---
title: "Latest reads"
description: "The engineering internet, summarised so you can actually read it."
---

# Latest reads
> The engineering internet, summarised so you can actually read it.

## Articles

### [Accelerating PyTorch distributed fine-tuning with Intel technologies](https://yomu.fyi/post/accelerating-pytorch-distributed-fine-tuning-with-intel-technologies.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Julien Simon
- Published: Nov 19, 2021
- Type: Tutorial

Transfer learning makes CPU-based clusters a viable, cost-effective option for shorter deep learning fine-tuning jobs. To accelerate PyTorch workloads, clusters can leverage Intel Ice Lake architecture featuring AVX-512 and Vector Neural Network Instructions alongside specialized software optimizations. Implementing the Intel Extension for PyTorch and the oneAPI Collective Communications Library (oneCCL) addresses communication bottlenecks across distributed nodes. In demonstrated tests fine-tuning transformer models on GLUE tasks across Amazon EC2 c6i.16xlarge instances, two-node jobs delivered roughly a 1.7x speedup, while four-node setups achieved a 3x speedup. These performance libraries and multi-node configurations maintain consistent scaling across both smaller datasets like MRPC and larger benchmarks like QQP.


### [Launching Livebook using LiveView](https://yomu.fyi/post/launching-livebook-using-liveview.md)
- Company: [Fly.io](https://yomu.fyi/company/fly-io.md)
- Published: Nov 18, 2021
- Type: Problem & solution

Fly.io’s Livebook launcher uses a single LiveView process to deploy a Livebook instance, but temporarily held deployment secrets could be lost when a user refreshed the browser or the server restarted. The secrets, including liveview\_password, were sent to deployment and stored in Vault, so the values could not be retrieved; keeping them in a database or Redis also conflicted with avoiding persistent server-side storage. The implementation encrypts the private state with Phoenix.Token.encrypt/4, stores the resulting blob in browser SessionStorage, and uses LiveView hooks and websocket events to restore it through Phoenix.Token.decrypt/4. After recovery, the UI resumes tracking deployment progress, then clears the cached data when deployment succeeds or fails, while closing the tab also clears SessionStorage.


### [Bootstrapping a UI component library](https://yomu.fyi/post/bootstrapping-a-ui-component-library.md)
- Company: [Ramp](https://yomu.fyi/company/ramp.md)
- Author: Nicolás Bevacqua
- Published: Nov 15, 2021
- Type: Explainer

Ryu is a component library created at a small startup to separate design concerns from feature development and business logic while product and conventions continue changing. It mirrors Figma's design system through shared colors, sizes, fonts, line heights, media breakpoints, z-indices, and components such as inputs, buttons, tooltips, menus, tables, and drawers. The article argues for a hard boundary between the library and main codebase, using a separate repository and an npm package with documented interfaces, hand-crafted changelogs, versioned releases, and extensive TypeScript. It emphasizes resisting narrowly tailored settings, postponing unresolved use cases, and prioritizing adoption over completeness, including preventing new code from using legacy components without good reason. Building on mature existing components while simplifying their interfaces, the library aims to improve maintainability, consistency, accessibility, theming, contribution, and coordination across applications.


### [Fine-Tune XLSR-Wav2Vec2 for low-resource ASR with 🤗 Transformers](https://yomu.fyi/post/fine-tune-xlsr-wav2vec2-for-low-resource-asr-with-transformers.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Patrick von Platen
- Published: Nov 15, 2021
- Type: Tutorial

Fine-tuning the multilingual XLS-R speech model adapts pretrained cross-lingual audio representations to automatic speech recognition tasks with limited labeled data. The demonstration configures a Wav2Vec2-XLS-R-300M checkpoint using Connectionist Temporal Classification on roughly four hours of validated Turkish audio from Common Voice. Building the pipeline requires pairing a Wav2Vec2FeatureExtractor for audio signals with a custom Wav2Vec2CTCTokenizer derived from dataset transcriptions. Training with Hugging Face Trainer over 3,200 steps decreases the validation word error rate from 0.7000 down to 0.3195. While the resulting transcription demonstrates recognizable phonetic alignment, output quality can be further improved by extending training schedules, refining preprocessing, and adding language model decoding.


