New SIGOPS Blog: The journey of real-life industry work behind an #OSDI paper — global capacity management for millions of servers by Marius Eriksen (marius eriksen) from Meta. sigops.org/2024/the-journ…

Ceph storage evolution paper that worth a carefully reading pdl.cmu.edu/PDL-FTP/Storag…

My optimizer/execution etc. knowledge is sub par. I’ve decided to do something about it. pingcap.com/blog/10x-perfo… Seems like people don’t know much about TiDB’s implementation, it’s a very sophisticated parallel implementation that is not only parallel by design but also

Border Gateway Protocol (BGP): How the Internet’s Routing Protocol Works - j2sw Blog blog.j2sw.com/networking/bgp…

Much of TigerBeetle is taking physical, non-deterministic interfaces and making them logical, deterministic interfaces. For example, replacing physical IPv4 addresses with logical replica identifiers in MessageBus, or replacing fds with grid addresses. Here's another example.

engineering team when you tell them we are now an "AI first" company

My talk on Zero's Query-Driven-Sync and IVM is live: youtube.com/watch?v=39CizI… 30 minutes was a bit tight to do this well (hence terrible ending 😢) so I'd like to do this online with like an hour and time for questions and go deeper. Let me know if you'd be interested in that

Upgrades are neat. TigerBeetle ships as a single, static binary. Each binary includes the code for several previous versions as well. You replace the binary on disk, and it continues to run the current version. Then, it coordinates with the rest of the cluster to roll out each

Failover is the foundation of high availability. At PlanetScale we perform over 100k each month, 99%+ of them on purpose. It's used for version upgrades, node TTLs, and unexpected failures. So what is a failover?

Done with this jam! Here's a detailed writeup of what I learned: handmade.network/p/688/buongior… There is so much potential for small-network tools like this...I need to keep exploring ideas.

Hands On Threading Practical Insights For Developers And System Designers

Caching is an Abstraction, not an Optimization buttondown.com/jaffray/archiv…

There are different reasons/benefits with every OLTP DB architecture. I would classify them into largely three types and I have run and managed all three versions in my lifetime. Pick your option based on what your company needs 1. Full local storage. This is kind of how we ran

Mutexes, Condvars, and Rwlocks are a scam. They introduce latency bottlenecks/dependencies from the OS scheduler. Instead, design data flow first, then optimize its channels second:

TigerStyle existed before TigerBeetle was created, and was intended for software development in general, as a methodology for “higher quality software in less time”. Enjoyed this post from Lewis Campbell, which gets it. lewiscampbell.tech/blog/250718.ht…

My talk from Better Software Conference last week is up on youtube. I present my findings on thread synchronization and job systems that I learned while parallelizing the physics solver. youtube.com/watch?v=Kvsvd6…

Here is why I like AMD’s Zen series of CPUs (right) over Intel’s CPUs (left). The NUMA microarchitecture of Intel results in each core being a different physical distance from the memory controller, resulting in very different memory latency performance per core, making it

what are large language models actually doing? i read the 2025 textbook "Foundations of Large Language Models" by tong xiao and jingbo zhu and for the first time, i truly understood how they work. here’s everything you need to know about llms in 3 minutes↓

No gonna lie, I'm the biggest Apple hater for very very good reasons, but their last research paper explaining why they weren't interested in pursuing LLMs seems more and more like its gonna be correct.