That's calculated with the assumption that you can cluster the events within a 1-month period and that the events happen at a fixed rate.

We are talking about a period of a few days here and about only a handful of services that tout 99.999(9)% uptime. I'm no mathematician but I don't think it's a great comparison.

AWS averages about one major outage per year[0].

Google's own status page[1] list near 100 incidents in the past 365 days, only one of which ended on HN frontpage.

[0] https://en.wikipedia.org/wiki/Timeline_of_Amazon_Web_Service...

[1] https://status.cloud.google.com/summary

Both of these Google outages made it to the HN frontpage:

https://news.ycombinator.com/item?id=20077421

https://news.ycombinator.com/item?id=20338263

Both of these AWS outages made it to the HN frontpage (and neither are listed in that AWS timeline):

https://news.ycombinator.com/item?id=20298719

https://news.ycombinator.com/item?id=19749781

Good correction: the outages do happen at a fixed rate. Regarding GCP, I only see a couple events that have lasted more than 8 hours, not sure how many of those affect more than one subsystem.

No the cluster size isn't fixed, nor is the event rate — the clusters are chosen purely based upon what an "unusually long" time between events would be. That's precisely how the rate ends up canceling itself out. Given this formulation, the expected cluster size is always e, regardless of how you define celebrity. That's what makes it fun.

Now the human psychology part that this doesn't cover is that typically when you have two or three A-listers die, well, then you start seeing all the B- C- and D-listers that also died in the same period that you would have otherwise ignored.

Can you calculate the chance of the cluster size being >5 when we can cluster events within a one-week period and the rate of events is one every 2 months (debatable, but feels right to me)?

I think that would just end this discussion. I don't know how to calculate that, but my intuition says the resulting chance is low.

I don't know how to do the math, but my python simulation says zero.

  import random

  service_providers = 20
  servers_threshold = 5
  up_time_threshold = 0.999**7 # prob down in week
  years = 20
  occurrences = 0

  # run sim 10,000 times
  for i in range(0, 10000):
      servers_down = False
    
      # simulate weeks in 20 years
      for j in range(1, 52*years):
          how_many_down_this_week = 0
        
          # run up times for service providers
          for s in range(0, service_providers):
              up_time = random.random()
              if up_time > up_time_threshold:
                  how_many_down_this_week += 1

          # did they go down?
          if how_many_down_this_week > servers_threshold:
              servers_down = True

      if servers_down == True:
          occurrences += 1

  print(occurrences)

Given those parameters, you'd expect to see something every 100,000 years or so. (poisson distribution, lambda=1/8, k=5, the result is like 7e-7).

Plus one big outage story then makes other minor outages get more headlines... when normally no one cares.

Same thing with hurricanes, when a bad one happens in the US they are more likely to document other minor hurricanes in the Caribbean and Latin America. But otherwise plenty of hurricanes hit those areas and it doesn't make major headlines.

Facebook, Google, Apple, Cloudflare, Azure, Amazon, WhatsApp, and we have noticed smaller bits of routing weirdness like servers in Los Angeles not being able to hit GitHub for a few hours a few days ago.

This is more like 6-8, not e. It's definitely odd.

One non-conspiracy explanation I can imagine is that maybe all these big providers have a bunch of hidden dependencies on each other.

Apple documents their dependencies on other Cloud providers. Facebook has its own DCs but I think they’re sometimes colocated with cloud providers’.

But the outages are sequential, not simultaneous

They are very dense compared to the usual rate.

I agree, this is a bigger and really dense cluster, but remember that e is just the average (mean) cluster size. There will be clusters bigger than it.

My point is simply that our intuition of randomness biases us to see meaning in clusters when there is none in the first place. In other words, we need to intentionally shift our prior on how unlikely this event is. Yes, it's still unlikely, but it's not as unlikely as you would think.

“When it rains, it pours” is the conventional wisdom distillation of Poisson variable behavior

2.718 is the Euler number

https://en.wikipedia.org/wiki/E_(mathematical_constant)