### [Using Propensity Score Matching to Uncover Shopify Capital’s Effect on Business Growth](https://yomu.fyi/post/using-propensity-score-matching-to-uncover-shopify-capital-s-effect-on.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Nov 11, 2021
- Type: Benchmark

Shopify’s data team sought to estimate whether accepting Shopify Capital for the first time increased merchants’ subsequent sales, without randomly denying funding in an A/B test. They compared US first-time adopters in January 2019 with Canadian shops that later pre-qualified for Capital, using propensity scores built from Shopify sales and performance characteristics and matching them with a caliper-based greedy nearest-neighbor algorithm. After matching about 600 treated shops against roughly 8,000 controls, they evaluated log-transformed cumulative GMV from February through July 2019 with a binary regression and checked balance using standardized mean differences, visual comparisons, and variance ratios. The estimated geometric-average GMV was 36% higher for US merchants, with a 95% confidence interval of 13% to 65%; robustness checks supported a positive treatment effect.


### [Building Blocks of High Performance Hydrogen-powered Storefronts](https://yomu.fyi/post/building-blocks-of-high-performance-hydrogen-powered-storefronts.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Nov 8, 2021
- Type: Explainer

Hydrogen is a React-based framework for custom Shopify storefronts, designed to balance fast loading, personalized and dynamic commerce, and developer experience. The central trade-off is that static generation and edge delivery can speed loading but defer personalization, while conventional server-side rendering can slow time-to-first-byte as data is fetched. Hydrogen addresses this with React 18 alpha streaming SSR powered by Suspense, which streams the page shell before data resolves and progressively hydrates components without extra client round trips. React Server Components keep server-only code out of bundles, support server-side data access, and enable component-level updates that preserve client state. Built-in fetching and configurable caching, including stale-while-revalidate, work with Oxygen’s colocated runtime, while developers can mix server rendering, client fetching, and edge delivery at page or component level.


### [Supabase Beta October 2021](https://yomu.fyi/post/supabase-beta-october-2021.md)
- Company: [Supabase](https://yomu.fyi/company/supabase.md)
- Author: Paul Copplestone
- Published: Nov 7, 2021
- Type: Announcement

Supabase announced its October 2021 beta release, introducing three new authentication providers for Slack, Spotify, and MessageBird phone logins. The platform rolled out database multi-schema support, enabling users to browse different database schemas directly within the Dashboard interface and query them through the PostgREST API after enabling schema access. Documentation and instructional resources were expanded to cover PostgreSQL functions, authentication server-side functions, RedwoodJS quickstarts, self-hosting architectures, and production readiness checklists. In addition to ecosystem releases such as supabase-py v0.0.3, nuxt-supabase v2.2.1, and vue-supabase v2.2.3, the project reached over 21,268 GitHub stars. The team also announced Launch Week III: The Trilogy, set to begin on November 29th on Discord.


### [The Vitality of Core Web Vitals](https://yomu.fyi/post/the-vitality-of-core-web-vitals.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Nov 4, 2021
- Type: Tutorial

Core Web Vitals were introduced by Google in 2020 as unified guidance for web UX, and their loading, interactivity, and visual stability metrics became one factor in ranking systems after rollout from mid-June through the end of August 2021. The article explains LCP, FID, and CLS, including how LCP identifies the largest visible element, FID measures delay before the browser responds to a first interaction, and CLS sums unexpected layout shifts. It gives good and poor thresholds of 2.5 seconds and 4 seconds for LCP, 100 milliseconds and 300 milliseconds for FID, and 0.1 and 0.25 for CLS, with intermediate scores needing improvement. Recommended improvements include updating dependencies and tree-shaking unused library code, preloading critical assets, improving server response time, lazy-loading below-the-fold content, compressing images, and reserving image space to prevent shifts.


