<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Mark Elliot on Medium]]></title>
        <description><![CDATA[Stories by Mark Elliot on Medium]]></description>
        <link>https://medium.com/@markelliot?source=rss-ee08a1ec175c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*dxAItbmv0sFhUdTnr9BMPQ.jpeg</url>
            <title>Stories by Mark Elliot on Medium</title>
            <link>https://medium.com/@markelliot?source=rss-ee08a1ec175c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 08:43:57 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@markelliot/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[2020/August]]></title>
            <link>https://medium.com/@markelliot/2020-august-f1a953c14274?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1a953c14274</guid>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Tue, 01 Sep 2020 14:25:04 GMT</pubDate>
            <atom:updated>2020-09-01T14:25:04.626Z</atom:updated>
            <content:encoded><![CDATA[<p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a> and recommendations from others. Here’s August’s list.</p><p><a href="https://medium.com/@steve.yegge/dear-google-cloud-your-deprecation-policy-is-killing-you-ee7525dc05dc">Dear Google Cloud: Your Deprecation Policy is Killing You</a><br>Deprecation timelines can be good for your organization but bad for your customer — there’s things that can be done to balance the work (providing automation to migrate), but in some cases it might be better to find a way to live with the current usage. One tactic I’ve found useful is thinking about leaving behind an adapter that converts old use to new use but otherwise moves the support for old functionality out of the primary development target.</p><p><a href="https://www.latimes.com/opinion/story/2020-08-16/covid-war-mash-doctor-alda-hawkeye">Op-Ed: As a doctor on ‘MASH,’ Hawkeye showed me how to get through this war on the coronavirus</a><br>MASH re-runs ran on TV between when I got home from swim practice and dinner (so pre-homework) when I was a kid, so the title caught my eye, and this op-ed doesn’t disappoint. The article is worth a read, and the show is certainly worth some time.<br><em>Humor is just shorthand for hope, a gamble we’ll live long enough to hear the punchline, whether in the face of hardship or the pathologic indifference to suffering that sows the seeds for war in the first place. And in a war of ideology, humor correlates to the most important thing of all: resistance.</em></p><p><a href="https://www.nejm.org/doi/full/10.1056/NEJMra1905136">Effects of Intermittent Fasting on Health, Aging, and Disease</a><br>Some of the scientific theory behind how and why IF works, plus evidence that it is in fact beneficial.</p><p><a href="https://ptx2.net/posts/unbricking-a-bike-with-a-raspberry-pi/">Unbricking a $2,000 Bike With a $10 Raspberry Pi</a><br>The steps to reverse engineer the data feed from an at-home Flywheel bike, and the steps to feed that data into Zwift using a Raspberry Pi. Some interesting insights on how to listen to the Bluetooth networking stack for packet-capture, and then a bit of experimentation to figure out what’s in those packets.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1a953c14274" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020/July — Reading List]]></title>
            <link>https://medium.com/@markelliot/2020-july-reading-list-34cfc4837dee?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/34cfc4837dee</guid>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Tue, 01 Sep 2020 14:11:19 GMT</pubDate>
            <atom:updated>2020-09-01T14:11:19.482Z</atom:updated>
            <content:encoded><![CDATA[<h3>2020/July — Reading List</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a> and recommendations from others. Here’s July’s list.</p><p><a href="https://svpg.com/product-vs-feature-teams/">Product vs. Feature Teams</a><br><em>If I had to pick the one thing I love the most about the feeling of working in a truly empowered product team, it is the form of magic that happens when you have people that are a) motivated and b) skilled in their respective discipline — product, design and engineering — and they sit around a prototype or watch a user interact with our prototype, and the engineer points out new possibilities, and the designer points out different potential experiences, and the product manager weighs in with the sales or financial or privacy related implications, and after exploring a bunch of approaches, they find one that actually works — it’s valuable, it’s usable, it’s feasible and it’s viable.</em></p><p><a href="https://svelte.dev/">https://svelte.dev</a> + <a href="https://svelte.dev/blog/svelte-and-typescript">https://svelte.dev/blog/svelte-and-typescript</a><br>In line with the discussion in <a href="https://tomdale.net/2017/09/compilers-are-the-new-frameworks/">Compilers are the new Frameworks</a>, Svelte provides an Angular2-like data binding model but with better-than-React performance and smaller overall code-size by generating the binding as imperative code at compile-time, and by skipping the virtual DOM. Over the years our use of Java has evolved in this direction and the benefits have been big, I’d expect to see more of Svelte or frameworks like it. Also…now with Typescript support.</p><p><a href="https://idk.dev/improving-performance-of-php-for-arm64-and-impact-on-amazon-ec2-m6g-instances/">Improving performance of PHP for Arm64 and impact on Amazon EC2 M6g instances</a><br>AWS-developer contributed fixes and improvements to PHP performance as part of diversification into ARM processor architectures. There are a few other examples of AWS developers providing improvements to critical libraries that meet or exceed performance on Intel processor architectures.</p><p><a href="https://www.blameless.com/blog/slo-adoption-twitter">SLO Adoption at Twitter</a><br>SLOs have the dual impact of defining/shaping understanding of service performance and plying cultural change to discuss team-to-team and service-to-service needs in the context of performance contracts. Traditionally, developers think of APIs as defining endpoints and messages, but in reality the APIs embody the full contract, to include limits, reliability, and yes, even performance.</p><p><a href="https://nicolodavis.com/blog/typescript-to-rust/">Moving from TypeScript to Rust / WebAssembly</a><br>WebAssembly is opening up Rust as the new same frontend-language-as-backend-language. The article here is light but I’m stashing it in this list as a pointer to a project that underwent a Typescript → Rust conversion, and as a data point for (a) this being possible and (b) the experience being reasonably positive.</p><p><a href="https://hbr.org/2017/07/the-more-senior-your-job-title-the-more-you-need-to-keep-a-journal">The More Senior Your Job Title, the More You Need to Keep a Journal</a><br>Self-introspection is an important aspect for work and for life, one pragmatic way to push on the introspection is to externalize experiences into a journal. Besides providing an outlet for sharing experiences where one may not otherwise be available, journals represent a way to review and understand history, trends, deepen memory and provide a basis for future learning. More repetitions with writing — even if only for self-consumption — also never hurt.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34cfc4837dee" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020/June — Reading List]]></title>
