A joint statement from UniSuper CEO Peter Chun and Google Cloud CEO Thomas Kurian
8 May 2024
UniSuper and Google Cloud understand the disruption to services experienced by members has been extremely frustrating and disappointing. We extend our sincere apologies to all members.
While supporting UniSuper to bring its systems back online, Google Cloud has been conducting a root cause analysis.
Thomas Kurian has confirmed that the disruption arose from an unprecedented sequence of events, where an inadvertent misconfiguration during provisioning of UniSuper’s Private Cloud services ultimately resulted in the deletion of UniSuper’s Private Cloud subscription.
This is described as an isolated, “one-of-a-kind occurrence” that has never before occurred with any Google Cloud client globally. This should not have happened. Google Cloud has identified the sequence of events and taken measures to ensure it does not happen again.
Why did the outage last so long?
UniSuper had duplication across two geographies as protection against outages and data loss. However, the deletion of the Private Cloud subscription triggered deletion across both geographies.
Restoring the Private Cloud required significant coordination and effort between UniSuper and Google Cloud, including recovery of hundreds of virtual machines, databases, and applications.
The instant cascading worldwide deletion upon closing or deleting a subscription sounds like a recipe for disaster. Why not mark it for deletion and delete say... a day or a week later?
It has been 0 days since GCP has taken down a startup (again).
You see this at least once a year. Never heard of this from AWS or Azure.
In all seriousness, this is why we don't use them. They have the most ergonomic cloud of the big three, then absolutely murder it by having this kind of reputation.
On the other hand i can’t remember when there was a serious outage on GCP, unlike AWS/Azure who seem to go down catastrophically a couple of times per year.
GCP has a lot of customers. But you wouldn't know the companies that do, unless you worked there and wanted to leak it, or it publicly comes out. Eg it's been publicly acknowledged that Apple uses GCP for iCloud, https://www.cnbc.com/amp/2018/02/26/apple-confirms-it-uses-g... , and Home Depot is another that's used as a case study, https://cloud.google.com/customers/the-home-depot but most customers don't want to make a big deal about being on GCP as it's none of our business who's hosting them.
I've been in AWS for almost twenty years at this point. It's been a long time since I've seen a global outage of the data plane on anything. The control plane, especially the US-east-1 services? Yes - but if you're off of east-1, your outages are measured in missile strikes, not botched deployments.
The impacts are usually partial. For example, scaling is impacted but everything already deployed contributes to work up to capacity. Or, you can't change configuration but the previous configuration works as configured. Often surprisingly not so impactful even if there can be limited work stoppage.
AWS goes down catastrophically but are back up in minutes/hours most of the time (as long as they aren't down because Iran blew up their data center). That's obviously REALLY bad for certain industries, but I suspect for the vast majority of their customers it's not a big deal. We've been able to isolate the damage almost every time just by having AZ failover in place and avoiding us-east-1 where we can.
There was a pretty bad one last summer - their IAM system got a bad update and it broke almost all GCP services for an hour or so, since every authenticated API call reaches out to IAM.
It had lasting effects for us for a little over 3 hours.
I still remember the one where they nuked all the storage of I think an Australian insurance company I think, luckily the it department had done a multi cloud setup for backups
During my 5 years of my startup, we had only 1 outage due to AWS because we picked us-west-2 as the primary reason. If anyone starting a company and picks us-east-1 as the primary reason, they should be fired. There's absolutely no reason to be in that region.
That’s an entirely different type of problem, and avoidable by just using us-east-2 (I still don’t understand why people default to us-east-1 unless they require some highly specific services).
Is it that easily avoidable? A lot of AWS's control plane seems to have dependencies on us-east-1, or at least that's what it's looked like as a non-us-east-1 user during recent outages.
It's AWS and Azure that are the outliers and tend not to care too much what their customers do with their infrastructure. AWS is perfectly fine with allowing me to run copies of 15 year old vulnerable AMIs copied from AMIs they've long since deprecated and removed. Even for removed features like NAT AMIs.
The only anecdotal thing I've seen is we hired a vendor to do a pentest a few years ago, and they setup some stuff in an AWS account and that account got totally yeeted out of existence by AWS if memory serves.
