Less reliable products aren't allowed to become bottlenecks in the first place, at least at any B-tier or better businesses.
Also, is anyone else getting the bitter taste of AI writing from this page?
Nah. Jim is just a decent writer (and historically has been fairly suspicious of AI)
Skip to the last 2 paragraphs
Which is to say that the significance of downtime depends on the user. Talking about nines only makes sense internally when you are evaluating your infrastructure and operations. It doesn't tell you squat about impact to your customer.
root@vixen:/fountain/crystals #
*** FINAL System shutdown message from dblrabbit@ ***
System going down IMMEDIATELY
System shutdown time has arrived
root@vixen:/fountain/crystals # uptime
3:05PM up 1931 days, 18:13, 0 users, load averages: 1.01, 1.03, 1.41
root@cookie:/srv/users/dblrabbit # uptime
3:07PM up 1931 days, 16:59, 1 user, load averages: 1.76, 1.17, 1.06
root@cookie:/srv/users/dblrabbit # poweroff
Shutdown NOW!
poweroff: [pid 47177]> GitHub Actions: 12 hours affected in the last 30 days (98.31% uptime).
This is trying to shine the most favorable possible light onto a deteriorating situation. It doesn't take away from the fact that most businesses have measurable missed revenue in downtime. Customers that shop somewhere else, ads that were never severed, leads that grew a little colder. 12 hours of downed GitHub results in millions of dollars of lost developer productivity that was externalized by Microsoft to other companies.
We shouldn't be trying to spin downtime as "just a few hours a month." Those hours cost real dollars.
Is trying to spin anything. It’s making easier to see the impact over the last 30 days. I agree with the article
Not all hours are created equal when it comes to downtime and my intuition is that most of these 12 landed within my working hours.
In terms of impact that then might mean they were down for 7.5% of the time I needed them, or had business hours uptime of 92.5% which is… both not very good and very disruptive.
On the other hand, downtime at 4AM would be much less impactful even if it happened every day and added up to more overall downtime.
You can obviously compute this for a particular customer, but being a global service, it's pretty much guaranteed that someone somewhere experienced the worse of those numbers
This is important because it's quite possible that the downtime is biased toward the times they have the most active users.
https://cloud.google.com/blog/products/gcp/available-or-not-...
They're also completely irrelevant, you as a customer of a service that is down can lose the same amount of money in a 5 minutes outage or 30 days outage, if you were only relying on this service for one operation that took 1 second and had to happen during the time where the outage happened.
Depending on the service in question, no amount of downtime is acceptable, however unrealistic this is.
A lack of ads is an argument for more downtime.
Keep the percentages, and regardless of that - GitHub fix your uptime
Some measure quite detailled but some just don't summarize the downtime from all providers up and below their own platforms.
For example we had a 6 9 (99.9999%) requirement from a customer for any given 3-6 month period. If we violated that, we owed them their money back (baring the outage wasn’t caused by us - I.e our cloud provider shit the bed).
That’s something like 7.5 seconds. For a contract over $1.5M. Am I the only one who thinks that’s outrageous expectations?
EDIT: The web app was for generating SBOMs of static assets.
This particular case was in cybersecurity- specifically static analysis of assets, for the purpose of providing a SBOM.
If you agree to those terms knowing it's unrealistic, you're agreeing to give away your service for free.
Well, someone on the business side of the house is free to negotiate. Whether engineering learns about the contract before sales has inked a 6-nines availability guarantee varies wildly by the company
See also: selling features that don't exist and cost more to implement and maintain than the contract is worth.
If it wasn't prorated anyone who approved the contract needs training and/or firing. If it is prorated, that is generally not a problem. Small outages aren't even worth the effort of trying to get the money back, and if you have a large enough one to make it worthwhile it is likely the prorated refund is still going to be laughably small.
