Unreliable vendors
Throughout my careers I have had to deal a lot with vendors, be that infrastructure, services, ... and one thing shines through, when you rely on someone else it is best to assess the impact of said vendor and have your contingency plans.
This post is in no way shape or form intending to talk about any vendor, it will mainly be an attempt to convey my perspective on how I have dealt with the issue of Unreliable vendors. When you are running a service, it's easy to shift blame to your vendor for being down or introducing a bug but you are responsible for the choices of your product. Let that sink in, we can't just shift accountability to a third party.
Reduce impact
One of the most important things would be reducing the impact of a third party having downtime. When my billing provider is down should unrelated API Calls that happen to have a level of overlap go down? No, they really shouldn't we should optimise for providing service without that x% of our API surface not working.
In the GraphQL world this could be done by i.e. having that part of the
API response be nullable and providing an appropriate GraphQLError to
highlight that billing is currently not available.
Reducing impact is ofcourse not possible for every vendor, if your infrastructure goes down well... You can convey that to your user but the impact will be hard to reduce.
Fallbacks
Another common thing might be having a fallback, a slower read-only database or another infrastructure provider deployed in parallel that just requires a DNS switch. A lot of these things can be controlled with little switches and should be as automatic as possible.
One example of this could be that you have a distributed KV store that sometimes goes down, when it does go down you fallback to some blob storage like S3 to get the data you need... Yes it will be slower but your customers won't be impacted.
A fallback will be a tradeoff to ensure resilience and create uptime guarantees.
Optimise for reproduction
One last thing we can do is optimise for reproduction, a vendor does not neccesarily go down... They might just introduce a bug in their systems which introduces unexpected interactions. For this case I have personally started thinking about optimising for reproductions.
It's very hard to contextualise what kind of customer you are, are you doing the most basic thing with their software or are you living on the bleeding edge. Are you the only one impacted by a certain issue or are more people affected but is just not that much of a problem for them?
In whatever bucket you fall here, optimise for reproduction, try to have some tool that generates all of the contact points with the third party provider so you can give all of the URL's, SDK calls,.. to their support team to show the issue.
For infrastructure providers I feel this is really crucial, being able to either spin up some piece of infrastructure that highlights the behavior in isolation is crucial for them to understand what part of their (distributed) system is buggy.
None of this is easy and building for resilience takes deep knowledge of what you're building and what your most critical pieces are but... It's better than having to blame your vendors at every incident.
All of this doesn't need to be isolated to vendors, heck optimising for reproduction might as well be applied to open-source.