You should not be conducting unauthorized penetration tests against third party infrastructure providers without permission. They have processes and systems and usually just wants a heads up of what you plan to test and t the duration / timestamps.
Cuz otherwise you look like a threat actor.
That’s assuming your vendor was pentesting AWS systems. If you meant you hired a vendor to pentest your own systems on AWS, that’s of course a totally different matter.
>That’s assuming your vendor was pentesting AWS systems. If you meant you hired a vendor to pentest your own systems on AWS, that’s of course a totally different matter.
Sorry for being unclear, the vendor was attacking our organization only, and any other company was expressly forbidden in the contract. As I recall it was a fake SSO sign-in page to collect credentials that they would try and social engineer our employees with.
Having done this for both Azure and AWS, there's a specific ticket that needs to be filed with each provider that documents the scope of your pen test, where you're coming from, and a time frame over which you're doing it (which ISTR was "not more than 24 hours")
Yup, I thought it was great. Although one concern I always had in the back of my mind was where is the line drawn. Such as if an adversary gains access to one of my orgs accounts and does something similar, do we get 100% taken out.
How the heck do these things happen, especially with companies with huge monthly spend? At my last job we had some suspicious workloads running on AWS and our TAM reached out to us before taking any action. Who wants to bet this was some AI automation gone wrong and because GCP seems to be allergic to actually contacting a human to get a response, this just sits in some support queue that outsourced workers look at after a few hours just to give a canned response?
Nothing surprises me with anything related to support on GCP. While we absolutely do not need them, I have been through no less than 12 different Account Executives over the last 6y and they're all ENTIRELY and COMPLETELY useless.
They all introduce themselves, beg me to setup a meeting w/them and some sort of engineering resource(s), and they come to a meeting with a canned slide deck that is so absurdly unrelated to us that I just laugh, and then the next time I hear from them it's because we have a new AE.
This is my most recent reply (right after Next '26):
> I really appreciate you reaching out; however, we have met with, I dunno at this point, more than a dozen GCP Account reps, execs, technical teams, etc over the years and there's little to no value for us or you, now or in the future. Please do feel free to invest your time on your other clients. We're good; truly.
I love GCP and its services; we have been very pleased with it over the years, but the human side of it? Fucking sucks and I just don't see why they even bother.
Don't know about GCP, but our AE on AWS was also continuously rotating, and as best I can tell, their job was to figure out what we are planning to build, and to ensure that we should always use <INSERT AWS SERVICE DU JOUR> for that, rather than a competitor product or build it ourselves.
Before I just cut them off entirely, I used to tell them my primary concern was cost savings and that I wanted them to recommend ways I could cut 25% off my bill every month and watch the glorified salespeople fumble over trying to avoid that conversation.
It’s ok though, Claude helped us cut >45% of our monthly costs. I’m surprised they haven’t been beating down my door after we made that level-shift. Probably in AE transition. ¯\_(ツ)_/¯
For what it's worth - I'm not sure what the criteria is (I assume we're "medium sized / not a big upsell opportunity"?) - our GCP rep quickly pushed us to switching to using a GCP reseller. They took over our billing so that we can pay via ACH, and provide both free first-line support/escalation and paid engagements for bigger projects; they don't charge a premium on top, apparently Google pays them for supporting us. Hasn't made much of a difference in how we operate, but at least we have a direct-ish line for issues when they come up.
huh- I guess there are two HN submissions with meaningful replies...
I said this in the other thread, we got access to our account back, but even with a Account Rep. and a CSM on our account- it still took them a while to figure out what was going on.
I'm sure it could have been worse if we didn't have a rep on our account.
What does blocked mean? Is there a different post that I am missing? There is shared infrastructure in GCP for networking (ex-googler here) and if only railway is affected, then it is not clear if it is only GCP or if there is something from Railway's perspective that needs to be addressed.
As someone who runs some public APIs, the amount of spam from Railway IPs is insane. They have horrible abuse prevention. Hopefully this encourages them to improve their operations.
I will never leverage GCP in an enterprise setting, it's honestly amazing how hard they fumble the bag. Will be interesting to see when GCP support started working with them, from the updates there was an hour and change from when they identified the issue and GCP support was confirmed.
In the cloud space it seems like AWS does nothing and wins.
