Principles of Reliable Systems
|Time:||3:30 pm - 4:20 pm|
This is unsexy like the 1980s Automotive Quality Wars. The lesson from the quality wars is that things that break suck. But things that keep going are awesome. Awesome like the Terminator who keeps going after getting shot in the face with a shotgun. People will spend money for quality, and they will develop loyalties to things that don't break. They will avoid things that break (or that are perceived to break).
Reliability (quality) is central to the Erlang community, in large part due to the [mythic] number of 9 9's of uptime put forward by Joe Armstrong. Erlang was a commercially motivated language, not academically motivated, so quality and reliability had an associated cost for the creators. Erlang came out of PLEX, another language Ericsson designed. PLEX was a real time, very parallel language, but it was very low level, and therefore very expensive to use. The idea was to find or develop a new language that had an OS independent VM and had great support for parallelism and concurrency. Erlang was the product of this, and because of its motivations it prizes pragmatism over purity.
Principles of Reliability
- Fault detection and Recovery
- Separation of concerns
- Black box design
- State management
- Avoid complexity
The Erlang VM is designed to operate processes. You can kill individual Erlang processes without impacting other processes on the VM. We see isolation all around us in the physical world, sort of by definition. So we need to think about how to apply the same ideas to software and think about how to keep things isolation from one another.
Fault Detection and Recovery
In order to recover from a failure, you need to be able to detect the failure first. This isn't as easy as it sounds, especially at the thread level. When thinking about how to detect failure, you want to detect this as quickly as possible to "fail fast". Once the failure has been detected, you need a strategy to recovery. Erlang addresses this by "turning it off and on again" -- restarting the piece of code that registered the failure.
Separation of Concerns
Separation of Concerns is the principle of focusing on one thing, and doing it well. [Cohesion, etc.] By keeping code focused, it's easier to reason about it, test it, and limit the scope for a change. This also means that if something fails, the scope of failure is limited.
Black Box Design
Black box design is an approach to designing things where you treat your components as an appliance. The appliances in your home may be quite complex (washer, microwave, etc), but the interface it presents is limited by design. Thinking about code as an appliance means you try to make it easy to set up (just plug it in?), push the start button, provide minimal controls, and reboot or replace to fix.
Erlang is effectively an "operating system" for your code. So you write "systems", and individual "programs" within that service handle some specific concern.
Erlang doesn't "hate" state, but it doesn't like it very much. Messing with data is costly, and as soon as state enters your application you have to deal with additional complexity during recovery, failover, repair, and synchronization. All of these are hard to get right. If you can avoid state, you should, either by avoiding it completely, or ensuring it's someone else's problem.
Reducing complexity means there's fewer edges to test. Things like dependencies, hierarchies, resource sharing, and fear all are indicators of complexity. Something simple is something reliable. And if something isn't completely obvious, spend some time making sure someone else could understand it.
How to Do This
OS Process Isolation
- No shared memory
- Communicate via message passing
- Process termination ("fault") detection
- Techniques: IO "servers", 0MQ, TCP/HTTP
- Processes have overhead, and at some scales aren't feasible
- Actors provide semantically isolated memory
- Inter-thread communication via message passing (queue inserts)
- Avoid defensive practices -- let things fail
- Let exceptions propagate and log them
- Use assertions -- and leave them in!
- Exiting the process isn't a bad idea if you're running under a supervisor
- runit, launchd provide process monitoring/supervision [speaker recommends runit]
- Narrow the scope as much as possible
- Aim for functional-style programming -- average functions are four lines long
- Think about a Micro SOA
- Avoid building for "the future"