I hope this won’t be counted as some form of self-promotion, even though I am sharing a post from my own blog.
As a tech worker who works in a Cloud shop, I wanted to elaborate the many reasons why I find working with Clouds terrible, from multiple points of view.
I tried to organize my thoughts in a (relatively long) post, in which both technical aspects and political aspects (which are very related) are covered.
I am sure many people will have different perspectives, and this could be potentially also a nice prompt for a discussion.
Complexity in this case should bring redundancy, not fragility. You are adding components in parallel, not in series, thus reducing fragility.
A raid 5 is more complex than a single drive, but it’s less fragile.
I wish it worked like that, but I donct think it does. Connecting clouds means introducing many complex problems. Data synchronization and avoiding split-brain scenarios, a network setup way more complex, stateful storage that needs to take into account all the quirks and peculiarities of all services across all clouds, service accounts and permissions that need to be granted and segregated for all of them, and way more. You may gain resilience in some areas, but you introduce a lot more things that can fail, be misconfigured or compromised.
Plus, a complex setup makes it harder by definition to identify SPOFs, especially considering it’s very likely nobody in the workforce is going to be an expert in all the clouds in use.
To keep using your simile of the disks, a single disk with a backup might be a better solution for many people, considering you otherwise might need a RAID controller that can fail and all the knowledge to handle and manage a RAID array properly, in addition to paying 4 or 5 times the storage. Obviously this is just to make a point, I don’t actually think that RAID 5 vs JBOD introduces comparable complexity compared to what multi-cloud architecture does to single-cloud.
Split brain are easily solved, there’s of the shelf solutions and if you have some custom code you can use plenty of well researched solutions, for instance raft. Putting bizantine fault in Google scholar yields thousands of papers,if you want something fancier.
Same for most problems you mentioned, they were an issue 10 years ago, nowadays you can federate, abstract or outsource most of it.
Making it harder to identify SPFOs doesn’t increase fragility. If you whole system a single instance it’s trivial to identify (the whole thing) but very brittle.
Of course the problem is solved, but that doesn’t mean that the solution is easy. Also, distributed protocols still need to work on top of a complicated network and with real-life constraints in terms of performances (to list a few). A bug, misconfiguration, oversight and you have a problem.
Just to make an example, I remember a Kafka cluster with 5 replicas completely shitting its pants for 6h to rebalance data during a planned maintenance where one node was brought offline. It caused one of the longest outages to date with the websites which relied on it offline. Was it our fault? Was it a misconfiguration? A bug? It doesn’t matter, it’s a complex system which was implemented and probably something was missed.
Technology is implemented by people, complexity increased the chances of mistakes, not sure this can be argued.
Making it harder to identify SPOF means you might miss your SPOF, and that means having liabilities, and having anyway scenarios where your system can crash, in addition for paying quite a lot to build a resilience that you don’t achieve.
A single instance with 2 failure scenarios (disk failure and network failure) - to make an example - is not more fragile than a distributed system with 20 failure scenarios. Failure scenarios and SPOF can have compensating controls and be mitigated successfully. A complex system where these can’t be fully identified can’t have compensating control and residual risk might be much harder. So yes, a single disk can fail more likely than 3 disks at once, but this doesn’t give the whole picture.
The only problem is that the single instance also has 20 scenarios (and keeps the 2 as well), making it more brittle.
A well design system removes points of failure, disk, power and network are obvious ones, and as long as you keep it byzantine safe, anything you added should be redundant so if one fails the system still runs. Ideally you remove all of them but if there’s one hidden it’s still better than “the whole thing is a single point of failure”.
No, it’s not true. A single system has less failure scenarios, because it doesn’t depend on external controllers or anything that makes the system distributed and that can fail causing a failure to your system (which may or may not be tolerated).
This is especially true from a security standpoint: complexity adds attack surface.
Simple example: a kubernetes cluster has more failure scenarios than a single node. With the node you have hardware failure, misconfiguration of the node, network failure. With a kubernetes cluster you have all that for each node (each with marginally less impact, potentially, because it depends for example on stateful storage, that if you mitigate you are introducing other failure scenarios as well), plus the fact that if the control plane goes in flames your node is useless, if the etcd data corrupts your node is useless, anything that happens with resources (a bug, a misuse of the API, etc.) can break your product. You have more failure scenarios because your product to run is dependent on more components to work at the same time. This is what it means that complexity brings fragility. Looking from the security side: an instance can be accessed only from SSH, if you are worried about compromise you have essentially one service to secure. Once you run on kubernetes you have the CI/CD system, the kubernetes API, the kubernetes supply-chain, etcd, and if you are in cloud you have plenty of cloud permissions that can indirectly grant you access to the control plane and to a console. Now you need to secure 5-6-7 entrypoints to a node.
Mind you, I am not advocating against the use of complex systems, sometimes they are necessary, but if the complexity is not fully managed and addressed, you have a more fragile system. Essentially complexity is a necessary evil to respond to some other necessities.
This is the reason why nobody would recommend to someone who needs to run a single static website to run it on Kubernetes, for example.
You say “a well designed system”, but designing well is harder the more complexity exists, obviously. Redundancy doesn’t always work, because redundancy needs coordination, needs processes that also depend on external components.
In any case, I agree that you can build a robust system within Cloud! The argument I am trying to make is that:
And mind you, everything you can do in Cloud you can also do on your own, if you invest on it.
You make it redundant, I thought I didn’t need to say that…
I am specifically saying that redundancy doesn’t solve everything magically. Redundancy means coordination, more things that can also fail. A redundant system needs more care, more maintenance, more skills, more cost. If a company decides to use something more sophisticated without the corresponding effort, it’s making things worse. If a company with a 10 people department thinks that using Cloud it can have a resilient system like it could with 40 people building it, they are wrong, because they now have a system way more complex that they can handle, despite the fact that storage is replicated easily by clicking in the GUI.
Redundancy should be automatic. Raid5 for instance.
Plus cloud abstracts a lot of complexity. You can have an oracle (or postgres, or mongo) DB with multi region redundancy, encryption and backups with a click. Much, much simpler for a sysadmin (or an architect) than setting the simplest mysql on a VM. Unless you’re in the business of configuring databases, your developers should focus on writing insurance risk code, or telco optimization, or whatever brings money. Same with k8s, same with Kafka, same with cdn, same with kms, same with iam, same with object storage, same with logging and monitoring…
You can build a redundant system in a day like Legos, much better security and higher availability (hell, higher SLAs even) than anything a team of 5 can build in a week self-manging everything.