My notes on information technology, technology, science and other matters, as taken from various podcasts, videos and other online flotsam. In effect, this page is a log of stuff I've learned over these past 30 days, often summarized on Twitter. Some of it also ends up in my reporting for my current job at The New Stack.
This Month: Bryan Cantrill on starting a hyperscale hardware company for the enterprise; Timnit Gebru on the problems of Google’s indiscriminate data harvesting; machine learning vs. reinforcement learning; what are feature stores; Netflix moves to a GraphQL Federation gateway; The unintended consequences of monitoring.
Data Centers"There was a need for on-premises compute that had not gone away [But] all these innovations in hyperscale computing were locked up in a small number of companies: Google, Facebook, Microsoft"--Bryan Cantrill on starting Oxide Computer, speaking on the Data Center Knowledge podcast. Cantrill goes on to say that the cloud has been around for awhile now. Most workloads that could have gone to the cloud are probably already there. There are good reasons for keeping them on-prem: economics being a good one. The cloud is expensive. Also, bandwidth, a high margin service for public clouds, ends up being really expensive for public-facing businesses, especially those that have gone beyond the grow-at-any-cost phase and are now looking to control costs for the long. CapEx is the new OpEx. Now that Moore's Law is dead, it becomes feasible to run the same server for maybe a decade or even longer. We used to replace servers every 18 months with something 2x as fast. “That is definitely not happening now,” Cantrill said.
Oxide is building its own next generation open source OS from Rust, called Hubris. Currently it uses the OpenSolaris Illuminos BSD kernel with BeeHive as the hypervisor and CockroachDB for the control plane.
Machine LearningIn the paper that allegedly got her fired from Google, AI researcher Timnit Gebru argues that large-scale learning models built on Google’s indiscriminate data harvesting have led to racial biases in the company's products. “Large datasets based on texts from the Internet over-represent hegemonic viewpoints and encode biases potentially damaging to marginalized populations,” Gebru said.
Machine Learning vs. Reinforcement Learning: Machine learning optimizes for a single decision, with all the data available at that moment, whereas Reinforcement Learning takes a longer view, based on previous attempts and reward signals.
One place where RL works better than ML is in Robotics. A ML robot in a maze gets stuck at a dead-end path that stops very near the end goal. Going backwards would be "suboptimal" in ML. "You get trapped in these dead ends that are almost optimal but not quite optimal" -- Dr Phil Winder, InfoQ Podcast
"Unity ArtEngine is a tool that uses AI to remove time-consuming and repetitive material creation tasks. The program is popular for scan-cleanup on photogrammetry scans, generating and manipulating [Physically Based Rendering] materials."--80.lv
Only 5% of ML processing software is actually specific to machine learning work. "Machine learning on Kubernetes is very unfriendly to data scientists"--Ian Hellström, on D2IQ's new Kubernetes MLops platform Kaptain. (Video) (Slides)
She is beautiful. He is clever...(Phoebe Tickell)
Towards a more equitable AI (Rachel Thomas)
& this is exactly why we’ll never have to worry about the singularity (Hal9000)
“It’s almost like we’re all starship captains now. We have become managers of ecosystems of complexity" -- André H., on whether or not the overabundance of abstractions in the IT Field is harmful, Usenix SRECon. “If your abstraction is compensating for a problem in another tool, should you write an abstraction for that, or should you solve that in the other tool?” (Video) (Slide).
"profanity is my number one debugging tool."--Yossi Kreinin.
“Oracle is set to mark Applet APIs for removal in JEP-398. Java Applets originally powered Rich Internet Applications at a time when browsers had less power and fewer standards for developing applications” --Erik Costlow,InfoQ.
