Service reliability math that every engineer should know
Uptime Downtime (Yearly)
99.00000% 3d 15h 39m
99.90000% 8h 45m 56s
99.99000% 52m 35s
99.99900% 5m 15s
99.99990% 31s
99.99999% 3s
Matt Rickard
5,467 posts
Matt Rickard
@mattrickard
tweeting about ai and software
- Obama's career advice for young people -- Just learn how to get stuff done. What I mean by that is I've seen at every level people who are very good at describing problems, people who are very sophisticated at explaining why something went wrong or why something can't get
- Wow, so Berkshire Hathaway is so expensive it broke NASDAQ data feeds. Apparently the feeds transmit stock prices as int32. The highest number they can handle is $429,496.7295. Berkshire class A hit $421,000 and disappeared from the feed. Fix is slated for May 17th
- GitHub Copilot, but for merge conflicts
- What was previously Google's biggest strength might prevent it from innovating in the future. Google technology used to be years ahead of the industry. In 2004, Google released a paper on its proprietary MapReduce algorithm, the system that powered Google's massively parallel
- Distributed Systems & AI Jeff Dean, Google’s SVP of Google Research and AI, started his career working on compilers and profiling tools. He would go on to work on hard distributed systems that formed the basis for Google’s infrastructure, like Spanner (distributed SQL), Bigtable
- Run LLMs entirely in the browser with a simple headless React hook, useLLM(). Live demo: chat.matt-rickard.com GitHub: github.com/r2d4/react-llm react-llm/headless lets you customize everything from the system prompt to the user/assistant role names. It manages a WebGPU-powered
00:00 - Great CI/CD = productive developers. What does the ideal CI/CD stack look like today? Serverless – A CI/CD system should be ephemeral and serverless. There's no reason to be managing Jenkins or individual nodes. Managing CI/CD machines makes it really easy to introduce
- You asked and we delivered. XML is the new default for Kubernetes and kubectl. All previous releases have been cherry-picked.
- I worked on machine learning infrastructure in the last AI cycle (compute + deep learning). Some lessons learned and other observations: • Human-in-the-loop doesn't work. The idea of "faking it until you make it" by being a negative margin services company with the hopes of
