Part 3 of 8 · 2 min read
Single points of failure
One part that takes everything down with it.
A single point of failure is any part whose loss takes the whole system with it. A system can have several.
Nygard (Release It!) adds a warning about integration points: every socket, call or data feed that joins one system to another. He names them the top killer of systems, because each one can hang at the worst moment and drag the caller down with it.
Single point of failure
A single point of failure (SPOF) is one component whose failure stops the whole system.
Example: a service that reads its settings from one database copy at start-up. If that database is down, the service cannot restart.
UDS gives a subtle example: an HTTPS service whose certificate must be renewed by hand. When the date passes, every client fails to connect. Nothing crashed. A date arrived.
The book's method for finding them is to look at every component and ask what happens if it fails. Some can be removed by adding redundancy (page 7). Those that cannot be removed must have their blast radius reduced: the amount of the system damaged when the part fails.
A worked example (made up). A shop has a load balancer, 6 web servers, 1 database primary and 1 login service. Ask the question of each. A web server fails: 5 are left, fine. The load balancer fails: nothing is reachable, so it is a SPOF unless there are two. The database primary fails: writes stop, so it is a SPOF until failover is set up. The login service fails: nobody can sign in, but browsing may still work, so the blast radius is smaller.
Check yourself: 1. Why is an expiring certificate a single point of failure? 2. What does blast radius mean? 3. In the example, which part has the smallest blast radius?