The Netflix Simian Army (2011)(netflixtechblog.com) |
The Netflix Simian Army (2011)(netflixtechblog.com) |
I think the companies I worked for were prioritizing working on no issue deployments (built from a series of documented and undocumented manual processes!) rather than making services resilient through chaos testing. As a younger dev this priority struck me as heresy (come on guys, follow the herd!); as a more mature dev I understand time & effort are scarce resources and the daily toil tax needs to be paid to make forward progress… it’s tough living in a non-ideal world!
I think that's why most companies don't do it. A lot of tedium and the main benefit was actually getting your ducks in a row.
A significant problem with early 'Web Scale' deployments was out of date or stale configuration values. You would specify that your application connects to backend1.example.com for payments and backend2.example.com for search. A common bug in early libraries was that the connection was established once at startup, and then never again. When the backend1 service was long lived, this just worked for months or years at a time - TCP is very reliable, especially if you have sane values on keepalives and retries. Chaos Monkey helped find this class of bug. A more advanced but quite similar class of bug: You configured a DNS name, which was evaluated once at startup, and again didn't update, Your server for backend1 had a stable address for years at a time, but suddenly you needed to failover to your backup or move it to new hardware. At the time of chaos monkey, I had people fight me on this - They believed that doing a DNS lookup every five minutes for your important backends was unacceptable overhead.
The other part is - Modern deployment strategies make these old problems untenable to begin with. If you're deploying on kubernetes, you don't have an option here - Your pods are getting rebuilt with new IP addresses regularly. If you're connecting to a service IP, then that IP is explicitly a LB - It is defined as stable. These concepts are not complex, but they are edge boundaries, and we have better and more explicit contracts because we've realized the need and you "just do" deploy this way now.
Those are just Chaos Monkey problems, though - Latency Monkey is huge, but solves a much less common problem. Conformity Monkey is mostly solved by compliance tools; You don't build, you buy it. Doctor Monkey is just healthchecks - K8s (and other deployment frameworks) has those built in.
In short, Chaos Monkey isn't necessary because we've injected the chaos and learned to control most of what that was doing, and people have adopted the other tools - They're just not standalone, they're built in.
If you know things will break when you start making non-deterministic configuration changes, you aren't ready for chaos engineering. Most companies never get out of this state.
In the stateful world, chaos testing is useful, but you really want to be treating every possible combination of failures at every possible application state, theoretically with something like TLA or experimentally with something like Antithesis. The scenarios that you can enumerate and configure manually are just scratching the surface.
Currently we do shadow shifts for a month or two first, but still eventually drop people into the deep end with whatever experience production gifts them in that time. That experience is almost certainly going to be a subset of the types of issues we see in a year, and the quantity isn’t predictable. Even if the shadowee drives the recovery, the shadow is still available for support & assurance. I don’t otherwise have a good solution for getting folks familiar with actually solving real-world problems with our systems, by themselves, under severe time pressure, and I was thinking controlled chaos could help bridge the gap.
I remember this getting a lot of buzz at the time, but few orgs are at the level of sophistication to implement chaos testing effectively.
Companies all want a robust DR strategy, but most outages are self-inflicted and time spent on DR would be better spent improving DX, testing, deployment and rollback.
It doesn’t test nearly as much as the real tools can, but it did find some bugs in our workflow engine where it wouldn’t properly resume failed tasks.
So ad-hoc, home-grown, chaos testing is still a useful exercise!
https://developer.chrome.com/docs/chromedriver/mobile-emulat...
I was reading this the other day looking for ideas on how to test query retries in our app. I suppose we could go at it from the network side by introducing latency and such.
However, it’d be great if there also was a proxy or something that could inject pg error codes.
It's been 15 years. Aws still sucks compared to your own hardware on so many levels, and total Roi is dropping.
Today, many of these ideas map directly to some of their managed services like AWS Fault Injection Simulator, AWS Resilience Hub, or AWS Config, AWS Inspector, Security Hub, GuardDuty, and IAM Access Analyzer for example.
There is also a big third-party ecosystem (Gremlin, LitmusChaos, Chaos Mesh, Steadybit, etc...) offering similar capabilities, often with better multi-cloud or CI/CD integration.
Some of these Netflix tools, I dont think they get much maintenance now, but as free options, they can be cheaper to run than AWS managed services or Marketplace offerings...
Hazing us a cycle of abuse that expresses in a magnification of the abuse inflicted in the hazing than was suffered in the previous cycle.
Maybe you are optimizing your personnel.