            <link>https://medium.com/@markelliot/2020-june-reading-list-c34eb9aa4fbc?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/c34eb9aa4fbc</guid>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Tue, 21 Jul 2020 19:06:03 GMT</pubDate>
            <atom:updated>2020-07-21T19:06:03.298Z</atom:updated>
            <content:encoded><![CDATA[<h3>2020/June — Reading List</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a> and recommendations from others. Here’s June’s list.</p><h3>2020/June</h3><p><a href="https://www.blameless.com/blog/slo-adoption-twitter">SLO Adoption at Twitter</a><br>Using SLOs as the basis for discussion around expectations, rationale, and informing not just when to shed load but the most appropriate way to shed load. In many ways, the article describes how SLOs assist in a culture of performance-focus, and as the domain/site header might imply, how to do that without seeing service owners as opposition, or placing blame.</p><p><a href="https://smartbear.com/blog/test-and-monitor/verification-and-validation-the-difference/">Verification and Validation: The Difference</a><br>Pulled this as up as an external reference to explain the difference in these concepts to some colleagues. Drilled in a Systems Engineering education, I remember these as <em>did we build it right?</em> and <em>did we build the right thing?</em> but likely the following definitions are a bit more portable, and on the face of it explainable:</p><ul><li>Verification — did we build what we said we would build?</li><li>Validation — did we build what we should’ve built?</li></ul><p><a href="https://danluu.com/metrics-analytics/">A simple way to get more value from metrics</a><br>Using offline analysis to find waste (and suggest improvements) for JVM processes by looking at standard JVM metrics across a large fleet. Notably embraces the idea that bulk analysis benefits from a more targeted approach than <em>all</em> metrics, choosing a common subset of likely-to-be-valuable metrics and feeding those into a data processing system capable of handling the reduced scale. Don’t miss the notes on using boring technology and its advantages near the end of the article.</p><p><a href="https://blog.twitter.com/engineering/en_us/topics/infrastructure/2019/metricsdb.html">MetricsDB: TimeSeries Database for storing metrics at Twitter</a><br>Evolving from using Twitter’s central storage system (Manhattan) to a fit-for-purpose timeseries database, largely to help keep developers focused on what matters while freeing up bandwidth from the Manhattan team. Like many systems of this family, MetricsDB specializes for time-indexed data and fits a partitioning model that balances the producers and expected consumption patterns, and because more is known about the data format, takes advantage of timeseries-specific compression strategies to make the solution not only more performant but much lower cost to run and operate.</p><p><a href="https://medium.com/apache-pinot-developer-blog/text-analytics-on-apache-pinot-cbf5c45d282c">Implementing full-text search in Apache Pinot</a><br>Using Lucene as a specialty index format for a columnar OLAP data system. We’ve been working on something similar and thought this spin was pretty interesting/informative for how one might go about integrating full-text indexes along-side columnar ‘big’ data.</p><p><a href="https://blog.roboflow.ai/yolov5-is-here/">YOLOv5 Is Here: State-Of-The-Art Object Detection At 140 FPS</a><br>Pretty much all in the title — YOLO v5 offers strong performance on real-time object detection, and is blazingly fast.</p><p><a href="https://www.technologyreview.com/2016/02/24/161885/google-unveils-neural-network-with-superhuman-ability-to-determine-the-location-of-almost/">Google Unveils Neural Network with “Superhuman” Ability to Determine the Location of Almost Any Image</a><br>Describes a paper from 2016 where Google trained PlaNet, a neural network that predicts the location of an image based on its pixels with error around 1000km and using only 377MB of memory.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c34eb9aa4fbc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020/May  —  Reading List]]></title>
            <link>https://medium.com/@markelliot/2020-may-reading-list-830ce3ebcf6f?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/830ce3ebcf6f</guid>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Tue, 21 Jul 2020 19:04:09 GMT</pubDate>
            <atom:updated>2020-07-21T19:15:32.206Z</atom:updated>
            <content:encoded><![CDATA[<h3>2020/May — Reading List</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a>, <a href="https://discoverdev.io/">discoverdev.io</a>, and recommendations from others.</p><p>As with April, May’s list is notably abbreviated and posted on a few months delay. There was and is lot happening in the world, and much of my reading and even my time (both at work and personal) was spent learning about the nature of the pandemic and getting more informed and engrossed in cultural discourse. Please consider skipping my list here in favor of the wealth of resources available about current world events.</p><p>Nevertheless, posting for completeness — here’s the brief summary of my technical and business related reading from May.</p><h3>2020/May</h3><p>State of Loom (<a href="https://cr.openjdk.java.net/~rpressler/loom/loom/sol1_part1.html">Part 1</a>, <a href="https://cr.openjdk.java.net/~rpressler/loom/loom/sol1_part2.html">Part 2</a>)<br>Great overviews of what, and more importantly <em>why</em> the Java team is investing in Loom, Java’s take on JVM-tier threads. Recently I’ve been impressed by Rust’s ability to observe, learn from, and build forward in a novel way from other systems — this article shares the Java team doing exactly that, embracing the idea that the tradeoffs between writing, observing and maintaining concurrent programs are at the wrong point on the spectrum.<br><em>Project Loom intends to eliminate the frustrating tradeoff between efficiently running concurrent programs and efficiently writing, maintaining and observing them. It leans into the strengths of the platform rather than fight them, and also into the strengths of the efficient components of asynchronous programming. It lets you write programs in a familiar style, using familiar APIs, and in harmony with the platform and its tools — but also with the hardware — to reach a balance of write-time and runtime costs that, we hope, will be widely appealing. It does so without changing the language, and with only minor changes to the core library APIs. A simple, synchronous web server will be able to handle many more requests without requiring more hardware.</em></p><p><a href="https://engineering.fb.com/web/facebook-redesign/">Rebuilding our tech stack for the new Facebook.com</a><br>Great walk-through of how developer-focused tooling enables a focused and strong user outcome — relevant content displayed to the user nearly immediately, and functionality filled in the most user-relevant ways. I found the developer tooling to both enable and maintain these outcomes particularly compelling.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=830ce3ebcf6f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020/April — Reading Digest]]></title>
