[dlenski]

The idea that AWS's services are fully regionalized or isolated has always been a myth.

All the identity and access services for the public cloud outside of China (aka "IAM for the aws partition" to employees) are centralized in us-east-1. This centralization is essentially necessary in order to have a cohesive view of an account, its billing, and its permissions.

And IAM is not a wholly independent software stack: they rely on DynamoDB and a few other services, which in turn have a circular dependency on IAM.

During us-east-1 outages it's sometimes possible to continue using existing auth tokens or sessions in other regions, while not possible to grant new ones. When I worked there, I remember at least one case where my team's on-calls were advised not to close ssh sessions or AWS console browser tabs, for fear that we'd be locked out until the outage was over.

[Roark66]

Anyone who thinks one cloud provider will provide them full resilience is fooling themselves. You need multicloud for true high availability.

But then you want to use the same stack across providers and all the proprietary technologies (even hidden from you with things like terraform) are suddenly loosing their luster.

[hnlmorg]

I don’t think any actually believes that.

What people usually think is “resilience up to a reasonable level of risk and cost”.

Multi-cloud is simply isn’t cost beneficial for 99.9% of problems.

And for a lot of businesses who talk about risk, saying “we followed AWS best practices but AWS went down” is an acceptable answer to the question of liability.

If you are in a position where AWS going down is a reasonable risk, then you’re already in a specialised enough domain to have engineers who understand how to deliver HA across different vendors.

[Roark66]

Depends on the product. I'd never guarantee 3 nines (a bit under 9h of downtime per year) to my customers of any product with moderate complexity without multicloud(except for a cloud provider issues that affect entire region).

On the subject of cost. Multicloud doesn't have to be expensive, but it requires foresight during the initial design not to lock yourself up in proprietary tech.

Things such as AWS cognito, lambda elastic beanstalk, SQS can make it completely uneconomical to do multicloud.

But let's say you run roughly equivalent services (load balances, DNS, vms, containers for example under GKE, s3 compatible object storage, etc) it doesn't have to cost an arm and a leg.

[doublerabbit]

I jest, Anyone who thinks multicloud will provide them full resilience is fooling themselves. You need colocated hardware for true high availability.

[stephenr]

> And IAM is not a wholly independent software stack: they rely on DynamoDB and a few other services, which in turn have a circular dependency on IAM.

When you dogfood your own Rube Goldberg machine.

[zaphirplane]

We should let the IAM service team know if this glaring gap the hn thread figured out /s

I’m 99% ;) certain dependencies of foundational services are a well discussed topic

[myroon5]

> outside of China

[Nitpick] There are a few more AWS partitions like GovCloud:

https://jasonbutz.info/2023/07/aws-partitions/

[dlenski]

Yes, I'm certainly aware of the other partitions. That's why I said all the public cloud regions outside China.

Yeah, "govcloud" is technically available to the public, although there are other partitions reserved for government use that are not, and the naming is a big hairy mess. Many service teams don't have any US-citizens-in-the-USA working for them, and they cannot in any way adequately support these regions.

My on-call experience improved significantly when I moved from the US to Canada, and I got taken off the (extremely thin!) list of engineers eligible to ssh into RDS instances in Govcloud. There were so few USA-citizen-in-USA engineers that I had been getting tickets for services and instances in Govcloud about which I had only the very thinnest knowledge… and then I was limited in my ability to consult with others who were actually experts. The customers in Govcloud paid a premium to be there, I got paged for a bunch of tickets which I was ill-prepared to handle, and it was generally a bad experience for everyone.

Working with the airgapped secret/top-secret partitions was even worse. You would get paged incessantly and then someone who was cleared for access but knew almost nothing about the service in question would have to go to a SCIF in the DC area, and you would exchange screenshots and text instructions with a turnaround time of hours or days.

[electroly]

Since this article was written, AWS also added European Sovereign Cloud as a partition: aws-eusc.

[sidewndr46]

Isn't this kind of circular dependency what lead to extended downtime a while back?

[superjan]

It reminds me of facebook. Staff was locked out of the office due to the outage they were supposed to fix.

[martin8412]

Luckily the plasma torch and bolt cutter didn’t require logging in with Facebook.

[jethro_tell]

It's basically what leads to extended downtime almost every time. There are just some things in the stack that are still single points of failure, and when they fail it's a mess.

[dlenski]

Yes, I concur.

Sometimes the circular dependencies get almost cartoonishly silly.