Is google allergic to humans or something? Cannot they just send an email or call the company before taking a wrecking ball to the entire company's infra? Are they stupid?
Well, as a 2 week tenured and very happy Railway customer until now, I am now a Render customer. Somehow DNS cut over within 1 min(!) and live after about 30 minutes of work. Not bad!
In my experience, DNS changes are a lot faster than they used to be. There’s some website that has a map that tries to resolve your domain with a bunch of name servers around the world that was pretty neat to look at last time I migrated something.
This is bad. Even their own website is down at railway.com. Looks like total dependency on google cloud. Surprising for a company of their scale with all this VC money.
Does anyone know how this even happens inside the walls of google? Is it an automated process? How is such a (presumably) high revenue account just magically blocked without human intervention? I'm quite perplexed.
There would have been efforts to contact them, but it would have been via their contact method, aka the email they set it up with.
Common ways this happens? They are using a credit card to run their business with no backup payment method. Then the company's contact person is on vacation.
Yeah, I'm not sure what to think here. We know Google is not the best at customer service and has automated account suspensions. But, what I'm curious about here is why this happened.
Railway hosts applications for customers. An uneducated guess for some possible reasons: 1) one of those customers hosted something they shouldn't have 2) railway had something spawn that took up too many resources 3) Or their account balance was too high 4) Or something...
But all of this probably culminates in someone needed to read an email that was missed.
Scaling a customer infrastructure setup like Railway is hard. This is one of the non-technical hard parts - how to make sure your account with your primary vendor is safe. But, I'm willing to wait to pass judgement here until more information is available. I'm sure the post-mortem will have lessons. I'd like to know more.
Honestly still insane to nuke a high-volume client's business after a single payment issue. There would be no reason for Google to believe that a single hiccup like that is evidence that they won't get paid and have to cut account access immediately.
The 3-2-1 backup rule is pretty outdated in the world of cloud. You could have 3 complete copies of your data in different S3 buckets, but if they're all under the same account you've lost your blast radius protection
Well having backups help, but I certainly can’t migrate my infra to rsync.net on moments’ notice (or ever since rsync.net does storage and nothing else) so my customers aren’t affected.
I've managed several accounts with GCP over the years and I've always maintained a great relationship with our contacts there. Some of these accounts were quite small, on the order of <$20k/mo, and even then we were kept abreast of anything that might be cause for concern. I always maintain a standing biweekly meeting with at least someone on the other side (account exec, technical staff, whatever) and I've yet to be blindsided by anything.
Is Google's communication good? No, not particularly. The only way something like TFA happens is if the relationship is neglected (by one or both parties). I'm not saying Railway did something wrong, but there are usually many flags and opportunities to correct long before drastic actions.
I get the impression that Railway plays fast and loose with a lot of their limits and resources and that Google may not be a fan of that.
Edit: would also like to say that if you put all your resources in one GCP project you are going to have a bad time. If you organize stuff over many projects it is very unlikely that they will ever take account wide action. I've had issues with, for example, a particular tenant's behavior, but it never jeopardized the other tenants.
I don't think you can ever trust one service with critical data. Some Claude instance deletes your prod database, you have to restore from an offsite backup because it also deleted your local backups. Even at small startups we did pg_dump to AWS from GCP because ... who knows what is going to happen to GCP, and we want to continue to be in business if that happens.
I don't feel safe with any one single point of failure. "Your credit card bounced", "you thought it was dev", "you got hacked", etc. are all the same problem to me and no cloud provider solves those merely by setting up an account.
At this point you can’t trust Google anymore, it keeps breaking things. Imagine having Google AI do this thins automatically. Will have apocalypse in in a day.
Wild to me that any tech sector business would want to rent an operating environment to park their entire infrastructure into. This is the equivalent to traveling shoe salesmen setting up a tent in the parking lot of a strip mall.
I didn’t knew Railway so with this misleading headline I thought a Google Cloud data centre was being built in the way of a railroad. That’d been a funny story to read..
There's a lot of, what seems to me, unfounded blame being directed at Google for this. Isn't railway the company that just blamed Anthropic for deleting their prod database?