            <link>https://medium.com/@markelliot/2020-april-reading-digest-6a9f74716b42?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a9f74716b42</guid>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Tue, 21 Jul 2020 19:02:33 GMT</pubDate>
            <atom:updated>2020-07-21T19:14:50.440Z</atom:updated>
            <content:encoded><![CDATA[<h3>2020/April — Reading Digest</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a>, <a href="https://discoverdev.io/">discoverdev.io</a>, and recommendations from others.</p><p>April’s list is notably abbreviated and posted on a few months delay. There was and is lot happening in the world, and much of my reading and even my time (both at work and personal) was spent learning about the nature of the pandemic and getting more informed and engrossed in cultural discourse. Please consider skipping my list here in favor of the wealth of resources available about current world events.</p><p>Nevertheless, posting for completeness — here’s the brief summary of my technical and business related reading from April.</p><h3>2020/April</h3><p><a href="https://slack.engineering/deploys-at-slack-cd0d28c61701">Deploys at Slack</a><br>A quick overview of how Slack’s software deployment system works and of its recent improvements, especially in terms of granularity of speed of deploy. Slack’s approach is to ensure deploys are atomic (changes apply consistently), fast (capable of handling 12 deploys/day to hundreds or thousands of servers) and tuning reliability (by using tiered gating systems). Atomic deploys function using directory swapping. Fast deploys happen through a pull-based model instead of a push-based model. Tiering functionality isn’t detailed in this post.</p><p><a href="https://community.bose.com/t5/Around-On-Ear-Headphones/Bose-QC-35-Firmware-4-5-2-Noise-Cancellation-Investigation/td-p/285738#anoverview">Bose QC 35 Firmware 4.5.2 Noise Cancellation Investigation Report</a><br>Bose provides transparency into their investigation around customer reports that a firmware upgrade resulted in worse perceived performance for hardware. I found the level of detail refreshing and illuminating, especially Bose’s description of its hardware test process, ability to isolate human perception as well as embrace it in a controlled environment.</p><p><a href="https://github.com/minio/sidekick">Sidekick: sidecar load balancer by MinIO</a><br>Middleware for balancing load at client call sites without a coordinated load balancer. Also supports limited caching for S3-compatible object stores with staleness checking to the underlying store, and the README includes a demonstration with Spark.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a9f74716b42" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020/March — Reading Digest]]></title>
            <link>https://medium.com/@markelliot/2020-march-reading-digest-69891304f164?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/69891304f164</guid>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Mon, 08 Jun 2020 18:20:01 GMT</pubDate>
            <atom:updated>2020-06-08T18:20:01.726Z</atom:updated>
            <content:encoded><![CDATA[<h3>2020/March — Reading Digest</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a>, <a href="https://discoverdev.io/">discoverdev.io</a>, and recommendations from others. Here’s this month’s digest.</p><h3>2020/March</h3><p><a href="https://labs.spotify.com/2020/03/17/what-the-heck-is-backstage-anyway/">What the Heck is Backstage Anyway?</a><br>Spotify explains their new open source software project — infrastructure for creating, deploying, and maintaining microservices.</p><p><a href="https://dropbox.tech/infrastructure/rewriting-the-heart-of-our-sync-engine">Rewriting the heart of our sync engine</a><br>Dropbox replaces one of the most fundamental and most critical parts of their infrastructure — the software that synchronizes files on a device to and from the servers. Amongst the many articulated lessons, the article helpfully provides a checklist that helps identify when a rewrite is a viable, and if it’s the best option. I especially appreciate the explicit explicit guidance for understanding and respecting the current system, basically a paraphrasing of <a href="https://en.wikipedia.org/wiki/Wikipedia:Chesterton%27s_fence">Chesterton’s Fence</a>. Pulling a couple of those lessons to the top:</p><ul><li><em>Do you deeply understand and respect the current system?<br>It’s much easier to write new code than fully understand existing code. So, before embarking on a rewrite, you must deeply understand and respect the “Classic” system. It’s the whole reason your team and business is here, and it has years of accumulated wisdom through running in production. Get in there and do archaeology to dig into why everything is the way it is.</em></li><li><em>What are your principles for the new system?<br>Starting from scratch is a great opportunity to reset technical culture for a team. Given our experience operating Sync Engine Classic, we heavily emphasized testing, correctness, and debuggability from the beginning. Encode all of these principles in your data model. We wrote out these principles early in the project’s lifespan, and they paid for themselves over and over.</em></li></ul><p>Besides the rewrite recommendations, Dropbox opted for a testing model that matches Google’s <a href="https://static.googleusercontent.com/media/research.google.com/en//archive/chubby-osdi06.pdf">Chubby</a> and FoundationDB’s development process:</p><ul><li><em>The Control thread is designed to be entirely deterministic when its inputs and scheduling decisions are fixed. We use this property to fuzz it with pseudorandom simulation testing. With a seed for our random number generator, we can generate random initial filesystem state, schedules, and system perturbations and let the engine run to completion. Then, if we fail any of our sync correctness checks, we can always reproduce the bug from the original seed. We run millions of scenarios every day in our testing infrastructure.</em></li></ul><p><a href="https://towardsdatascience.com/scalable-efficient-big-data-analytics-machine-learning-pipeline-architecture-on-cloud-4d59efc092b5">Architecture for High-Throughput Low-Latency Big Data Pipeline on Cloud</a><br>A decent overview of what it takes to build a reliable big data pipeline.</p><p><a href="https://blogs.oracle.com/javamagazine/java-14-arrives-with-a-host-of-new-features">Java 14 Arrives with a Host of New Features</a><br>Switch expressions, text blocks, and records are just a few of the exciting features in this release.