RSS System Design ArticlesSystem design decisions, architectures and failures from engineering blogs. Summaries and Classifier leverages AI; may contain errors.urn:rss-system-design:system-design-articles2026-10-28T12:00:00Zrss-triageAutomating agentic AI Well-Architected reviews with the AWS Agentic AI Lensurn:rss-system-design:abe3c925b82026-10-28T12:00:00ZAWSSmarsh built a Claude skill that reviews an agentic workload against 150 AWS Agentic AI Lens best practices, using a JSON dataset, an evidence rubric and a deterministic workbook generator.How Revinate accelerated API adoption with Amazon API Gatewayurn:rss-system-design:e9194a74772026-10-28T11:59:59ZAWSRevinate explains why it chose Amazon API Gateway and its portal over self-managed or enterprise gateways, and why it connected to EKS using TargetGroupBinding with Network Load Balancer target groups.How Uber Built A Gateway For AI Agents To Access Toolsurn:rss-system-design:ccdcae58272026-10-28T11:59:58ZByteByteGoThe article explains how Uber built an MCP Gateway with a registry, a proxy that translates to HTTP, gRPC and TChannel, AutoCrawler tool discovery, tool-level access control and context-saving features.Adopting Cloud SQL Backup Vault to meet strict backup requirements (DeNA Infra SRE) (translated from Japanese)urn:rss-system-design:536f6576612026-10-28T11:59:57ZDeNADeNA describes adopting Cloud SQL Backup Vault with a cross-project setup for immutable backups, estimating about 77% lower monthly cost than daily GCS exports and noting plan and billing limits.2026 Summer Engineering Intern Projects – Part Oneurn:rss-system-design:325dce954f2026-10-28T11:59:56ZDoorDashSummarizes three summer 2026 intern projects on trustworthy LLM systems: a data-quality check suggester, an agent that builds dashboards, and an automated root cause analysis pipeline for multi-agent failures.Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Designurn:rss-system-design:39127629a92026-10-28T11:59:55ZMetaMeta describes the next Andromeda ads retrieval system, using a mixture-of-experts gating model, a fused GPU kernel and single-GPU co-location to score over one billion ads per second per server.Behind the scenes, two months later: How we ship Google Agent Skills like productsurn:rss-system-design:813bd371b02026-10-28T11:59:54ZGoogleGoogle describes the pull request pipeline for its Agent Skills repository, with linters, semantic versioning checks, eval-suite validation, link checks, an AI review agent and human approval.How GKE networking scales to support high densities of concurrent AI agentsurn:rss-system-design:2d0629110a2026-10-28T11:59:53ZGoogleGoogle explains how moving high-cardinality pod labels to annotations and filtering labels in GKE Dataplane V2 avoids the 65,536 Cilium identity limit for high-density agent sandboxes.Hill-Climbing Agent Quality with Agents CLI Evaluationurn:rss-system-design:2a8d977a352026-10-28T11:59:52ZGoogleA walkthrough of the Agents CLI eval commands on a LangChain agent, covering baseline scoring, a prompt regression, a model swap, nightly scheduled runs, threshold alerts and prompt optimization.Agentic harnesses (epilogue): working day to day with a frontier agent (translated from French)urn:rss-system-design:e67302ad6a2026-10-28T11:59:51ZOCTO TechnologyThe epilogue of an OCTO series describes a Claude Code workflow of agent interviews, Opus planning and Sonnet implementation, plus methods to re-audit the harness and reduce token use.A JSON schema Is not an AI governance layerurn:rss-system-design:c4728b3f992026-10-28T11:59:50ZStack OverflowThe author of the PIC protocol explains how duplicate identifiers, input normalization and self-asserted trust made agent action verifiers disagree, and proposes schema, semantic and policy layers with conformance vectors.How Anchorage Digital and Turnkey verify code with AWS Nitro Enclavesurn:rss-system-design:8778145b0a2026-10-27T12:00:00ZAWSAnchorage Digital and Turnkey run the VisualSign transaction parser inside AWS Nitro Enclaves so wallets can verify through attestation and reproducible builds