Nope, Railway was the company who was hosting PocketOS, which is the company that blamed Cursor for deleting their prod database. Railway is only involved insofar as their API allowed an instant delete of the prod database.
Why does Railway deserve any blame here at all? It was an MCP with elevated infra access, that the user willingly connected through Cursor, which allowed an LLM Agent to manage infra on Railway. The user would first have gone through oAuth confirming the access level scope. So obviously it has access to all commands the user would also have access to. From my perspective the blame is entirely on the user, and partly on Cursor for not enforcing HITL correctly across their agents.
Putting AI aside, people make mistakes. One of the most common mistakes people make is deleting the wrong thing. After they realize the mistake, people want to restore the thing they deleted from backups. Thus deleting the thing and deleting the backups of the thing should always be separate operations.
fairly certain you are remembering the goofy article that was going around where a railway user allowed an agent to delete his db. iirc he questioned the agent after and the agent told him it should have read the file that told him not to do things, so just sounds like he deleted his db and blamed his tools.
Yeah but until you find that the new cloud provider won't approve your compute quota or doesn't have enough capacity in the region or you hit fraud flags for stagnant account spinning up lots of compute.
How to handle domains? The rest is easy, but your domain registrar blocking you sounds like a pain. My current solution is to use a local small provider, just for the domain. Then if there is a problem with your play account it is out of any blast radius.
Looks like they were sold at the beginning of the year to a company without a Wikipedia page whose parent company doesn’t have one either https://en.wikipedia.org/wiki/Markmonitor
Acquired in November 2022 by Newfold Digital, it was later announced that the firm would be sold to Com Laude, a company owned by PX3 Partners.
PX3 stands for purpose, passion, and performance. It is a pan-European private equity firm with headquarters in London. It invests behind transformative themes and targets companies operating within select segments of the business services, consumer and leisure, and industrials sectors with strong business fundamentals.
https://www.unisuper.com.au/about-us/media-centre/2024/a-joi...
A joint statement from UniSuper CEO Peter Chun and Google Cloud CEO Thomas Kurian
8 May 2024
UniSuper and Google Cloud understand the disruption to services experienced by members has been extremely frustrating and disappointing. We extend our sincere apologies to all members.
While supporting UniSuper to bring its systems back online, Google Cloud has been conducting a root cause analysis.
Thomas Kurian has confirmed that the disruption arose from an unprecedented sequence of events, where an inadvertent misconfiguration during provisioning of UniSuper’s Private Cloud services ultimately resulted in the deletion of UniSuper’s Private Cloud subscription.
This is described as an isolated, “one-of-a-kind occurrence” that has never before occurred with any Google Cloud client globally. This should not have happened. Google Cloud has identified the sequence of events and taken measures to ensure it does not happen again.
Why did the outage last so long?
UniSuper had duplication across two geographies as protection against outages and data loss. However, the deletion of the Private Cloud subscription triggered deletion across both geographies.
Restoring the Private Cloud required significant coordination and effort between UniSuper and Google Cloud, including recovery of hundreds of virtual machines, databases, and applications.
You see this at least once a year. Never heard of this from AWS or Azure.
In all seriousness, this is why we don't use them. They have the most ergonomic cloud of the big three, then absolutely murder it by having this kind of reputation.
Azure nerfed the front door of all Azure and O365 services last year.
All it these companies are great at what they and occasionally fuck up.
https://status.cloud.google.com/incidents/ow5i3PPK96RduMcb1S...
They had a really bad global outage a year ago. At least with AWS outages are contained to a single region.
It had lasting effects for us for a little over 3 hours.
AWS does it more efficiently; it takes down many startups at a time when us-east-1 goes down.
https://news.ycombinator.com/item?id=46731498 https://news.ycombinator.com/item?id=33360416
Then I recall https://news.ycombinator.com/item?id=45798827
https://news.ycombinator.com/item?id=33737577
It's AWS and Azure that are the outliers and tend not to care too much what their customers do with their infrastructure. AWS is perfectly fine with allowing me to run copies of 15 year old vulnerable AMIs copied from AMIs they've long since deprecated and removed. Even for removed features like NAT AMIs.
The only anecdotal thing I've seen is we hired a vendor to do a pentest a few years ago, and they setup some stuff in an AWS account and that account got totally yeeted out of existence by AWS if memory serves.