</p><p><a href="https://www.filingdb.com/pdf-text-extraction">What’s so hard about PDF text extraction?</a><br>Like many widely-used and well aged file formats, PDFs are composed of a wandering specification with limited consistency, esoteric features and complex semantics focused on the format’s original, primary objective, at the cost, ironically, of easy machine readability. In the end, PDFs are designed to consistently show pixels on the screen regardless of the device or operating system. To read the text those pixels represent perfectly every time, in the limit of all the diverse ways the text makes it to the screen and all the diverse layouts the text my appear in, it’s simplest, and often easiest, to fall back to OCR.</p><p><a href="https://blog.acolyer.org/2020/03/02/firecracker/">Firecracker: lightweight virtualization for serverless applications</a><br>I’ve mentioned Firecracker here before, but this is a good technical walk-through of its mechanics, architecture and performance, worth a read for those interested in a layer or two down.</p><p><a href="https://blog.acolyer.org/2020/01/17/synthesizing-data-structure-transformations/">Synthesizing data structure transformations from input-output examples</a><br>I found this to be a fun exploration of learning applications — programs can be seen as boxes that transform inputs to produce outputs, what if we could show a learning machine a bunch of inputs and corresponding outputs, can we get a good, idiomatic program out?</p><p>Applications of something in this space would meaningfully improve the repeatability and reliability of complex data pipelines by enabling non-software-engineer users to provide iterative examples and verify a program’s correctness without actually having to author the program, and, at the same time, to allow machines to ensure the reliability, observability and security of that program as enforced, separate concerns.</p><p><a href="https://blog.acolyer.org/2019/09/30/small-world-with-high-risks/">Small world with high risks: a study of security threats in the npm ecosystem</a><br>We’re seeing increased exploits of package management systems as a means for breaking through security boundaries — deep dependency chains with huge numbers of packages and widely-used packages as a way to amplify reach make npm a high return target. We’ll see more attacks in the future. And, in this game of cat-and-mouse, we’ll see more protection and more theory around <a href="https://bytecodealliance.org/articles/announcing-the-bytecode-alliance">securing and authorizing dependencies</a> in a granular way to manage otherwise untrusted code.</p><p><a href="https://www.infoq.com/presentations/snowflake-architecture/">Snowflake Architecture: Building a Data Warehouse for the Cloud</a> + <a href="https://blog.acolyer.org/2020/03/09/snowflake/">Building an elastic query engine on disaggregated storage</a><br>A pair of good overviews for how Snowflake makes use of the unique parts of its runtime environment. Notably, uses isolated, elastic compute based on canonical data stores with differing read/write optimized instances and enables the isolated compute to re-render data in best-for-access formats. I suspect we’ll increasingly see systems repeat the idea that data has to be delivered in a format that’s fit-for-purpose, and systems will compromise not on performance or scalability, but on storage cost (as it continues to drop this matters less) and on small forms of data latency (because performance and consistency of performance will matter more than up-to-dateness, and up-to-dateness can be added back later).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69891304f164" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020/February — Reading Digest]]></title>
            <link>https://medium.com/@markelliot/2020-february-reading-digest-fdde6210df1d?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/fdde6210df1d</guid>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Tue, 17 Mar 2020 16:00:36 GMT</pubDate>
            <atom:updated>2020-03-18T18:48:38.031Z</atom:updated>
            <content:encoded><![CDATA[<h3>2020/February — Reading Digest</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a>, <a href="https://discoverdev.io/">discoverdev.io</a>, and recommendations from others. Here’s this month’s digest.</p><h3>2020/February</h3><p><a href="https://coded.io/2020/02/early-impressions-of-rust-from-a-go-programmer/">Early impressions of Rust from a Go programmer</a> &amp; <a href="https://pingcap.com/blog/early-impressions-of-go-from-a-rust-programmer/">of Go from a Rust programmer</a><br>A pair of articles that examine a Rust developer experimenting with Go and a Go developer experimenting with Rust on new projects. First impressions share common feedback about both languages:</p><ul><li>Go: fast compilation, fast learning curve, challenges due to lack of generics/built-in data structures</li><li>Rust: slow compilation, slow learning curve, high performance results with little effort, language and libraries make it easy to do things that might be hard (incl. generics and lots of helpful built-in data structures)</li></ul><p><a href="https://blog.acolyer.org/2020/02/26/meaningful-availability/">Meaningful availability</a><br>Traditionally, availability is measured either through time or count-based availability (and reported as proportional not-available-to-all-requests), but while this works well for services/APIs, it’s a relatively poor approximation of hosted, frontend applications. In particular, the application surface areas are large, and can result in scenarios of “one person’s uptime is another person’s downtime”. Using poor metrics might mean missing critical user-facing failures amongst the noise or overweighting non-critical failures during development.<br>Google proposes using sum-over-users uptime measurements as a way to ensure user-facing downtime counts sufficiently. Further, the user uptime measures are then computed as a function of varied windows to account for times between user actions, and then examined as a distribution of the minimal uptime for a given period over all window sizes. The resulting curve helps create actionable investment by revealing what duration of outages are leading to poor uptime.</p><p><a href="https://netflixtechblog.com/dblog-a-generic-change-data-capture-framework-69351fb9099b">DBLog: A Generic Change-Data-Capture Framework</a><br>Netflix describes both the challenges and solution for capturing changes from relational systems consistently, performantly, and under various kinds of stress events when attempting to populate downstream streaming systems.</p><p><a href="https://itnext.io/k8s-vertical-pod-autoscaling-fd9e602cbf81">K8s Vertical Pod Autoscaling</a><br>Suggest or automatically grow pods’ resource requests vertically as a function of metrics — in some cases it makes more sense (architecturally, economically, pragmatically) to scale up a pod rather than scale out pods/add additional replicas. Noteworthy that it’s not possible to combine this with a horizontal scaling strategy, though, editorializing, I think it’d be valuable to set both vertical and horizontal constraints to optimize over the economic/pragmatic (including time-to-execute) deployment sizing problem.