October 12, 2026 — EventBridge gets replay, but not for yesterday

Host0:46

AWS lets you set retention on its new EventBridge bus anywhere from one to three hundred sixty five days. Here's the catch. Changing that setting only applies to future events. Older ones don't come back. That's one detail inside a bigger launch. AWS announced enhanced custom event buses for EventBridge, with cross-account sharing, ordered delivery, replay, content-based deduplication, event transformation, and Avro and Protobuf support. AWS calls it a relaunch of the original custom event bus resource, on a new architecture. So the replacement for the custom event bus is the custom event bus, enhanced. InfoQ quoted Gregor Hohpe calling it almost like Kafka but a big step ahead, naming aside, which he said is in good tradition. Here's the interface. In the platform account, you create the bus with the AWS CLI. That's events v two, create event bus, with a storage configuration that sets retention to seven days. AWS says seven days covers a working week. You get back a bus ARN, and you share that with publisher and subscriber accounts. The design splits the labor. A platform team governs the bus. Application teams create subscribers in their own accounts. InfoQ says a new Subscriber resource combines filtering, targets, retries, and dead-letter handling. Publishers send structured JSON with PutEvents. For binary Avro or Protobuf, they use PutRawEvents and name a schema registry in the request. Ordering holds within a key the publisher supplies, not across the whole bus. It replaces the old pattern, forwarding events through rules from bus to bus across accounts. AWS says that adds ingestion hops and spreads governance and observability across separate bus configurations. Micah Walter, an AWS senior solutions architect, said in InfoQ, “This workaround reintroduces the operational complexity that serverless architectures are meant to eliminate.” The number that matters is price. InfoQ reports that pricing separates publisher ingress from subscriber egress. Cloud economist Corey Quinn put ingest at eighteen cents per gigabyte, high but not unreasonable for typical EventBridge use. AWS also claims scale and cost improvements. That's a vendor claim, and nothing I read includes an independent benchmark. Expect pushback. Teams with existing routing will ask what migration costs. Kafka users will point out that a commentator's one-liner isn't a benchmark. Nobody's forced, though. Existing buses stay unchanged, because this is a new resource. What's unclear is how ordering and replay behave under failure, so read those docs before you commit a topology to it. To try it, run that seven-day create command in a sandbox platform account, then share the ARN with a second account.

4:00

The last story's numbers came from the vendor. This one comes with a confession. The WordCraft README says, flatly, that it's in alpha, and the project's own roadmap puts it about sixty percent of the way to replacing Word for real work. WordCraft is an open-source, clean-room reimplementation of Microsoft Word, written in Rust. It runs as a native desktop app and in the browser through WebAssembly. It reads and writes DOCX, and handles ODT, RTF, HTML, Markdown, LaTeX, and plain text. It imports legacy DOC files and exports PDF. Here's the part for builders. A CLI and an MCP server expose four hundred twenty two commands for editing, inspecting, rendering, and saving documents. You send a command, it changes or renders the document. Under the hood, it's layered crates. Document and editing, layout and file formats, rendering and PDF, an engine, then replaceable front ends, including an egui desktop app and the WebAssembly one. Paragraph layout is cached and pages render on demand. The README claims about one point four milliseconds to re-layout typing in a one hundred eighty eight page document. That's the project's number, not an independent test. The limits are the project's own list. Real-world DOCX testing, charts and SmartArt round-tripping, pagination fidelity, stability, and native printing. It's licensed MIT or Apache two point oh, and says it's built from public specifications and observation. The team even takes its screenshots with the app's own offscreen test harness, so the app takes its own headshots. To try it, clone the repo and run cargo run with the release flag, package wordcraft, and the sample flag. That launches the desktop app with a sample document. Just don't open your contracts in it yet.

6:06

WordCraft's browser version runs on WebAssembly, inside the same kind of browser that, according to our next story, is quietly telling the web about your GPU. Open Roy Fish's demo and it hands you a fingerprint. Refresh, same one. Clear cookies, same. Turn on a VPN, same. Go incognito, same. It changes across devices, but not across any of those. Fish argues in his post that fingerprints have become what cookies used to be. Here's how it works. The demo measures audio-rendering output, WebGL renderer and drawing behavior, canvas text rendering, and locale and time-zone data. It hashes selected measurements together. Beyond the browser, Fish lists network and environment signals, including IP address, browser and OS headers, language, GPU, CPU, display, fonts, and system information. The number is seven. His example is a trait found in one of one hundred twenty eight candidates, which yields seven bits of surprisal. Then comes the catch he stresses. Correlated traits don't simply add up. And a fingerprint needn't be globally unique. It only has to separate you within a particular sample. Firefox's own feature page backs up the mechanism. Fingerprints can survive cleared cookies and private browsing. Firefox says it blocks known fingerprinters, reduces the usefulness of some signals, and in some settings randomizes canvas readback data. Fish names two defenses. Farbling, which adds noise to measurements, and making browsers look alike. Fish writes, “The ironic truth is, privacy extensions make you more unique, since fewer people have them.” Treat this as an illustration, not a study. It's one author's demo, and Firefox is a vendor describing its own product. Fish's claim that only Brave, Tor, and Mullvad Browser offer advanced protection hasn't been independently evaluated. My favorite detail. The script can't run with JavaScript disabled, but Fish cautions that disabling it does not block network-based identification. You can switch off the script, but network-based identification is another matter. To try it, run EFF's Cover Your Tracks, which reports entropy for measured data points. If you've ever told a user that clearing cookies resets them, run it first. What nobody has measured here is how distinctive these signals are in real populations, or how well defenses hold once trackers adapt to a blocked measurement.