But six nines gives you 7.9 seconds a quarter. If you run a multihost system, that translates to ~1s dead host detection and switch and 3-4 switches per quarter. It's acheivable with reliable hardware and reasonable software. Otoh, it's very hard to hit if you need to move traffic to a different location to respond to a no notice location failure (failed automatic transfer switch, all fiber paths severed by construction because the redundant paths were in the same bundle, etc). If you have an out for 'cloud provider failure', that probably covers location failures.
Often times a tight uptime promise like that also comes with maintenance windows. Depending on the application, degraded service or no service may be acceptable within the maintenance window.
These companies are happy that you don't know the difference between 99%, 99.9%, and 99.99% and that you think they all sound pretty good.
One vendor in particular we deal with has a powerful feature which we use to a large extent. Unfortunately, that particular feature is all too often not working. The servers are up and the rest of the platform is working, but we need that feature, so if it's down, it doesn't help much that the rest of the platform is up.
If you do something 100 times a day against a four-nines service, you can reasonably expect that everything will succeed.
If you do something 10,000 times a day against a two-nines service, you can expect to hit a substantial number of errors during that day, or even have long periods where your work cannot happen at all.
People aren't frustrated with Github because Github has 98% uptime or whatever the specific number is. They're frustrated because it regularly interferes with their ability to work. The 98% number is just a concise way to say it.
Before 9's became a thing, people built systems that required outages, and those outages would happen outside of business hours.
However that's also why some transactions had to complete the next business day after they were registered. Because the whole business was running on offline processing (aka batch processing) that could be interrupted for upgrades, but had to be completed by the start of business the following morning.
My dad did one of those jobs, and my brain has made a bigger deal out of the times I awoke in the middle of the night to find him on the phone at the kitchen table at 2 am dealing with a war room call because an upgrade broke things that needed to be done in 5 hours. It probably only happened 3 times that I know about, and I probably knew about at least half of them, but it felt like it happened twice a year.
Also, 0.1% downtime in the form of a 45-minute outage per month is very different from 0.1% of requests failing in brief bursts. You often see downtime reported as "increased error rates" which is so vague as to be meaningless.
Google's "windowed user-uptime" attempts to deal with this a bit better, by exposing different views of the data instead of trying to condense uptime into a single number: https://www.usenix.org/system/files/nsdi20-paper-hauer.pdf
$ python3
Python 3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import math
>>> def nines(num):
... return -math.log10(1-num)
...
>>> nines(.9)
1.0
>>> nines(.99)
1.9999999999999996
>>> nines(.999)
2.9999999999999996
Which then smoothly covers the entire space:
>>> nines(.9321)
1.1681302257194985
>>> nines(.2)
0.09691001300805639
If I'm tasked with getting a tool to be used at least 2x as much, my first task is to reduce the failures per 1000 runs by 4x. In that way, if adoption increases by 2.5, 3x instead of the the minimum we are looking for, then the number of errors reported per week still goes down instead of up.
We like to think of things as percentages but the moment they start increasing in the time domain everyone gets mad, because they asked for and received the wrong thing.
And that's on top of the fact that when people are 'forced' to use a tool, any errors they experience will be lumped onto the Learned Helplessness theater they've been engaging in to avoid being made to use a tool that is going to make all of our lives easier.
The suggested format is equally unhelpful.
You can get 12 hours of downtime by being down once for 12 hours, or 144 times for 5 minutes. The user experience is VERY different in those two cases.
Ultimately the graphs are the most useful format.
I am now _required_ to consult status page of github, circleci or MS services etc because i need to know why a build is not passing, why i cannot open a repo, why is my work stalling.
Percentages matter, it is just so much more obvious why they matter when it comes down to important pieces of the internet like github. And i highly doubt the number of 12 hours in the last month. MS has been downplaying the issues they have with GH performance for a while now and i don't think it is time to start to believe them yet. Maintaining these pieces of infrastructure is responsibility and a burden.
Overall i would be careful with "nonlinear significance of numbers near 100%" we are talking gh being well into the 90's this year and one number that infra people are also often being reminded about is that "1% is 3.5 days".