</p><p><a href="http://btorpey.github.io/blog/2014/02/18/clock-sources-in-linux/">Measuring Latency in Linux</a><br>Somewhat surprisingly to me, the CLOCK_MONOTONIC source in Linux can be slewed forward as a result of time-adjustment code (such as NTP). Since System.nanoTime() in Java is based on CLOCK_MONOTONIC, this means that nanoTime calls are actually subject to NTP. While Linux includes CLOCK_MONOTONIC_RAW to gain access to processor-level monotonic time, that might be subject to environmental skews in a way which make it less reliable than an NTP adjusted monotonic clock. Bottom line: measuring time is hard — we’d all do well to have access to more reliable (and synchronized) clock sources.</p><p><a href="https://blog.acolyer.org/2020/02/17/software-20-migration/">Migrating a privacy-safe information extraction system to a Software 2.0 design</a><br>Interesting discussion around the migration path from an expert system (“Software 1.0“) to a machine-learned system (”Software 2.0“), both in the context of delivered advantages (less code, better results) and the challenges. Unsurprisingly, challenges are heavily focused on how to balance the need for labeled training data and pragmatic issues with privacy of the training data. Google opted for generating their own training data, and achieved great results. Even with generated data, though, the realities of an ML-based ecosystem are the same (and something <a href="https://www.palantir.com/palantir-foundry/">Foundry</a> explicitly <a href="https://vimeo.com/351744697">designs for</a>):<br><em>We argue that a key component for any such effort is a system to manage training data — including acquiring, debugging, versioning, and transforming it.</em></p><p><a href="https://rachelbythebay.com/w/2020/02/09/horizonta/">Trying to sneak in a sketchy .so over the weekend</a><br>Importance and impact of architecture constraints on design decisions — notably, sometimes development has to take a slightly longer path in the near-term for safety of existing operations. (As a bit of an editorial, to the extent development always has to take the long-path there’s probably something up with the architecture, but not the case in the linked article.)</p><p><a href="https://blog.acolyer.org/2020/02/07/cloudburst/">Cloudburst: stateful functions-as-a-service</a><br>New possibilities of colocating logic with data, and colocating logic with cooperative logic as a function of underlying data sharding. Principally: in a world of well-sharded databases, localize copies of related code next to/on the relevant database shards to improve performance and reduce overall network traffic, and colocate related logic to further improve performance/reduce traffic.</p><p><a href="https://blog.discordapp.com/why-discord-is-switching-from-go-to-rust-a190bbca2b1f">Why Discord is switching from Go to Rust</a><br>Discord took a metrics-first approach to tuning a core service that tracks read state of messages by user. Despite lots of focused, hand-tuned optimizations in their Go service, Discord is unable to get past a mandatory every-two-minutes garbage collect in Go that moves p95 response time from &lt;10ms to 300–400ms. Interestingly, the service carefully managed its memory and effectively should never have required any kind of collection to run, but Go has a backstop of GC-every-two-minutes regardless. Experimenting with a rewrite in Rust, Discord’s developers achieved the same top-line performance with almost no tuning, and because of Rust’s ownership-based memory-management didn’t suffer the bottom-line performance.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fdde6210df1d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020/January — Reading Digest]]></title>
            <link>https://medium.com/@markelliot/2020-january-reading-digest-9ed23e37cd72?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ed23e37cd72</guid>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Thu, 20 Feb 2020 20:57:37 GMT</pubDate>
            <atom:updated>2020-03-18T18:49:01.195Z</atom:updated>
            <content:encoded><![CDATA[<h3>2020/January — Reading Digest</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a>, <a href="https://discoverdev.io/">discoverdev.io</a>, and recommendations from others. Here’s this month’s digest.</p><h3><strong>2020/January</strong></h3><p><a href="http://web.eecs.utk.edu/~azh/blog/notebookpainpoints.html">What’s wrong with computational notebooks?</a><br>Notebooks are challenging for lots of reasons — non-linear execution, difficulty collaborating, difficulty understanding changes and history, and all of that leading to challenges with reproducing work, and we haven’t even gotten to security, yet. Rob Fink and I share some additional thoughts in <a href="https://medium.com/palantir/on-dataset-versioning-in-palantir-foundry-8f23de22cc4c">On dataset versioning in Palantir Foundry</a>.</p><p><a href="http://m.nautil.us/issue/81/maps/what-happens-to-google-maps-when-tectonic-plates-move">What Happens to Google Maps When Tectonic Plates Move?</a><br>I’m always pretty fascinated by this sort of thing — most of us wander around under the impression that latitude/longitude pairs uniquely locate a point on the Earth…but the ground is moving! Not a hugely in-depth article, but a good exploration of the conceptual challenges of what it means to locate something, somewhere, and what it takes to manage and rectify that knowledge over time.</p><p>Related: <a href="https://www.openstreetmap.org/user/StephaneP/diary/390290">You thought OpenStreetMap data uses the WGS84 datum? No it doesn’t!</a></p><p><a href="https://www.ultrasaurus.com/2020/01/writing-c-library-in-rust/">writing c library in rust</a><br>A great, in-depth walk-through of what it takes to write a library in Rust that’s ergonomically and natively accessible from C — just the kind of thing worth reading if you’re looking to eat your C codebase from the inside out.</p><p><a href="https://github.com/kubernetes/kops">https://github.com/kubernetes/kops</a><br>Tooling to launch production-grade Kubernetes (for some definition of production grade) across clouds with a number of convenient options for declarative state changes, including generation of Terraform code.</p><p><a href="https://iamturns.com/yarn-vs-npm-2018/">Why Are You Still Using Yarn in 2018?</a><br>Quick reminder that new technology that’s better now might not be better later. We’re active users of yarn, but maybe we should reconsider our choice and look at whether npm’s improvements now outpace the improvements we’ve seen with yarn. On the other hand, a lot of this ecosystem seems surprising in its behavior, maybe instead of better tools that replace each other we should think about different tools and tweaking the behavior.