4.2 Fault tolerance and redundancyBuilt to survive broken parts

Elements of complex systems fail, at unexpected times and often in groups. A system that can support failures and still continue to function is fault tolerant. The internet was engineered to be fault tolerant, with abstractions for routing and transmitting data, so you can keep using it while parts of it are broken.
The tool is redundancy: the inclusion of extra components that can be used to mitigate failure if other components fail. For a network, the main kind is having more than one path between devices. If a connection or device on one path fails, routers send data along a different path, if one exists. Redundancy costs something, since extra cables and routers must be bought and kept running even when they're not needed, but it buys reliability. The redundancy of routing options also helps the internet scale to more devices and more people.
?Wait, how is a path different from a connection?
Look at the ring. From A to F there are exactly two paths, one around each side: A-B-D-F and A-C-E-F. Break any one connection in the ring and the other side still works. To cut A off from F, you need at least two failures, one on each side. That's fault tolerance.
Now look at G. Its only connection is D-G. If that one connection fails, or if device D fails, G can't reach anyone. A part whose failure alone breaks the system is called a single point of failure. The fix is more redundancy: one extra connection.
?Wait, what happens to the connections when a device fails?
I thought about it. Show me
- (a) Try one failure: any single ring connection leaves the other side of the ring. So 1 is not enough.
Why?
Always start at 1 and go up. The first number of failures that works is the minimum. - Try two: A-B and A-C together cut A off. So the minimum is 2. (B-D with E-C would also work, one failure on each side of the ring.)
- (b) Connection D-G and device D: either failure alone cuts G off. No other single connection or device breaks anything.
- (c) With F-G added, list paths to G by the last step. Into G from D: A-B-D-G and A-C-E-F-D-G. Into G from F: A-C-E-F-G and A-B-D-F-G. That is 4 paths.
- Check for single points of failure again: G now has two connections, to D and F, so losing D-G or device D leaves the route through F. None are left.
Why?
Python agreed when I brute forced every possible single failure. On the exam you test a few likely ones, starting with the device that has the fewest connections.
Answer: (a) 2. (b) Connection D-G and device D. (c) 4 paths from A to G, and no single point of failure remains.
Stuck? Get a hint
- For (a), test single failures first. Does any one ring link stop A from reaching F?
- For (b), which device has only one connection?
- For (c), every path to G must enter through D or through F. Count each kind.
- ARedundant paths make packets arrive out of order, which cannot be fixed
- BExtra connections reduce the bandwidth of every other connection
- CExtra hardware must be bought and maintained, even though it may sit unused until something fails
- DRedundancy removes the need for protocols
Show the answer
C. Redundancy means extra components, and those cost money and upkeep. The payoff is that the network keeps working when parts fail. The first choice is a trap: out-of-order packets are normal and are fixed with packet numbers.
- Fault tolerant: the system keeps working when some parts fail.
- Redundancy: extra components, mainly extra paths, that take over when others fail. It costs extra hardware.
- A single point of failure is a connection or device whose failure alone cuts devices off.
- A failed device takes all of its connections with it.