Things are tough for gh people and i feel for them but they are not a startup or a underdog of some sort to receive sympathy in that case.
I've yet to find a status page that wasn't lying about the actual status.
Also 97% up is bullshit for the 3% of people who are offline.
Saucelabs was doubly bad for this because I'm absolutely certain based on traces that they had some sort of demux bug where they would send events from their tunnel to the wrong job. I could see it in the logs that a test timeout was often the cause of an event firing that was looking for something that never happened, because the event immediately preceding it in the script was never fired. Which meant it was either dropped or went somewhere it shouldn't.
Then it stopped one day and there was nothing in their release notes about it. Lies compounded by further lies.
That's just the most memorable example I have. Stuff like this happens all the time and with many services it plays out the same. There's a perverse incentive not to be transparent about problems with the service, so the status pages play down the intensity of the situation.
What's going on at Microsoft? Are they just copy-pasting their github issue reports into copilot and hitting send it without doing code reviews?
It was the interview equivalent of the multi-headed dragon meme, where the last one looks absolutely stupid. The contrast was insane, microsoft was an absolute shit show compared to the other two companies in terms of talent, personality, organization and more.
I think that illustrates the author's point quite well. 99% uptime sounds good, but when you think about a 3+ day outage that doesn't sound very good. Imagine Facebook or TikTok being down for 3 days.
Of course most of the time it's not all one outage, but a bunch of short ones. Still, it might communicate the impact better, especially depending on the argument you're trying to win.
Three nines reliability is great for most purposes. 8 hours downtime a year.
If your system produces money at a constant rate, it captures 99.9% of the available money. Even two nines or one nine might be pretty good on that basis, when the alternative is spending 2x or 10x as much - let's build another unreliable system with that money that captures some other independent market opportunity.
Poor reliability is a problem where you need to chain many systems together, or where the cost of a single failure is very large compared to a success. Or - as happens commonly because of load - if your periods of unreliability are correlated with periods of maximum opportunity, like an e-commerce site failing on Black Friday or a trading system failing when the market is most busy. But if you don't have one of those cases, evaluate whether investing in reliability is actually worth it to you.
GitHub is an example where two nines of reliability ought to be OK. The argument against it is that it's bad marketing to have an unreliable service, especially one aimed at software engineers. And if GitHub is largely a marketing play by Microsoft anyway (do they really make back its cost in enterprise subscriptions?) then marketing considerations need to drive its reliability.
Probably I already ask them to spend 2+ days attending HR or compliance training or listening to senior management tell them about sales targets.
But the point is, that 1% extra productivity requires the sometimes staggering cost of making the software 10x or 100x more reliable.
A short outage can snowball very easily in a lot of lost time. What I learned when working with enterprises is that above all they value reliability. This is for a reason.
A short outage might at best trigger loads of paperwork for multiple hierarchies, big meetings etc. The org has no choice. It needs to evaluate if whatever happens is a threat to their business.
In the worst case it is that, plus missing some crucial windows of delivery. This is because a system that is unavailable for a short time can cause backlogs that, like traffic jams, cascade as everyone has to slow down and then synchronously speed up again.
Orgs have the option to create more resilience, but that is overhead similar to compliance. You need to drill all your backup plans all the time, otherwise they are worthless. The drills cost time and money. At scale it is infeasible to be robust to all failures. Therefore, enterprises (at least) often prefer reliable systems over sophisticated systems. Because this delegates the risk management to the vendors rather than adding an overhead to every employee. Because at some point the employee would just do drills all the time instead of work.
Downtime really matters if it's at a time you need something to be up, and Github is big enough to have users for that to be all the time. That moves the conversation from 'It's down for a few hours a month' to 'Github is failing a significant number of it's users'.