</p><p><a href="https://hacks.mozilla.org/2019/04/pyodide-bringing-the-scientific-python-stack-to-the-browser/">Pyodide: Bringing the scientific Python stack to the browser</a><br>Pyodide is a Mozilla initiative to get scientific Python into the browser, not just at the cursory <em>running of Python</em> but with strong interop between the Python runtime and JS. Check out a live notebook — <a href="https://alpha.iodide.io/notebooks/300/">https://alpha.iodide.io/notebooks/300/</a> — for a good walk-through of the capability. Spoiler: your Python can mutate the DOM, and your JS can invoke your Python functions. Under the covers Pyodide is CPython + NumPy + SciPy emscripten’d to WASM with sufficient WASI implementations for the filesystem to live resident in browser memory, and a few extra goodies like Matplotlib that renders correctly into the browser.</p><p>We’re starting to explore how we might lean on some of this capability as we think about our web-based IDEs, and have an interesting demo of r<a href="https://github.com/iodide-project/pyodide/pull/602">unning the Python Language Server we maintain with Monaco, all running in the browser</a>.</p><p><a href="https://hollow.how/">Hollow (Netflix OSS)</a><br>Netflix-built single-producer many-consumer Java library for read-optimized workflows. The single-producer model helps with providing low-latency updates to the data and the Java library offers a number of strategies for getting medium-scale data and its indexes to fit sufficiently well in memory so that applications with a need for read performance see low-latency replicas in formats that fit their own access pattern.</p><p><a href="https://blog.twitter.com/engineering/en_us/topics/infrastructure/2019/daperture-load-balancer.html">Deterministic Aperture: A distributed, load balancing algorithm</a><br>Twitter engineers describe a client-side (that’s the <em>distributed</em> part) load-balancing change that dramatically improves the total service load through more balanced call traffic. Over-simplifying: clients use a token-ring-like assignment scheme with limited <em>aperture</em> (i.e. choose 3 of N) for failover — the token-ring assignment helps distribute load evenly amongst clients, limiting aperture helps ensure that in failure/restart/upgrade scenarios that load doesn’t consolidate to the same few nodes, while still providing sufficient breadth for actual failover.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ed23e37cd72" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2019/December — Reading Digest]]></title>
            <link>https://medium.com/@markelliot/2019-december-reading-digest-fed2254a22e7?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/fed2254a22e7</guid>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Mon, 06 Jan 2020 13:55:04 GMT</pubDate>
            <atom:updated>2020-01-06T13:55:04.626Z</atom:updated>
            <content:encoded><![CDATA[<h3>2019/December — Reading Digest</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Keeping with the public version of <a href="https://medium.com/@markelliot/2019-november-reading-digest-4a74c4341769">that series</a> I started last month, here’s December’s digest. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a>, <a href="https://discoverdev.io/">discoverdev.io</a>, and recommendations from others.</p><h3>2019/December</h3><p><a href="https://medium.com/pinterest-engineering/gokul-extending-time-series-data-storage-to-serve-beyond-one-day-52264307364d">GokuL: Extending time series data storage to serve beyond one day</a><br>Pinterest talks about extending Goku (their timeseries database) to support rolled-up histories of otherwise dense timeseries data and the data tiering and compaction associated with maintaining these rollups. Generally we’ve eschewed these things at Palantir in favor of reasonably smart compression of timeseries data, though the rollup and data tiering resonates as we look at schemes to assist with pruning data from queries — and in fact, many of our systems keep roll-up like statistics to do just that.</p><p><a href="https://samcurry.net/filling-in-the-blanks-exploiting-null-byte-buffer-overflow-for-a-40000-bounty/">Filling in the Blanks: Exploiting Null Byte Buffer Overflow for a $40,000 Bounty</a><br>Exploiting form inputs via encoded null bytes (%00) and gaps between the sanitized string parsing and un-sanitized string length functions to read whatever happens to be nearby in memory.</p><p><a href="https://acoup.blog/2019/05/10/collections-the-siege-of-gondor/">Collections: The Siege of Gondor, Part I: Professionals Talk Logistics</a><br>Professional historian Bret Devereaux analyzes the film version of LotR’s Siege of Gondor for accuracy, intent, and some pop-culture amusement. The full six-part series is well worth the read, and covers several aspects of pre-modern warfare that might be less obvious. The cultural elements were particularly insightful — that pre-modern war was as much (if not more) in the mind than on the field, and found the lessons especially in the later parts of the series an interesting lens when watching the latest Star Wars installment.</p><p><a href="https://mobile.twitter.com/david_perell/status/1208930197702995968">These Amazon writing tips are fantastic…</a><br>Summarizing this Tweet’s attached image:</p><p>Use less than 30 words per sentence.</p><ul><li>Due to the fast that → because</li><li>Totally lacked the ability to → could not</li></ul><p>Replace adjectives with data.</p><ul><li>We made the performance much faster → we reduced server side tp90 latency from 10ms to 1ms</li></ul><p>Eliminate weasel words.</p><ul><li>Nearly all customers → 87% of Prime members</li><li>Significantly better → +25 basis points (bps)</li></ul><p>Does your writing pass the “so what” test?</p><p>If you get a question, reply with one of the four Amazon answers:</p><ul><li>Yes.</li><li>No.</li><li>A numbers.</li><li>I don’t know (and will follow up when I do).</li></ul><p><a href="http://paulgraham.com/fp.html">Fashionable Problems</a><br><em>If you can find a new approach into a big but apparently played out field, the value of whatever you discover will be </em><a href="http://paulgraham.com/sun.html"><em>multiplied</em></a><em> by its enormous surface area. … Then you’ll continue to work on it even if you make the same mistake as other people and think that it’s too marginal to matter.</em></p><p><a href="https://blog.rust-lang.org/2019/12/19/Rust-1.40.0.html">Announcing Rust 1.40.0</a><br>Pretty cool to see non_exhaustive enums as a language-level feature. non_exhaustive on enums forces consumers of a library to deal with potentially unknown future values, and explicitly informs developers that there could be new values. Besides the signaling, this helps library developers convey that something is still a work in progress. We did something similar for Conjure enums with the idea that APIs evolve, and consumers (and the libraries they use to consume Conjure-defined APIs) need to safely handle potentially unknown values.