### [Scaling up BERT-like model Inference on modern CPU  - Part 2](https://yomu.fyi/post/scaling-up-bert-like-model-inference-on-modern-cpu-part-2.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Ella Charlaix, Jeff Boudier, Morgan Funtowicz, Michael Benayoun
- Published: Nov 4, 2021
- Type: Problem & solution

Scaling transformer inference on modern CPUs requires coordinating hardware capabilities with software optimizations across memory allocation, parallel execution, and mathematical kernel libraries. Intel Ice Lake Xeon processors deliver up to seventy-five percent faster inference across natural language processing tasks compared to Cascade Lake by combining Sunny Cove architecture improvements with dedicated extensions like Intel Extension for PyTorch. Fine-tuning software knobs such as alternative allocators, threading runtimes like Intel OpenMP, and core allocations can involve thousands of possible configuration combinations. Rather than evaluating these combinations through exhaustive brute force, Bayesian optimization via Intel SigOpt achieves comparable latency within roughly twenty trials and reveals experiment-dependent parameter importance. These software tuning strategies enable low-latency production transformer serving without dedicated accelerators.


### [Always on Calibration with a Quarterly Summary](https://yomu.fyi/post/always-on-calibration-with-a-quarterly-summary.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Nov 2, 2021
- Type: Problem & solution

Annual or infrequent performance reviews can create surprises, incomplete records of contributions, recency bias, and missed opportunities for timely feedback, correction, and career growth. The proposed alternative is “always-on” calibration, which treats individual development like continuous deployment: managers and team members repeatedly align expectations, discuss impact, and identify just-in-time opportunities. Each quarter, team members create a living document with intended outcomes, top and other accomplishments, upcoming opportunities, coworker feedback, and a quarterly review connecting impact to organizational expectations for role and level. They align on outcomes at the quarter’s first one-on-one and update accomplishments and feedback weekly before subsequent one-on-ones. Quarterly rollups provide smaller contribution snapshots, enable performance ratings and earlier course correction, and distinguish missed personal-development outcomes from missed role or level expectations.


### [Lessons from my internship at Ramp](https://yomu.fyi/post/lessons-from-my-internship-at-ramp.md)
- Company: [Ramp](https://yomu.fyi/company/ramp.md)
- Author: Nathan H. Leung
- Published: Oct 28, 2021
- Type: Explainer

The author reflects on a three-month software engineering internship at Ramp from March to June, describing lessons about execution, company choice, and location. Ramp’s charge card appeared to be a commodity, but the author credits its bias for shipping quickly, technical and product strength, differentiated technology, and favorable capital secured through finance expertise with turning it into a distinctive product. Compared with more formal, lower-stakes internships, Ramp offered flexibility, mentorship, and exposure to a fast-growing organization, although its pace and limited documentation could be challenging. The experience also shaped the author’s view that working on large problems at such a company can magnify impact, while New York provided experiences and a tech community distinct from San Francisco’s.


### [Supabase $30m Series A](https://yomu.fyi/post/supabase-30m-series-a.md)
- Company: [Supabase](https://yomu.fyi/company/supabase.md)
- Author: Paul Copplestone
- Published: Oct 28, 2021
- Type: Announcement

Supabase announced a $30 million Series A funding round led by Coatue, bringing its total funding to $36 million. The company aims to provide an open source alternative to Firebase while addressing the scalability limitations typical of abstracted database platforms. Rather than building a compatibility layer, Supabase provisions a full PostgreSQL database for every project alongside integrated authentication, storage, RESTful APIs, and realtime data streaming. Platform features rely directly on core PostgreSQL primitives, including Row Level Security policies, database triggers, and logical replication streams. Funding will support long-term development of a cloud-native PostgreSQL platform offering database branching, distributed read-write nodes, scalable storage, and ephemeral compute that scales down to zero.