One thing I have noted over time is a lot of these AWS, Azure et el downtimes is they occur in the middle of everyones day, millions of people are impacted by them. Same with github its getting in the way of work. Whereas when we hosted services on our own equipment the downtime was usually out of main hours. The percentages are in many ways the wrong measure of downtime because hours aren't equal in impact to businesses.
The problem with load balancing also shows up in employees mass quitting. It's the same queuing theory problem but I don't know the name of it.
Essentially, when you run a set of 'resources' near capacity, if one fails then the next most highly loaded one also goes past capacity and also fails. The work keeps getting foisted onto other units that also fail.
In retail, the beleaguered employee quits in frustration, and the three other people who can do the same onerous tasks the quitter did now find their jobs have gotten 33% worse with no extra remuneration for the insult, and one of them quits too. The last one or two simply refuse to do the task more than they used to, and the boss can't threaten them with being fired or reported for it because if they quit then the business ends. And still the whole thing goes off the rails.
That's why consistent hashing divides the work of one failed node and spreads it evenly over the remaining nodes, so that a spike is softened as much as possible instead of being concentrated on one or two fallbacks who also immediately fail. It's not that it's a great solution, it's that it's the least bad option amongst some pretty terrible ones.
The solution is to undersubscribe the hardware, which some beancounters hate. But most bean counters don't understand Queuing Theory.
Similar for LLM measures from an ideal 1.0 mark.
In the electric utility world we have a few IEEE standardized metrics (with appropriately IEEE'd acronyms) for tracking service reliability that I like much better and always wish for when I'm looking at a status page. Pie in the sky stuff for sure, nobody wants to do this analysis and publish the results without a regulator telling them have to, but c'est la vie.
SAIDI - System Average Interruption Duration Index. How many minutes an average customer experienced service interruption in a year. This is the big one I'd want to see on your service status page IMHO. For the power grid, we consider any outage longer than five minutes to be an interruption ("non-momentary outage").
SAIFI - System Average Interruption Frequency Index. How many total periods of interruption occurred for the average customer in a year.
CAIDI - Customer Average Interruption Duration Index. How long it takes service to be restored for the average customer when there is an interruption.
For the US, here is what these numbers look like: https://www.eia.gov/electricity/annual/html/epa_11_03.html. If you're outside the US look up yours and have a good laugh at us. :)
For the "right now" aspect you have probably visited your utility's outage map, but here I would say we do much better than most utilities. The level of detail on the investigation and resolution is often more detailed, and we usually know better than to bother providing much in the way of a concrete estimate for restoration time of a current outage (though this is getting better in the utility space).
If you want to feel better about that, check out South Africa. You know you have a problem when "load shedding" is a household word.
(I should note that things have improved recently: it was a big milestone when they went a full year without load shedding, as of May 16.)
Point the AI at your logs, tell it to fix things, rinse and repeat. It's faster than debugging, and fast is good.
I think the point is that the metric being meaningless right now to most people makes it easier to shift the reliability-Overton-window. 99.9999 vs 98.0 seems not to bad to a layman (I.e. executives), but 4m19s to 12 hours seems pretty intuitively bad. Sure, we may lose some shift in the immediate future, but it's easier for that to continue happening with just percentages, is the point.
It's not like we have to stop showing percentages as well, but time is a much more accessible expression. Right now the lay-ness of company leadership has already allowed the shift to happen pretty markedly.
This attitude of modern tech claiming 98% is good just doesn’t work in the old tech acceptance. We had individual components fail all the time. We’re still looking at a 230ms outage to a branch office last week caused by a power failure combined with a badly plumbed power distribution.
Modern software people don’t consider 230ms to be an outage. Glad they don’t work in electricity.
(The failure we had was only on the services we guarentee at 99.1%, our lowest sla. After that there’s 99.95 and 99.999.
(In reality we reach five nines year after year on even the lowest levels, but there are major concerns like “large bomb in data centre” which could cause some of our less critical units to drop way more than 5 minutes a year.