</p><p><a href="https://lemire.me/blog/2019/12/19/xor-filters-faster-and-smaller-than-bloom-filters/">Xor Filters: Faster and Smaller Than Bloom Filters</a><br>Quick introduction to an <a href="https://arxiv.org/abs/1912.08258">academic article</a> by Daniel Lemire on Xor filters, that offer better memory density and better performance at the same false positive rate for probabilistic set membership. This introduction mostly explains the motivation, points to the article, and to the implementations in many common programming languages. The paper itself is extremely accessible, skip to section 3 for the algorithm details.</p><p>One notable change from Bloom filters that’s likely to match many use-cases: Xor filters are immutable, constructed once based on the known membership rather than mutated over time for membership additions.</p><p>Not discussed: whether Xor filters naturally extend as counting Xor filters in the same way Bloom filters extend to <a href="https://en.wikipedia.org/wiki/Counting_Bloom_filter">counting Bloom filters</a> (from a very quick skim it seems like it should be possible).</p><p><a href="https://ionutbalosin.com/2019/12/jvm-garbage-collectors-benchmarks-report-19-12/">JVM Garbage Collectors Benchmarks Report 19.12</a><br>Good and very approachable investigation into JVM garbage collectors and their performance under a number of potentially hard-for-GC circumstances. Maybe unsurprisingly, Shenandoah and ZGC (Java’s newest GC implementations) tend to win out in performance focused benchmarks, and Serial GC (Java’s oldest GC implementation) tends to win out in size/footprint focused benchmarks.</p><p><a href="https://hbr.org/2017/11/many-strategies-fail-because-theyre-not-actually-strategies">Many Strategies Fail Because They’re Not Actually Strategies</a><br><em>A real strategy involves a clear set of choices that define what the firm is going to do and what it’s </em>not<em> going to do. Many strategies fail to get implemented, despite the ample efforts of hard-working people, because they do not represent a set of clear choices. …<br>Another reason many implementation efforts fail is that executives see it as a pure top-down, two-step process: “The strategy is made; now we implement it.” That’s unlikely to work. A successful strategy execution process is seldom a one-way trickle-down cascade of decisions.</em></p><p><a href="https://cloud.google.com/blog/products/compute/understanding-dynamic-resource-management-in-e2-vms">Performance-driven dynamic resource management in E2 VMs</a><br>Pretty cool advancement in resource management to be able to shift a VM’s underlying execution machine in order to both better utilize resources and to help that VM execute with better performance (for some definition of better). Also notable that the first tactic is to improve scheduling based on past behavior — I suspect a lot of cloud workload doesn’t vary significantly over medium timespans.</p><p><a href="https://queue.acm.org/detail.cfm?ref=rss&amp;id=3375635">API Practices If You Hate Your Customers</a><br>Fun satire about the qualities of a good API…and a good API experience. Notably, a lot of what gets discussed here is about how customers interact with the API, the documentation, the sign-up process and the authentication/authorization procedure. Too often we look at APIs as just the <em>interface</em>, but our customers experience our APIs as so much more.</p><p>Among the many reasons we built <a href="https://palantir.github.io/conjure/">Conjure</a>, we wanted to (a) standardize the authn/z setup and (b) make the API definitions more obvious to passive consumers. We have some work to do in terms of documentation rendering and general tooling for editing/consuming Conjure definitions, though.</p><p><a href="https://github.com/docker-slim/docker-slim">docker-slim</a><br>Kind of like tree-shaking for your Docker container. Likely some pitfalls to work through, but maybe an interesting approach to lighter-weight images. More generally, I suspect Google’s <a href="https://github.com/GoogleContainerTools/distroless">distroless</a> containers offer some advantages — adding what you need vs. removing what you don’t means the additions are known, justified (through versioned changes), and makes it more likely that the containers only contain what’s actually needed, not just reducing size but also reducing risk surface.</p><p><a href="https://blog.jonlu.ca/posts/spf-dkim">Email authentication: SPF, DKIM and DMARC out in the wild</a><br>Great introduction to SPF, DKIM and DMARC — what they are and how to use them. <br>Plus: some basic research for uptake in popular domains.<br>Also, this <a href="https://www.quora.com/What-are-many-tech-companies-getting-wrong-right-now-when-it-comes-to-data-security">resonates</a>:<em> The ramifications of not building authentication and security into your protocol when it’s designed can be far reaching.</em></p><p><a href="http://adventures.michaelfbryan.com/posts/how-to-riir/index.html">How to RiiR</a><br>A good, specific walk-through of translating a real library from C to Rust, one function at a time. Like many migrations, this approach benefits from two specific advantages:</p><ul><li>Focusing on direct replacement rather than replacement + improvement (I think this is perhaps the most common pitfall of these sorts of translations, and an easy trap since embellishing something while in the weeds is easy to do).</li><li>Enabling invocation of old code while writing new code.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fed2254a22e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2019/November — Reading Digest]]></title>
            <link>https://medium.com/@markelliot/2019-november-reading-digest-4a74c4341769?source=rss-ee08a1ec175c------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a74c4341769</guid>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Mark Elliot]]></dc:creator>
            <pubDate>Mon, 02 Dec 2019 13:20:25 GMT</pubDate>
            <atom:updated>2019-12-02T13:20:25.127Z</atom:updated>
            <content:encoded><![CDATA[<h3>2019/November — Reading Digest</h3><p>Starting a few years ago I began keeping a list of interesting articles I’d come across, roughly sorted by month, sometimes with notes. Over time, friends and colleagues have asked me to share the list, and roughly monthly I send an email containing a digest — a friend asked me if I’d consider publishing on Medium, so in addition to sharing over email, I’ll also post monthly digests here. Articles are typically sourced from some sampling of <a href="https://news.ycombinator.com/">Hacker News</a>, <a href="https://lobste.rs/">Lobste.rs</a>, <a href="https://discoverdev.io/">discoverdev.io</a>, and recommendations from others.</p><h3>2019/November</h3><p><a href="https://www.usenix.org/system/files/conference/atc18/atc18-yang.pdf">NanoLog: A Nanosecond Scale Logging System</a><br>In the same vein: logging commonly produces the same strings repeatedly, so one simple idea is to maintain a dictionary of all the strings and to emit dictionary keys rather than the values, using common approaches to optimize the assignment and compression of keys. One could go a step further and make use of a <a href="https://github.com/palantir/metric-schema">Metric Schema</a> style approach for defining, at compile time, that dictionary.