October 11, 2026 — Cloudflare tracing is in beta, and the meter is coming

Host0:30

Here's a trace to chew on. InfoQ reports one example where a cache miss spent five hundred twenty-seven milliseconds of a five hundred thirty-nine millisecond request waiting on the origin. That's almost the whole request. The other twelve milliseconds were everything else. Cloudflare Traces draws that picture for you. It entered open beta on October second, extending automatic tracing from Workers across the request path. Security rules, transformations, caching, routing, Worker execution and origin handling. You don't add instrumentation or plugins. Cloudflare records each supported platform operation as a span, with timing, outcome and attributes. The sampling controls are the part I like. You set a baseline rate, and Cloudflare's example is one percent. Then Trace Rules, which reuse Cloudflare's rules language, trace matching traffic at a higher rate. Match on paths, methods, headers, IP addresses or geography. So you can trace every request from one customer, or anything carrying a temporary debug header, without raising sampling on the whole domain. It plays well with others. Traces accept W3C traceparent context and can forward it to your origin. Spans export over OTLP to a compatible backend, so they sit beside the rest of your telemetry. Now the meter. Cloudflare says new pricing takes effect December first. Free gets half a gigabyte of ingestion a day and seven days of retention. Paid and Enterprise get fifty gigabytes ingested and ten gigabyte-months of storage per billing cycle. Past that, it's twenty-five cents per gigabyte ingested and ten cents per gigabyte-month stored. The shift, in Cloudflare's telling, is away from charging per span and toward charging by how much data you ingest and how long you keep it. InfoQ notes an earlier per-span plan had been reported for October first. Different schedules, not a contradiction. Here's the catch. It's a beta. Cloudflare itself says broader coverage, authenticated context propagation and ad hoc tracing are still planned. So if you'd accept trace context from callers you don't control, pause there. The breadth and OpenTelemetry claims are Cloudflare's own assertions, not independent evaluations. And how the new bill compares with per-span pricing for high-volume workloads isn't clear yet. To try it, enable tracing for a domain from the dashboard, the API or Terraform, and set an OTLP destination if you're exporting. Then check your ingest volume before December first. Traces show you what happened on a server. This next repo is about what happens on the glass in your hand.

3:32

The README's first line, translated from Chinese, says use Codex to operate your real iPhone over USB. That's the pitch, and the repo showed about three thousand stars and three hundred two forks when I read it. It's called iPhone-use, a Codex plugin with an MCP service, a setup skill and a live-screen widget. A local Python service talks to Appium WebDriverAgent on the phone over a USB-forwarded connection. It uses appium-ios-device for USB, port forwarding and screen streaming, and needs no separate Appium Server. You tell Codex what to do, and it calls seventeen MCP tools, plus two more for the widget. Tap, swipe, press buttons, run batch actions, scroll to find an element, collect lists. When it can't identify an element, it can return a screenshot so the agent recovers by coordinates. The draw is a real device instead of a simulator. The catch is setup. You need macOS, full Xcode, an Apple developer account for signing, a recent Node, and a plugged-in iPhone. The installer can build, sign, install and start WebDriverAgent. But you handle Apple sign-in, device trust, Developer Mode and unlocking. My favorite warning is about uncertainty. After a timeout or disconnect, an action may or may not have happened. The agent should check the actual phone rather than blindly repeat a tap or a submission. It also sends anonymous usage analytics to PostHog by default. The project says that excludes phone images, typed content and device identifiers, and setting do-not-track to one turns it off. Those privacy claims are the maintainers' own. To try it, clone the iPhone-use repo from GitHub and run its install script with sh. That repo is all about waiting to learn whether a tap landed. The next story is about a CPU that waits on memory, and a fix that stops it.

5:41

Four gibibytes. That's how large the sorted array got in the benchmark behind NumPy two point five's big number. The test machine's shared L two cache is twelve megabytes. The Scientific Python article reports that NumPy two point five made searchsorted up to twenty-five times faster than two point four. You hand it a sorted array and a batch of keys, and it hands back where each key belongs. The old version ran one binary search per key, and each memory access depended on the previous comparison. On a huge array, the CPU sits waiting for memory. The new native version advances several independent searches in batches, so the CPU has multiple memory accesses in flight. Think of a cook who starts several pots instead of staring at one. It also reformulates the search, tracking each query's base and one shared interval length. The output array doubles as search state, so extra memory beyond the output stays constant instead of growing with per-query arrays. Now, who measured it. The article's authors did, on random thirty-two-bit integer arrays, with ten thousand query keys, each run fifty times and the minimum reported. The machine was a MacBook Pro with an M one Pro and thirty-two gigabytes of RAM. That's not an independent reproduction. So twenty-five is a best case. It depends on array sizes, query count, hardware and cache behavior. Against the article's vectorized Python version, the C++ one can be up to twice as fast on smaller arrays, which is a different comparison. What you'll see on small arrays or already sorted keys is unclear. NumPy two point five point four landed October tenth. It's a patch release fixing issues found after two point five point three, and it supports Python three point twelve through three point fifteen. It adds a workaround for an MSVC bug and changes how svd and lstsq treat infinite inputs. The speedup isn't that patch's headline. It arrived in two point five. My verdict: take two point five point four, then time searchsorted on your own data shapes before you promise anyone twenty-five times.