With tech, like, it's crazily difficult to make sure every http request succeeds, so you build in retry, and look at that 230ms doesn't cause real disruption now. Not to excuse 98%, that's awful.
You work in grid/industrial?
There are projects like The Missing GitHub Status Page [2] that attempt a more accurate reporting, but taking 98.31% for granted, I think the more impactful framing would be _a floor_ of 1.5 business days lost every month. At 22 business days per month, that can feel more like 7+% downtime.
I saw a User's Group presentation by a guy at Speakeasy, they had made a whole logistic system that looked a lot like a CRM manager but not for customers, for dealing with how flaky Covad was at the time. Covad only had to be less shitty than Qwest (who we called, "Qworst", because if you hadn't dealt with Covad, they were the worst) and they couldn't even manage that.
So if Covad promised to have some line work done by 10 am on Tuesday, the SE techs would get a reminder to verify in had been completed shortly after the deadline so they could call Covad before the customer noticed the install was behind schedule and ask why it hadn't been done by the deadline. Otherwise SE was telling people they'd have Internet in 4 business days and it would be more like 10 - on average.
Around the time Qworst became CenturyLink, Covad bought SpeakEasy, for a net loss of beauty in the world.
1. These are core to the (service level agreements) SLA's for the enterprise companies when they evaluate and sign the contract (must for government RFPs)
2. The contracts generally have provisions for payback or penalties for missed SLAs
3. When your product depends on any of the products directly then the availability of your product has to take into account the availability of the dependent services. Eg, I can easily sell a SLA of two nines for my service if I am building it on a platform which has SLA of three nines, other way around is always questionable (though possible)
4. All of this is what used to happen may be up until 5 years ago. As of today even with good number of outages, company like AWS has not updated their availability to reflect those outages. Ideally all the companies built on top of AWS infrastructure would have to update their availabilities in a cascading manner but some how everyone decided to skip the beat
5. At this point, calculation of the availability itself has become so opaque that no one understands it anymore so no one questions it either
So a production outage happens, and we go to look at our runbooks to figure out what to do. The Wiki is also down, so no runbooks. What the actual fuck?
So it turns out the alleged clowns put all of our internal infrastructure onto the same SAN system in different partitions. They got a lot less judgy and the other Ops team got a lot more autonomy after that event.
Anyone with an Operational IQ above room temperature knows that you don't put offline and online resources onto the same hardware. Not only do they have separate duty cycles but also offline services can end up taking out the online services in unexpected ways, in part because they are assumed to be a bit sloppy and so they get less operational scrutiny.
I sped up the average runtime of a batch job by quite a humongous amount by throttling the request rate we made to a service that was used by user-facing services. If several unrelated batch jobs kicked off during the wrong time of day, we would start getting circuit breakers opening everywhere, including occasionally production.
I throttled us to use something like 8 or 10% of the official capacity of the service and no more. But I did it as limiting the number of in-flight requests, so that created back pressure if the servers were under heavy load and sending responses slowly, or let us go faster if the servers were returning results faster than usual. That peak shaving saved the other team a load of grief and let them push some capacity planning work down the roadmap to work on other things. And it dropped our failure rate by more than a factor of six. Low enough that we no longer had to babysit that process, which meant we started using it much, much more.
100%!
“It’s just a few minutes of downtime, nobody noticed, best not draw attention to ourselves by placing something on our status page”
Status pages should be tied into (external/simulated) monitoring to update in near real-time, without hiding incidents. Transparency is the only way. By reporting the _real_ hours of downtime, including the small stuff you think people didn’t notice.
Which 12 hours? One block of 12 hours? Twelve blocks of 1 hours? Important hours or unimportant ones? Which hours are important to whom?
The audience for the status page is everybody, and everybody will want things their way. There is no pleasing everybody. That's why there's color. If you see a lot of red, that's bad. If there's not that much red, it's not that bad.
Maybe that's not what you're referring to, but do i mess that simple uptime metric... Finding out your server has been trucking along for 5 years without a single reboot.