Notesfoundations🌱 seedling
Single Point of Failure
Also: SPOF
A single point of failure is any component whose failure alone takes the whole system down. Removing them means [[Redundancy]] at every layer, including the layers that provide redundancy, like load balancers and discovery.
Why it exists
A system is only as available as its least redundant dependency on the request path. Three servers don't help if every request passes through one balancer, one database or one registry.
How it works
Find SPOFs by tracing every request path and asking "if this one thing died, would users notice?" The usual fixes:
- Active-passive: a standby watches heartbeats and takes over (a floating IP via VRRP, or Failover of a database leader).
- Active-active: several peers serve at once and traffic is spread across them (ECMP routing to many balancers, or many replicas behind a Load Balancer).
- Graceful degradation: keep working on cached or last-known data when a dependency is gone.
Trade-offs
| Approach | Gains | Costs |
|---|---|---|
| active-passive | simple, no shared state | idle capacity; slow or wrong takeover (Split Brain) |
| active-active | uses all capacity, fast | peers must agree (consistent hashing, shared state) |
In practice
- Load balancers: Google's Maglev runs many active balancers behind ECMP and uses consistent hashing so any of them sends a connection to the same backend.
- Hidden SPOFs are usually shared infrastructure: DNS, a service-discovery cluster, a config store, one cloud region.
Pitfalls
- Redundant servers behind a non-redundant front door just move the SPOF.
- A standby that's never been failed over to is an untested assumption.
- Correlated dependencies: two copies sharing one power feed, network or config push are still one point of failure.
Connections
- Part of
- Fault Tolerance
- Related
- FailoverRedundancy
- Mentioned
- Load BalancerSplit Brain
Recall
These cards are also in your review deck, scheduled with spaced repetition.