### [The AWARE Development Plan](https://yomu.fyi/post/the-aware-development-plan.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Oct 28, 2021
- Type: Tutorial

The AWARE Development Plan presents a structured way to pursue long-term goals in software development, education, dance, or other areas of interest. David Ward developed it after seeing team members struggle to define their growth paths and describes using a similar process for his own goals. AWARE stands for Aim, Work Back, Achieve, Reflect, and Encore: define the desired outcome, work backward into smaller aims, execute measurable objectives, learn from the result, and maintain the plan. Short- and medium-term aims act as checkpoints, allowing people to identify missing experience, seek opportunities, and course-correct. Ward recommends reviewing the plan with a lead or manager, including during weekly one-on-ones, to select next steps, connect opportunities to aims, and ensure completed aims are reflected on.


### [Shopify's Path to a Faster Trino Query Execution: Custom Verification, Benchmarking, and Profiling Tooling](https://yomu.fyi/post/shopify-s-path-to-a-faster-trino-query-execution-custom-verification-b.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Oct 27, 2021
- Type: Problem & solution

Shopify needed safer ways to evaluate Trino configuration changes and software upgrades as its forked, hundreds-node cluster handled more data and data scientists depended on fast queries. The Data Reliability team built a lightweight Python library that centralizes Trino query orchestration, then exposed it through PyTest fixtures for repeatable verification, including correctness checks and candidate-versus-control comparisons. The broader framework also supports benchmarking and profiling concepts such as replaying historical queries, reproducing traffic patterns, and testing autoscaling, although the project ultimately focused most of its performance-testing scope on verification. Extensive correctness tests helped the Data Foundations team complete a successful Trino upgrade, while p95 query execution time remained stable over the reported period; the library was designed to leave room for more advanced interfaces later.


### [Course Launch Community Event](https://yomu.fyi/post/course-launch-community-event.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Sylvain Gugger
- Published: Oct 26, 2021
- Type: Announcement

Hugging Face scheduled the release of Part 2 of its machine learning course for November 15th alongside a collaborative community event. While Part 1 covered text classification and Model Hub uploads, Part 2 expands into token classification, causal and masked language modeling, translation, summarization, question answering, and ecosystem tools like Datasets and Tokenizers. The event includes two days of technical talks from industry practitioners covering transfer learning, ethics in machine learning development, and workflows across Keras, PyTorch, and TensorFlow. Participants collaborate on team projects to fine-tune natural language processing models and develop interactive portfolio demos, with AWS supplying complimentary compute through Amazon SageMaker. Attendees who successfully complete and demonstrate a project receive a certificate of completion.


### [Large Language Models: A New Moore's Law?](https://yomu.fyi/post/large-language-models-a-new-moore-s-law.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Julien Simon
- Published: Oct 26, 2021
- Type: Problem & solution

Recent advancements in generative artificial intelligence have spurred the rapid growth of large language models like Megatron-Turing NLG 530B, which require vast financial investments and substantial energy consumption. Training these massive parameter architectures demands hundreds of multi-GPU servers and generates significant carbon footprints for relatively modest benchmark gains. Rather than relying on brute-force scaling, engineering teams can adopt smaller, more frugal architectures through knowledge distillation and transfer learning techniques. Practical alternatives include fine-tuning existing pretrained models, deploying on energy-efficient cloud infrastructure, and applying optimizations like pruning, layer fusion, and quantization. These methods deliver low-latency inference and high task accuracy while significantly decreasing hardware requirements, development time, and environmental impact.


