New post: I added Direct I/O support to Cassandra for compaction reads, cutting p99 read latency by 5x (and p50 a little). Walks through the page cache mechanics, benchmarking, and when to enable it. Some funky charts along the way. lightfoot.dev/direct-i-o-for…

If you're interested in SSD internals and how to use them efficiently, our paper, “How to Write to SSDs,” has been accepted to VLDB and is currently on the Hacker News front page. vldb.org/pvldb/vol19/p1469-l…

An excellent take

As a manager, it took me far too long to learn that meetings aren't actual work. I thought meetings drove output. Turns out, most meetings are about control, not collaboration. About signaling ambition versus pursuing it. If you can't pass these three tests, cancel the meeting:

Someone really built this. A VS Code extension that turns your AI agents into pixel art characters working inside a virtual office.

New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel. Here's what it taught us about the future of autonomous software development. Read more: anthropic.com/engineering/bu…

Linus vibing. Things I did not expect.

Interested in io_uring? Check out our new paper, joint work with my former colleagues from TU Darmstadt (@cbinnig and @m4ttze) and from TU Munich (@melhindiCS and @viktor_leis). arxiv.org/abs/2512.04859v1

The Path of a Packet Through the Linux Kernel net.in.tum.de/fileadmin/TUM/…

Tales of the Tail is one of my all time favorites research.google/pubs/tales-o…

An excerpt from the OSTEP book regarding distributed file systems. Sounds familiar - is it just scalability the protocol affects @jorandirkgreef? :-) youtube.com/watch?v=yKgfk8lT…

TIL: The Chrony NTP daemon fixes clock drift fast — slew rate of 100,000 ppm vs ntpd’s 500 ppm. Sparked by the 'Order and Time' chapter from @DominikTornow's Distributed Systems book.

It still has to be reviewed, but finally the cursor based C* compaction patch is in the open: issues.apache.org/jira/brows… "3-5x faster [compaction] in most scenarios and allocates ~20mb vs. multiple GB[per compaction]" :-)

Latest post: Fintech Devcon 2025 lightfoot.dev/fintech-devcon…

How to develop taste: Step 1: Ignore all the hustle-culture get-rich-quick BS. If someone isn't motivated by quality they won't develop it. Step 2: The essence of taste is simplicity. Simplicity is hard. Any time a system is getting messy or complicated ask yourself "how could we have designed this differently to make it more simple?" Step 3: Never dismiss a decision you don't agree with as someone being an idiot. They had a reason for making that decision and by learning what it was you'll learn perspective plus more about how humans think and use systems. Step 4: Good taste doesn't show up on short timescales. You won't know if a decision was good until years of usage and changing requirements. If you hop between jobs too fast or don't own projects long enough you won't learn this. Step 5: Never be satisfied with any work you do. This isn't a recipe for happiness but it's a recipe for introspecting every day on what could have been better. Good taste evolves from a constant desire to build things that solve real problems in the most elegant way possible. Step 6: You can't derive taste from data-driven decision making. You need to make decisions based on intuition. Develop the confidence to make your own decisions, stand behind them, and learn when they could have been better. Step 7: You also won't learn good taste from an LLM. LLMs traffic in competency. Taste is about excellence. This one is on you. Step 8: There is no step 8, you've just got to try really hard every day, ignore most of what you read on the internet, and get slowly slowly better over time.

Or we can do better: To hedge physical requests for consistency and tail tolerance, and with logical availability rather than CAP’s reduced definition of physical availability. Plus, tighter P100s on latency than what CAP would give. TB recently started hedging for all these reasons.