which code parsed a transaction.Vessel-specific voyage optimization with machine learning: Manta on AWSurn:rss-system-design:a4b986c1e22026-10-27T11:59:59ZAWSManta Marine Technologies trains a model per vessel on AWS, using Airflow to run ETL, training and voyage optimization as ECS tasks on EC2 with per-instance-type capacity providers.How Apple Fits an LLM on Your iPhoneurn:rss-system-design:2862aed6ff2026-10-27T11:59:58ZByteByteGoA summary of how Apple runs a 3-billion-parameter model on the iPhone using palettization, quantization-aware training, cross-layer KV sharing, LoRA adapters and Private Cloud Compute.From spot fixes to measured gains: improving document extraction accuracy with ai_extract and MLflow Prompt Registryurn:rss-system-design:4aed764eb72026-10-27T11:59:57ZDatabricksDatabricks describes improving ai_extract accuracy on about 40 fields by versioning prompts and schemas in MLflow Prompt Registry and scoring candidates against reviewer-approved ground truth, cutting errors by 32%.How to make postMessage type-safe in under 350 bytesurn:rss-system-design:f5a08e05ce2026-10-27T11:59:56ZEvil MartiansEvil Martians describes typed-channel, a 271-byte library that makes postMessage, event buses and Figma plugin messaging type-safe, and its 0.11 request API with AbortSignal cancellation.Beyond the Prompt: Engineering the Loop for Software Change at Scaleurn:rss-system-design:9314dfe02e2026-10-27T11:59:55ZExpedia GroupExpedia Group describes an agent loop built on a Modularization MCP server, deterministic scripts, build and test gates and persistent memory that modularized its Shared iOS Components Stack in about three months.PrivacyLib: Privacy Controls at the Data Access Boundaryurn:rss-system-design:a4839def382026-10-27T11:59:54ZMetaMeta describes PrivacyLib, a library inside the clients of hundreds of data systems that runs privacy modules such as Policy Zones on billions of checks per second, including agent tool calls.Securely connecting Gemini Enterprise to Snowflake dataurn:rss-system-design:73d9e45bab2026-10-27T11:59:53ZGoogleA guide to connecting a Gemini Enterprise agent built with Google ADK to a Snowflake Cortex Agent over MCP, running queries with each user's OAuth session and role.Building Conversational Commerce with Multi-Agent Swarms on GCP: A Working Prototypeurn:rss-system-design:dc1ded807a2026-10-27T11:59:52ZGoogleThe HappyDelivery prototype implements conversational grocery ordering with a Pub/Sub-backed webhook ingress, a Redis session store and a LangGraph agent swarm on Gemini, reporting latency figures and about 4,240 tokens per session.Fast, reliable A2UI, with swappable inference formatsurn:rss-system-design:b8729281aa2026-10-27T11:59:51ZGoogleGoogle introduces swappable inference formats for A2UI, making the Python-like A2UI Express the default after comparing four formats, and shows the Jev decision model building fixed-layout screens in the Apparite example.How Spec-Driven Development Shaped My New Way of Engineeringurn:rss-system-design:c78cb9d9322026-10-27T11:59:50ZGustoA Gusto engineer describes adopting spec-driven development with AI agents, moving from spec-kit to OpenSpec, and adding a reviewer agent that scores spec and implementation alignment before pull requests.[Intern report] Re-architecting the Settlement Service that supports merchant settlement (translated from Japanese)urn:rss-system-design:f4549b7e392026-10-27T11:59:49ZMercariAn intern describes the Settlement Service v1 to v2 re-architecture and the design of Transaction Clearing, which uses idempotency keys, ordered side effects, error detection and the Magician workflow engine for retries.3 Ways to Boost MoE Inference on NVIDIA DGX Sparkurn:rss-system-design:538db15dde2026-10-27T11:59:48ZNVIDIANVIDIA shows how Humming MoE kernels, FP8 LM-head quantization and MTP speculative decoding raised Ling-3.0-flash decode throughput by 42.5% on GSM8K on a DGX Spark.Fine-Tuning NVIDIA Nemotron to Know Local Rules, Not Just Local Languagesurn:rss-system-design:2a5c8ec1ff2026-10-27T11:59:47ZNVIDIANVIDIA shows how to fine-tune