Skip to content
About
Main content

What overload defences forbid

Why does every defence cost work, optimality or fairness, and why is each one a new loop?

What this article forbids

  • It forbids a defence that preserves full service. Every mechanism costs work, optimality or fairness.
  • It forbids defence by capacity. Every pathology here is a ratio, and capacity is a level.
  • It forbids purely local defence. The three most valuable mechanisms (deadlines, , priority) are global invariants that need in-band carriage. This is the third confirmation of the global-invariant principle.
  • It forbids leaving defences untested. Each is a rarely run path whose first real execution happens during an incident.

The defences as one table. The comparison is the point.

DefenceAttacksTime scaleCost
Store (excess demand)msRejected users; needs classification
Concurrency limitStore (L, hence W)msCapacity must be known or adapted
Store (moved upstream)ms to sNeeds a producer that can slow down and a complete chain
Retry budgetGainsShared state; some valid retries denied
Gain + caller resourcessBinary; synchronized across clients
CorrelationsAdded
Wasted workper requestViral; conjunctive correctness
Hedging (not overload)per requestAbout 5% load; must disable under overload
Cost per requests to minPre-built degraded paths
AutoscalingCapacityminToo slow to be a defence at all

Keep the time-scale column in mind. Overload dynamics run in seconds. Anything in the "minutes" row is a cost mechanism wearing a defence's clothing.

Mitigations as new loops. Every defence is itself a control loop, so every defence has gain, delay and a failure mode:

  • Shedding generates retries unless clients respect the signal.
  • Breakers synchronize and starve a recovering callee.
  • Health-check ejection reduces capacity and so causes more ejection.
  • Hedging adds load in the conditions where it fires most.
  • Autoscaling is briefly negative capacity.
  • Backpressure plus a durable broker turns latency into retention debt.

The general rule: A system with many independently designed defences is a system with many interacting controllers, and usually nobody has checked their composition.

The discard principle, confirmed. Work is always discarded by someone. You can discard it at admission, cheaply and by policy. The client can discard it at , expensively and at random. Or the failure can discard it. Shed early costs about 0. Shed at timeout costs the full service cost. Shed by collapse costs everything. The only choice is who decides and what it costs.