Cuz otherwise you look like a threat actor.
That’s assuming your vendor was pentesting AWS systems. If you meant you hired a vendor to pentest your own systems on AWS, that’s of course a totally different matter.
Sorry for being unclear, the vendor was attacking our organization only, and any other company was expressly forbidden in the contract. As I recall it was a fake SSO sign-in page to collect credentials that they would try and social engineer our employees with.
They all introduce themselves, beg me to setup a meeting w/them and some sort of engineering resource(s), and they come to a meeting with a canned slide deck that is so absurdly unrelated to us that I just laugh, and then the next time I hear from them it's because we have a new AE.
This is my most recent reply (right after Next '26):
> I really appreciate you reaching out; however, we have met with, I dunno at this point, more than a dozen GCP Account reps, execs, technical teams, etc over the years and there's little to no value for us or you, now or in the future. Please do feel free to invest your time on your other clients. We're good; truly.
I love GCP and its services; we have been very pleased with it over the years, but the human side of it? Fucking sucks and I just don't see why they even bother.
It’s ok though, Claude helped us cut >45% of our monthly costs. I’m surprised they haven’t been beating down my door after we made that level-shift. Probably in AE transition. ¯\_(ツ)_/¯
I said this in the other thread, we got access to our account back, but even with a Account Rep. and a CSM on our account- it still took them a while to figure out what was going on.
I'm sure it could have been worse if we didn't have a rep on our account.
Implement anti-abuse measures and you will hit some loud false positives (this may be the case with GCP here).
I don't envy anybody running a hosting co - the internet is a really ugly place under the surface.
edit: to add - AWS are really good here. Must be the ~30 years of retail fraud and abuse experience.
> The fact of the matter is, you simply cannot build a cloud on someone else’s cloud.
Indeed…
[0] https://blog.railway.com/p/launch-week-02-welcome
In the cloud space it seems like AWS does nothing and wins.
Common ways this happens? They are using a credit card to run their business with no backup payment method. Then the company's contact person is on vacation.
Sign up for terms. It will get you payment terms!
Railway hosts applications for customers. An uneducated guess for some possible reasons: 1) one of those customers hosted something they shouldn't have 2) railway had something spawn that took up too many resources 3) Or their account balance was too high 4) Or something...
But all of this probably culminates in someone needed to read an email that was missed.
Scaling a customer infrastructure setup like Railway is hard. This is one of the non-technical hard parts - how to make sure your account with your primary vendor is safe. But, I'm willing to wait to pass judgement here until more information is available. I'm sure the post-mortem will have lessons. I'd like to know more.
If it's anything like AWS, that may be just one of hundreds of emails they send every day, most of which are just noise.
… on the Unix command line …
… to a cloud older than AWS…
… if only …
I'm exaggerating but someone said they got "auto banned"
what if that happens to a small account which hosts some really important data/services there?
Is Google's communication good? No, not particularly. The only way something like TFA happens is if the relationship is neglected (by one or both parties). I'm not saying Railway did something wrong, but there are usually many flags and opportunities to correct long before drastic actions.
I get the impression that Railway plays fast and loose with a lot of their limits and resources and that Google may not be a fan of that.
Edit: would also like to say that if you put all your resources in one GCP project you are going to have a bad time. If you organize stuff over many projects it is very unlikely that they will ever take account wide action. I've had issues with, for example, a particular tenant's behavior, but it never jeopardized the other tenants.
Pray to @dang that you will make the front page of HN?
Agreed. Railway are probably not far off a billion dollar company though!
I don't feel safe with any one single point of failure. "Your credit card bounced", "you thought it was dev", "you got hacked", etc. are all the same problem to me and no cloud provider solves those merely by setting up an account.
"Absolutely. The Railway network is a mesh ring between AWS, GCP, and Metal
So: - High availability interconnects - High availability path routing between clouds - Database itself is high availability
However, Google's VPC itself is not. So we will add a shard to Metal and AWS"
https://x.com/JustJake
Looks like they were sold at the beginning of the year to a company without a Wikipedia page whose parent company doesn’t have one either https://en.wikipedia.org/wiki/Markmonitor
-Edit-Private equity apparently https://px3partners.com