothing wrong and has no visibility into the cause. Investigations begin with "our service got slower" and no local explanation, which is why per-tenant metrics matter more here than aggregate ones.

Controls, in increasing order of strength: quotas and rate limits per tenant; resource requests and limits at the container or query level; admission control with per-tenant concurrency caps; shuffle sharding, so any two tenants rarely share a full set of resources; and finally dedicated capacity for the largest tenants, which is where multi-tenancy economics stop paying.

The measurement that makes it tractable: attribute resource consumption per tenant, so the noisy one can be identified in minutes rather than by elimination.