8:51

Unlike fingerprints, releases come with changelogs. Mostly. Codex CLI zero point one sixty-two point one. Startup failures from a mismatch between a running background server's feature settings and your CLI defaults are fixed. Compatibility checks now apply only when you pass an explicit feature override. It also fixes a terminal interface crash when an asynchronous question spans several lines, and keeps line breaks and complete hyperlink destinations. Upgrade if you've hit either one. A link keeping its whole destination is a low bar, but nice to see it cleared. Pydantic two point fourteen point one. It avoids two temporary allocations during validation, a Python tuple in model-fields validation and an extra dictionary in model validation. The notes publish no measured speedup, so don't promise your team a faster API. It also fixes Decimal max digits enforcement, out-of-range timedelta integers, inherited TypedDict support for closed and extra items, and a ForwardRef TypeError on Python three point ten point zero specifically. Not three point ten in general. Somebody's very exact bug. Upgrade with pip install, dash U, pydantic. CrewAI one point fifteen point twenty seven. Evaluation runs now track cost and time per run for crewai eval with the models flag, and record why an evaluation stopped before its verdict, without repeating progress lines. It also adds DeepInfra as an OpenAI-compatible provider. Other fixes cover stop sequences sent to GPT six, GPT five point six, and gpt-oss, and SQLite flow initialization contention. Upgrade if you compare models, but check the cost numbers against your provider invoices. OpenAI-compatible doesn't mean identical. Bun one point four point three. I can't tell you what changed. The release page thanks eight contributors, lists install commands, and sends you to Bun's blog for the notes. The page I read doesn't show the fix list. Until you've read those notes, there's no sourced reason to call it urgent. Install options are a curl script, a global npm install of bun, or PowerShell.

11:15

CrewAI will now tell you what a run costs. Here's a laptop with opinions about what your RAM should cost. MSI's Prestige N16 Flip AI+ is a sixteen-inch convertible running Nvidia's RTX Spark N1X, a twenty-core processor paired with a GPU of six thousand one hundred forty four cores. The screen is a three thousand eight hundred forty by two thousand four hundred Tandem OLED touchscreen, thirty to one hundred twenty hertz, one thousand nits at peak HDR. All of that is per TechRadar. It comes in three configurations. Thirty two gigabytes with one terabyte, sixty four with one terabyte, and one hundred twenty eight gigabytes with two terabytes. U.S. pricing starts at three thousand two hundred ninety nine dollars. The top Eurozone model is seven thousand nine hundred ninety nine euros and ninety cents, more than twice the entry price of three thousand eight hundred ninety nine euros and ninety cents. Why care? One hundred twenty eight gigabytes is a lot of headroom for a heavy dev environment, though that's my reading, not MSI's. It weighs one point nine kilograms with a ninety nine point nine watt-hour battery. TechRadar says MSI had not specified battery endurance, and the report includes no measured performance comparison, so this is a spec sheet, not a verdict.

12:46

Now for the part of the show that costs nothing. Fast Boot, three items, no migration plan. Karaoke malware. Malware on one contractor's computer at Nippon Columbia Group may have exposed eight million six hundred thirty one thousand karaoke customer records, with no theft confirmed. The malware's family is unknown, so it's the only act on the bill without a songbook. Dwarf Fortress. Its creator skipped version control in twenty thirteen because it felt like a black box with no immediate upside, then mentioned using it by twenty twenty-three. Simon Willison says the Steam release made the creators millionaires—an upside, at least, for the game. D-Bus Hell. David Bushell registered dbushell dot com in two thousand four, two years before D-Bus shipped its first release, and readers keep reading it as D-Bus Hell. He kept the domain anyway, because the real hell is redirects. Tarn Adams once wouldn't commit to version control, and David Bushell won't commit to a new domain. Some people just have commitment issues.

13:56

Today: EventBridge got a shared bus with ordered delivery and replay, WordCraft is a Rust Word clone that admits it's in alpha, and browser fingerprints are doing the job cookies used to do. Your try-it for today. Create that EventBridge bus in a sandbox account, or run Cover Your Tracks on your own browser and see what it says. Every link is in the show notes. Follow The Daily Reboot wherever you get your podcasts, so tomorrow's episode shows up on its own. I'm your host. Go commit something.