</p><p>This paper runs a compile-time analysis to back out the “log schema” and create the dictionary and shows that programs can get to an order of magnitude fewer instructions per log line, gaining an order of more magnitude of throughput in the logging system, making logs free-enough to run consistently in production.</p><p><a href="https://github.com/fuzzitdev/javafuzz">Javafuzz: coverage-guided fuzz testing for Java</a><br>Coverage-guided fuzzing can be helpful in rooting out bugs, nice to have some tooling to do this in Java.</p><p><a href="https://bytecodealliance.org/articles/announcing-the-bytecode-alliance">Announcing the Bytecode Alliance: Building a secure by default, composable future for WebAssembly</a><br>Library security is <a href="https://github.com/dominictarr/event-stream/issues/116#issuecomment-441759442">something</a> that’s <a href="https://hackernoon.com/im-harvesting-credit-card-numbers-and-passwords-from-your-site-here-s-how-9a8cb347c5b5">come</a> <a href="https://hackernoon.com/part-2-how-to-stop-me-harvesting-credit-card-numbers-and-passwords-from-your-site-844f739659b9">up</a> <a href="https://github.com/rest-client/rest-client/issues/713">recently</a>, and the idea here goes a bit further by:</p><ol><li>Explicitly provisioning access rights to libraries</li><li>Isolating libraries and invoking over a new RPC system designed to keep memory safety while adding minimal copy/call latency</li></ol><p>The idea seems ambitious, and generally speaks to a principal around including 3rd party code being risky in a way that the industry is under-addressing right now. Our general approach: assume the whole package is unsafe and limit the outer-bounds and what it can do — one should generally not allow all code to access the internet if at all possible, but similarly not all code should have access to storage, and it’s surprising we’re not also addressing ways to perform this lockdown (perhaps because containers take us so much further on that path?).</p><p><a href="https://vcstarterkit.substack.com/p/how-vcs-make-money">How VCs Make Money</a><br>Makes good on the title and provides a good rundown of the dynamics used for VCs to make money. Note that a lot of the return for VCs comes from the compounded 2 in the 2-and-20, much like hedge funds, but that the ability to leverage that mechanic does generally require some base reputation (and thus some base kind of fund performance). OTOH, the returns are over such a long period, the performance doesn’t necessarily have to persist.</p><p><a href="https://devblogs.nvidia.com/grcuda-a-polyglot-language-binding-for-cuda-in-graalvm/">grCUDA: A Polyglot Language Binding for CUDA in GraalVM</a><br>Truffle (GraalVM’s universal language adapter) provides natural interop between languages so that polyglot implementations can easily and efficiently pass data and cross-call compute, and provides a natural way to make it easy to embed CUDA instructions fluently into any Truffle supported language. Cool opportunities here to expose CUDA in Java, Python, Ruby and even JS/TS.</p><p><a href="https://blog.acolyer.org/2019/11/11/snap-networking/">Snap: a microkernel approach to host networking</a><br>Summary of Google’s replacement for the TCP/IP stack in user-space, including a discussion of how to accomplish the user-space replacement without embedding a library in services — the gist is to use unix sockets or shared memory as a local process bridge. Some additional discussion on rollout procedure, coordination, and the challenge of having multiple versions in the wild due to sequenced rollouts.</p><p><a href="https://engineering.linkedin.com/blog/2019/an-inside-look-at-linkedins-data-pipeline-monitoring-system-">An inside look at LinkedIn’s data pipeline monitoring system</a><br>LinkedIn’s take on dataset-centric data monitoring system that measures not only whether data is flowing/jobs are executing but also whether the data content meets expectation. Notably talks about a three-state job outcome (succeeded, failed, late/stuck) and documents a number of automated remediations based on symptoms through a system called Nurse. Similar to Foundry, simple rules are encoded as post-build checks, more complicated rules (like those involving more than one dataset) seem to be built themselves as datasets.</p><p><a href="https://www.kalzumeus.com/2019/10/28/tether-and-bitfinex/">Tether: The Story So Far</a><br>Patrick Mackenzie’s take on Tether, cryptocurrency and prediction of impending doom in the crypto markets similar to 2008 doom in the traditional financial markets. See also his notes on the <a href="https://gist.github.com/patio11/598ec35c6c1675c97d93383f41b39b0b">insolvency of Mt. Gox</a>. Both are good reads and insightful leaning on evidence to tell the story. Plus, see some opinionated predictions.</p><p><a href="https://www.theguardian.com/lifeandstyle/2019/nov/06/i-was-an-astrologer-how-it-works-psychics">I was an astrologer — here’s how it really works, and why I had to stop</a><br>Most things that seem like magic are sufficiently sophisticated to be hard-to-perceive by the human mind, and that’s worth remembering when looking at advanced technology, too.</p><p><a href="https://blog.acolyer.org/2019/11/06/ceph-evolution/">File systems unfit as distributed storage backends: lessons from ten years of Ceph evolution</a><br>Ceph finds large performance improvement and overall less complexity by using raw disks directly instead of fighting the abstractions of various filesystem implementations, and providing a way to optimize for the kind of raw disk. Possibly see this as re-learning the “you should own one-level below what you do”, and if you build a distributed file store, the level below is the filesystem.</p><p><a href="http://august.nagro.us/optimized-openjdk.html">http://august.nagro.us/optimized-openjdk.html</a><br>Comparing compilation of the JDK targeting a specific architecture and the impact of -O3 (what the JDK typically ships with) and -Ofast (OP locally compiled).</p><p><a href="https://arxiv.org/pdf/1711.06368v2.pdf">Mobile Video Object Detection with Temporally-Aware Feature Maps</a><br>LSTMs embedded in a CNN to create temporal relationships between subsequent frames + bottlenecked LSTM for computational cost reduction.</p><p><a href="https://arxiv.org/pdf/1903.10172v1.pdf">Looking Fast and Slow: Memory-Guided Mobile Video Object Detection</a></p><ul><li>On slower sample rate, use a slow CNN to perform initial object detections</li><li>On faster sample rate, feed the results of the slow-path into a fast-path with far more limited features</li></ul><p><a href="https://arxiv.org/abs/1603.05279">XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks</a><br>Quantizing weights to 1-bit for a computational speed-up that outpaces the need for a larger network, plus quantized input filters for even larger gains.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a74c4341769" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>