I build and run things on the web, mostly infrastructure and the unglamorous parts that hold it up.
This is where I keep the links worth remembering, with a sentence or two on why.
The clearest framing I have read of why agent security is its own problem. Private data, untrusted content and the ability to communicate externally are each fine alone; the combination is what gets you. Worth keeping as a checklist rather than a blog post.
The interesting part is not the storage engine, it is that a point read stopped being an await against a remote KV. A lot of designs that were awkward because every read cost a round trip are suddenly reasonable.
Marc Brooker on why the p99 you measure at the load balancer is not the p99 your customer experiences. The maths is short and the conclusion is uncomfortable.
Not about software performance. About the habit of treating every part of a life as a system with a throughput number attached.
I keep coming back to this one when someone proposes a fourth datastore. Usually the honest answer is a table and an index.
Spent the afternoon reading our own runbooks as though I had never seen the system. Three of them start with a step that assumes you already know which cluster is affected. That is not a runbook, it is a reminder.
Old, still right. If a page does not require a human to act immediately, it is not a page — it is a dashboard that learned to interrupt you.
Good explanation of the interaction between Vary, cache keys and origin headers. The section on why Vary: User-Agent effectively disables caching is the bit I have had to explain most often.
Two lines of CSS and full-page navigations get a transition. No router to intercept, no JavaScript to ship. This is the version I want; the single-page one always cost more than it returned.
Reread after a build produced different output on two machines in the same office. The answer was a date formatted in local time in exactly one place.
Sliding window counters are the compromise nobody is happy with and everybody ships. This is the best explanation of why the approximation is acceptable.
The advice to write the timeline before writing anything else is doing most of the work here. Every review I have seen go wrong started with someone drafting the conclusion.
And then the follow-up, which is that for most workloads it is fine as one anyway. Both things are true and the article is honest about where the line sits.
Enclosure is the right word. Nothing is being destroyed; it is being fenced and rented back.
The initial congestion window has been ten segments since 2013 and most of the internet has moved on around it. Good data here on what that costs a small response.
The component library is the artefact. The agreement about who can change what is the system.
Sara Soueidan on 2.5.8. The 24 by 24 minimum sounds trivial until you audit a real navigation bar and find half of it fails on the spacing exception rather than the size.
Migrated a personal project off a build step that existed only to inline three kilobytes of CSS. Twelve dependencies removed. The page got four milliseconds slower and I do not care.
Not a paper — an interactive bit-level view of a float. I have linked the Goldberg paper at people for years and this explains more in thirty seconds.
Expand, backfill, contract. There is no clever version of this, only a disciplined one, and pgroll makes the discipline mechanical.