### [Train a Sentence Embedding Model with 1B Training Pairs](https://yomu.fyi/post/train-a-sentence-embedding-model-with-1b-training-pairs.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Antoine SIMOULIN
- Published: Oct 25, 2021
- Type: Announcement

Sentence embeddings map sentences into dense vectors to capture semantic meaning for downstream tasks such as clustering, asymmetric question answering, and text search. To train general-purpose models at scale, the team utilized a dataset compiled from multiple sources containing up to one billion sentence pairs. Training leveraged contrastive learning with Multiple Negative Ranking Loss and in-batch negatives, running across seven TPU v3-8 accelerators with JAX and Flax. Batches combined samples from at least two datasets to balance hard negatives within topics against broader cross-topic semantic structure. The project resulted in twenty trained models—including variants of RoBERTa, DistilBERT, and MPNet—alongside eight specialized evaluation datasets and an interactive demo covering sentence similarity, search, and gender bias.


### [Designing products and services based on Jobs to be Done](https://yomu.fyi/post/designing-products-and-services-based-on-jobs-to-be-done.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: Tim Lange
- Published: Oct 21, 2021
- Type: Problem & solution

The Jobs to be Done (JTBD) framework helps organizations understand consumer goals by defining a job as the progress a consumer seeks to achieve in a specific context. To prioritize new features, GrabFood applied this framework and force diagrams to investigate consumer motivations and friction points. Interviews revealed that working parents ordering family meals experienced anxiety from navigating excessive individual choices during checkout. To address this obstacle, Grab developed an algorithm within the GrabMerchant application that identified complementary food items and automatically suggested ready-made bundles. Following the release, thousands of merchants adopted one-tap bundle creation, streamlining ordering for families while supporting restaurant growth.


### [How Safari Ruined My Tuesday](https://yomu.fyi/post/how-safari-ruined-my-tuesday.md)
- Company: [Fly.io](https://yomu.fyi/company/fly-io.md)
- Published: Oct 21, 2021
- Type: Incident / postmortem

Safari 15 introduced an obscure CSS bug that made pointer-events: none persist in LiveView applications. LiveView uses connected and disconnected modes, and its default app.css had included .phx-disconnected \* { pointer-events: none; } to disable interaction while disconnected. In Safari 15, the rule could remain effective after reconnection or class changes, leaving links and buttons unresponsive even though the web inspector reported pointer-events: auto. The workaround removes pointer-events: none from app.css, although LiveView already ignores clicks on Phoenix links and buttons while disconnected; LiveView 0.17 instead uses a different disconnected-state class. The change avoids applying the faulty rule from existing user-land CSS while retaining custom loading-state styling, and the post says a Safari Development Preview build might fix the underlying bug.


### [Debugging Systems in the Cloud: MySQL, Kubernetes, and Cgroups](https://yomu.fyi/post/debugging-systems-in-the-cloud-mysql-kubernetes-and-cgroups.md)
- Company: [Shopify](https://yomu.fyi/company/shopify.md)
- Author: 2023-10-18
- Published: Oct 21, 2021
- Type: Incident / postmortem

KateSQL, Shopify’s Database-as-a-Service platform on GKE, investigated why otherwise comparable MySQL Pods took between 10 and 30 minutes to be created during instance replacement. Measurements found extreme differences for Pods with identical resources and datasets: one required 2,120 seconds for initialization and 1,104 seconds for startup, while another took 74 and 17 seconds, respectively; affected Pods also accumulated more slow queries involving temporary memory tables. Profiling showed InnoDB’s single-threaded buffer pool initialization spending substantial time in mmap calls and the Linux memory cgroup function mem\_cgroup\_commit\_charge, while affected nodes had roughly 50,000 memory cgroups. The team linked the slowdown to a Linux kernel cgroup memory leak and slab reparenting behavior, then improved performance through node replacement, memory.force\_empty testing, and GKE 1.18 upgrades with relevant fixes.


[Newer posts](https://yomu.fyi/posts/page/216.md) · [Older posts](https://yomu.fyi/posts/page/218.md)