"As good as Tailwind is, one downfall can be the complexity of the resulting CSS for build tools and browsers. No fear.. Tailwind is getting a just in time compiler to compile your CSS on demand as you author things." (Front End Focus)
Scraping a Web page in Python (FreeCodeCamp)
"Infrastructure is best modeled not as code, nor in a GUI, but as a text-based, middle-ground, data-driven policy" -- (Michael DeHaan)
Docker won the container wars, Kubernetes won the orchestration wars, and it is looking like Envoy Proxy is winning the proxy service space and Istio is becoming the dominant Service Mesh -- Christian Posta, Solo.io, Solocon21.
How WebAssembly could work with a Service Mesh: A WASM binary, such as a traffic filter, can be packaged into an OCI image & shipped to a hub (much like Docker), so that any Envoy can find the extension/ download into its own mesh. -- Yuval Kohavi, Shane O'Donnell Solocon21.
IstioMesh 2021 development will focus less on building new features and more on easing Day 2 operations: lower overhead, easier upgrades (Thx! Helm), self-testing docs, better debugging. Feature dev will move to WASM extensions -- Louis Ryan, Lin Sun, SoloCon21.
Netflix is moving user services to a single GraphQL Federation gateway, which breaks queries into multiple API requests that can be sent to 100s of separately-managed domain services (>40k nodes) (InfoQ)
While our primary mission is freedom for software users, we want to be clear, the FSF board unanimously condemns misogyny, racism,and other bigotry as well as defamation, intimidation, and unfair attacks on free thought and speech. pic.twitter.com/B9lq9Hq8z0— Battle Programmer Yuu (@netspooky) March 30, 2021
"These holes entered Linux when virtual socket multi-transport support was added. This networking transport facilitates communication between virtual machines (VM) and their host." Steven J. Vaughan-Nichols on the latest Linux vulnerabilities. (ZD Net)
Open source log parser LogReduce uses numerical variable compression to trim log size, sorely needed by Alibaba to parse 1PB daily cloud logs. It cut 1.7TB to 34GB — 4x the compression ratio of LogZip. Also by switching from #Python libraries to C++, the researchers were able to cut the compression time by 2x. --Junyu Wei, Usenix Fast21. (Video) (Slides) (Slides)
“Whenever you try to force the real world to do something that can be counted, unintended consequences abound.” The New Yorker on statistics.
"But it’s exhausting work. And repetitive, to continue to offer the same basics to white people of what’s wrong with a country, an economy, and a tech industry that’s systemically built on anti-Blackness."
"...mediocre, unremarkable white dudes in tech creating products and services that are harmed fundamentally because they don't have the lived experiences of seeing those blind spots."--Kim Crayton, (The New Stack)
Some 5G coverage (Verizon, AT&T) is actually *slower* than the 4G/LTE networks, due to less-than-ideal spectrum allocation. Big performance winner here? TMobile. (Verge)
"A feature store is a data management layer that saves features specifically designed for machine learning use cases. What is a feature? A feature is a measurable property of an entity, which is a model or representation of a domain object."-- Databaseline
Dutch tax authorities created "risk profiles of residents who were supposedly more likely to commit fraud and then used automated systems with little oversight to scan through benefits applicants and flag likely fraudsters who were then forced to pay." (Vice)
"Where we look as we move broadcasts details about where we intend to go next. Without that, it’s harder for passers-by to avoid us gracefully."--The New York Times on the dangers of looking at your phone while walking"This teaches us that—when it’s a big enough deal—Amazon will lie to us. And coming from the company that runs the production infrastructure for our companies ... this is a nightmare." -- Corey Quinn,on the recent AWS Tweetstorm targeting members of the U.S. Congress.
Amtrak is proposing New lines to Nashville and Scranton, which would be nice, but no still no direct "high-speed" NYC <--> Chicago route? Seems like that one could actually be a feasible alternative to the headaches of flight.
100 years ago trains could run NYC to Chicago in 16 hours, and that time could have been cut to 10. (Darienite)
"You can't write a simplified explanation if you are not qualified to write the complex explanation"-- Robert Graham.