8:08

Pydantic two point fourteen drops Python three point nine and adds three point fifteen, so check your interpreter first. It also adds ordered-dict and counter core schemas, and multiple of constraints for Fraction and timedelta. The changelog already lists two point fourteen point one, dated October eleventh, so check that patch before upgrading. The point zero release lasted three days. SQLite three point five four is about four percent faster in the project's own Valgrind cycle count on Linux x64, and seven and a half percent faster than three point fifty-one. SQLite says mileage varies. What breaks: Windows XP and CE support is gone, so Vista is the minimum, and Phoronix notes Unix builds are unaffected. The shell also no longer ends a command on a line holding only go or a slash, unless you build with the legacy flag. Goodbye, go. Copilot for JetBrains now needs IDE twenty twenty-five point two or later, per GitHub's October tenth changelog. Admins can set any available agent model as the default for new conversations, and a user's explicit pick still wins. A new setting stops MCP servers for Copilot and Claude starting automatically. On twenty twenty-five point one? Upgrade the IDE first. Claude Code two point one point two ninety six adds a code key to the Claude apps gateway's managed policies. It applies the same settings as the cli key to Claude Desktop's Code tab. Anthropic's docs say managed settings beat local config. The fixes include an escape sequence older MATE Terminal printed after copying. If you run the gateway, upgrade. Vite eight point three point four adds bundled-dev support for acceptExports, a hot module replacement call the docs aim at framework and tooling authors. The real content is the fixes: a restart request dropped during another restart, and static aliases matching outside path boundaries. Most apps can skip it unless they hit one. Transformers five point nineteen adds EmbeddingGemma two, Google's multimodal embedding model. It maps text, code, images, video and audio into one seven hundred sixty-eight dimension space, and Google describes the license as Apache two point zero. Google's model card reports code retrieval NDCG at ten of seventy-eight point six eight, versus sixty-eight point seven six for version one. That's Google's own number. Skip float16, which can quietly return degraded embeddings, and re-embed before swapping it into an existing index.

10:58

All that software assumes somebody built the machine. Oxide builds the machine, and says it can't build them fast enough. Oxide announced a four hundred forty-five million dollar Series D on October ninth, and SiliconANGLE reports Eclipse led it. The product is a pre-integrated rack of twenty-four sleds. Each sled can host an AMD CPU of up to one hundred forty-four cores, about one point six four terabytes of memory and up to three hundred terabytes of flash. SiliconANGLE quotes CEO Steve Tuck saying Oxide scaled manufacturing twenty times in twelve months and demand still exceeds supply. For software people, that's on-prem cloud arriving as a rack, with demand exceeding supply. Oxide also says it paid income tax this spring from ordinary computer sales. That's its own account. The next three items are about things that were right enough.

12:00

Fast Boot. Three items, and at least two were shipped by squinting. Iframe sizing. Chrome one fifty four can size an iframe to its content, but the embedded page must opt in and only Chromium has it. The author's form height was set by hand until it looked right, which is engineering by squinting. Lookalike domains. Researchers registered twenty lookalike domains using two odd characters that Chromium's lists missed, though no real attack has been reported. One Chromium defense is a guest list, and these two characters weren't on it. Driving physics. Pat Kerr rebuilt in JavaScript his nineteen ninety-six Atari ST prototype, which was later ported to C for use in Grand Theft Auto, and you can drive it in a browser. He calls the tyre model a semi-educated guess he knew wasn't accurate, which is more honest than most code comments. A form sized by eye, and a tyre model that was good enough. Squinting remains the most widely deployed technology in this episode.

13:07

Today: Cloudflare Traces and its new meter, an iPhone driven from Codex, NumPy's faster searchsorted, and Oxide's big round. Your try-it for today: run searchsorted on your own data under NumPy two point four and two point five point four, and see what number you get. Links to everything are in the show notes. Follow The Daily Reboot in your podcast app so the next episode finds you. I'm your host. Go build something.