Like, "One of the two guys who has the physical keys to the server cage in us-east-1 is on vacation. The other one can't get into his apartment because his smart lock runs into the AWS cloud. So he hires a locksmith, but the locksmith takes an extra two hours to do the job because his reference documents for this model of lock live on an S3 bucket."

I made that example up, but only barely.

[somat]

We had a pair of machines. And some bright spark set them up to mount each others NFS shares. after a power outage "Holy mother of chicken and egg NFS hangs batman"

That was a weird job, fun, it was a local machine room for a warehouse that originally held the IBM mainframe, it still held it's successor "the multiprise 3000" which has the claim to fame as being the smallest mainframe IBM ever sold. But now the room was also full of decades of artisanal crafted unix servers with pick databases. the pick dev team had done most the system architecture. The best way to understand it is that for them pick is the operating system, unix is a necessary annoyance they have to put up with only because nobody has made pick hardware for 20 years. and it was NFS mounts everywhere, somebody had figured out a trick where they could NFS mount a remote machine and have the local pick system reach in and scrounge through the remote systems data. But strictly read-only. pick got grumpy when writing to NFS not to say anything about how the other database would feel about having it's data being messed with. Thus the circular mount.

Still was not the worst thing I saw. I liked the one system with a SMB mount. "Why is this one SMB?" "Well pick complains when you try to write to a NFS mount, but it's NFS detection code does not trip on SMB mounts." ... Sighs "Um... I am no pick expert but you know why it does not like remote mounts right. SMB does not change that, Do you happen to get a lot of corrupt indexes on this machine?" "yes, how did you know"

[roryirvine]

Oh, yeah, re-exporting NFS mounts via SMB was very much a thing in the early 2000s - something to do with their different approaches to flock() vs fcntl() handling. If you ran into locking issues with nfs, then re-exporting via SMB was standard advice.

At some point, the behaviour changed and locks starting conflicting. IIRC, we hit it when upgrading to Debian Etch and took the time to unwind the system and make pure NFS work properly for us. Plenty of people took the opposite approach, and fiddled with the config to make locking a noop on SMB. I know of at least one web hosting company who ended up having to restore a year's worth of customer uploads from backups as a result...

[wgjordan]

A real example, from Facebook's 2021 outage [1]:

> Our primary and out-of-band network access was down, so we sent engineers onsite to the data centers to have them debug the issue and restart the systems. But this took time, because these facilities are designed with high levels of physical and system security in mind. They’re hard to get into, and once you’re inside, the hardware and routers are designed to be difficult to modify even when you have physical access to them. So it took extra time to activate the secure access protocols needed to get people onsite and able to work on the servers. Only then could we confirm the issue and bring our backbone back online.

There was one (later denied) report that a 'guy with an angle grinder' was involved in gaining access to the server cage.

[1] https://news.ycombinator.com/item?id=28762611

[MichaelZuo]

Why would such a critical server even be accessible with only one set of keys?

I’ve always thought mission critical stuff needs two independent key holders, with key holes placed far apart enough to make it impossible for 1 person to reach both.

[bigfatkitten]

Other than for certain nuclear missile launches[1], that only happens in the movies.

[1] https://www.nationalmuseum.af.mil/Visit/Museum-Exhibits/Fact...

[MichaelZuo]

I dont know how it is in the datacentre industry, but certainly in other industries that is how its done for anything truly mission critical and also easily tampered with.

I guess it shows very few care enough to pay enough to make that a reasonable upgrade.

[michaelt]

They're not actually accessible with 'only one set of keys' in my experience.

You actually have to present your photo ID at the site entry gatehouse, then again to the building entry guard (who will also check you have a work permit and a site-specific safety induction) then you swipe a badge at a turnstile to get from reception into the stairwell, then swipe your badge at a door to get into the relevant floor, then swipe your badge and key in a code to enter the room with the cages then you use the key.

[sidewndr46]

a circular dependency and a single point of failure are not the same thing. If I have a single point of failure and it is down, I fix that and things work again. If I have circular dependency, there is no obvious way to fix anything that is broken any longer.

[grogenaut]

when you have a circular dependency, one strategy employed, is to have it be circular but interruptible for 18 or so hours. Call it an oh shit bar.

I'm glad I never had to get that deep into the failure chain.

[dlenski]

> Call it an oh shit bar.

Amazon's equivalent of this (sort of) was the "andon cord."

Not only was the physical metaphor that led to this name never properly explained (it's basically an "emergency stop" string in a Toyota factory), but the actual use of this mechanism was so heavily discouraged that I never saw it used in 4+ years at Amazon. except once very performatively by a VP who had already been paged awake at 2am or something like that.

In my experience, a lot of the AWS engineers live in continuous fear of screwing up by using the huge array of extremely powerful, dangerous, poorly-explained, and ever-changing tools that they have access to.

[zaphirplane]

Services outside of us-east-1 don’t call us-east-1 for IAM data plane thou right ?

[cmiles8]

They’re talking about the backbone and what goes on behind the scenes. There have been issues with services in other regions when us-east-1 has issues.

Folks built in other regions believing they were fully isolated only to discover later during an outage that they were not.

[master_crab]

IAM isn’t even really the most painful dependency. Route53 is. The control plane only runs out of use1.

Better make sure the only DNS operations you run during an outage are data plane queries and health check failovers.

[easton]

They actually kind of fixed this recently, you can ask them to move your route53 control plane to another region in the event of us-east-1 breaking: https://docs.aws.amazon.com/Route53/latest/DeveloperGuide/ac...

There’s a bunch of caveats but it’s worth enabling if you’re changing dns all the time (as most AWS networking doodads like to do).

[trollbridge]

Is there an architectural reason it’s not for replicas in the other AZs?

[jmsgwd]

> The idea that AWS's services are fully regionalized or isolated has always been a myth.

This is highly misleading. It's true that there's a handful of global AWS services - but only their control planes operate from a single region (e.g. us-east-1). Their data planes are regionally isolated or globally distributed.[1]

The only time you'd normally use a service control plane is to deploy changes, e.g. when you create, read, update or delete service resources or update configuration during a change window.

Workloads should be designed for "static stability", as recommended by AWS.[2] A statically stable workload only depends upon the data planes of the services it uses at runtime. Statically stable workloads are designed to continue operating as normal even if there's a service event impairing one or more control planes (including for global services).

> During us-east-1 outages it's sometimes possible to continue using existing auth tokens or sessions in other regions, while not possible to grant new ones.

This is just plain wrong! The IAM Security Token Service (STS), which grants IAM tokens, is a data plane-only service and runs independently in each region [3]. The IAM data plane, which enforces access control, is also regional.

If the IAM control plane is impaired, you might not be able to create new IAM roles (a control plane operation) - but you can continue generating and using temporary credentials for existing IAM roles (data plane operations) within the region your workload is running in. This allows statically stable workloads to continue using IAM without interruption.

[1] https://docs.aws.amazon.com/whitepapers/latest/aws-fault-iso...

"Global AWS services still follow the conventional AWS design pattern of separating the control plane and data plane in order to achieve static stability. The significant difference for most global services is that their control plane is hosted in a single AWS Region, while their data plane is globally distributed."

[2] https://docs.aws.amazon.com/whitepapers/latest/aws-fault-iso...

"...eliminating dependencies on control planes (the APIs that implement changes to resources) in your recovery path helps produce more resilient workloads."

[3] https://docs.aws.amazon.com/whitepapers/latest/aws-fault-iso...

"STS is a data plane-only service that is separate from IAM, and does not depend on the IAM control plane."

[dlenski]

You're right of course to distinguish the control plane and data plane, and it sounds like you know more about this than I do for IAM.

I disagree, though, that my post was "highly misleading" despite this omission.

As a practical matter, some services fail to achieve the "static stability" you describe, in terms of not depending on other services’ control planes.

And also, many on-calls ops and firefighting tasks (to say nothing of canaries and other automated tests) depend on other services’ control planes.

And above all, many AWS engineers (myself very much included even after years there) don't have a clear understanding of the boundaries of other services’ control planes. https://news.ycombinator.com/item?id=48078254

> > During us-east-1 outages it's sometimes possible to continue using existing auth tokens or sessions in other regions, while not possible to grant new ones.

> This is just plain wrong! The IAM Security Token Service (STS), which grants IAM tokens, is a data plane-only service and runs independently in each region.

I didn't mention STS in the service to which you're responding. The service that I worked on the most, RDS, required ssh'ing into live instances to solve basically all non-trivial problems (I'd guess 80% of the tickets that I saw actually resolved required it). And I have no idea if it how STS was involved in generating the ephemeral Midway-signed ssh keys required for it… but whenever there were us-east-1 IAM outages we'd have big problems opening new sessions, while less-capable web-console-based ops tools with long-lived credentials would keep working.