Nemotron 3.5 Lightning with LoRA on synthetic conversations from Nemotron-Personas-France to build a French public-service support agent and evaluate it.Memory-Enforced GPU Sharing in Kubernetes with NVIDIA CUDA 13.4urn:rss-system-design:3a92d8ebc02026-10-27T11:59:46ZNVIDIANVIDIA introduces the open-source gpu-fractioning operator and KAI Scheduler NvFractions mode, which use CUDA 13.4 to enforce per-workload GPU memory limits and compute modes in Kubernetes.What happened with npm packages over the last 12 months? (translated from French)urn:rss-system-design:482a0ef3c22026-10-27T11:59:45ZOCTO TechnologyA French article reviewing npm supply chain attacks from August 2025 to August 2026 and proposing release-age delays, install-script blocking, CI secret separation and OIDC publishing.How a beagle-shaped AI agent ships code to madewithvuejs.comurn:rss-system-design:f13e4469942026-10-27T11:59:44ZSentryThe madewithvuejs.com team describes Milly, an AI agent with access to GitHub, Linear, Sentry and Discord that scouts content, builds features and fixed production bugs under defined safeguards.Freight Search: Composing a Shared Search Index from Independently Owned Fragmentsurn:rss-system-design:6f583eae592026-10-27T11:59:43ZUberUber Freight explains Freight Search, which builds OpenSearch indices from fragments owned by separate data sources, ingesting over 70 topics into 300 fragments with Spark bootstrap and Flink CDC.Data governance with annotations in Amazon S3urn:rss-system-design:d301412f3a2026-10-26T12:00:00ZAWSAn AWS reference design for keeping classification, retention, residency and PII context as Amazon S3 annotations, written by a Lambda and Bedrock pipeline and queried through an S3 Metadata annotation table with Athena.Recommended resilience testing in the next generation of AWS Resilience Huburn:rss-system-design:ed009e81b12026-10-26T11:59:59ZAWSAWS introduces recommended resilience testing in Resilience Hub, four FIS-based tests including dependency validation, which uses DNS-based dependency discovery and packet loss injection to check soft and hard dependencies.A Deep Dive into Salesforce’s Agent Tech Stackurn:rss-system-design:8b1ade10302026-10-26T11:59:58ZByteByteGoSalesforce replaced a single-prompt Agentforce design with Agent Script, Data 360 identity resolution, the Agent Context Engine and Agent Identity, and exposed its platform to outside agents through AIforce and an MCP server.Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLangurn:rss-system-design:929fdafa8a2026-10-26T11:59:57ZCloudflareCloudflare explains how a prefill KV cache pool smaller than the context window and an SGLang disaggregation bug let 155 GLM 5.3 responses on Workers AI include other users' data, and the fixes that followed.Easy to integrate, compliant by design: Bringing the Swiss Franc to Google Cloud Universal Ledgerurn:rss-system-design:14a4e4e5be2026-10-26T11:59:56ZGoogleCrypto Finance Group and Google Cloud describe a pilot on Google Cloud Universal Ledger enabling instant transfers between customers of two Swiss banks, using deferred settlement and integration completed in under three months.Scale agents securely with IAM Helper Agent Skills and MCP Toolsurn:rss-system-design:bbb9119ec92026-10-26T11:59:55ZGoogleGoogle Cloud introduces IAM Helper, Agent Skills and Google-managed MCP tools that let coding agents troubleshoot 403 errors, author policies, simulate changes against 90 days of logs and request just-in-time access.Unifying enterprise context and semantics: LookML in Knowledge Catalog is now GAurn:rss-system-design:26e44b92be2026-10-26T11:59:54ZGoogleGoogle Cloud announces general availability of LookML semantic indexing in Knowledge Catalog, so Gemini Enterprise and custom agents can use indexed join paths, metric formulas and lineage when generating SQL.Making Failure Diagnosis Reliable with Pensive AIurn:rss-system-design:e411cc279f2026-10-26T11:59:53ZNetflixNetflix describes Pensive AI, which adds LLM error extraction by line ranges and on-demand cached debugging to its rule-based failure diagnosis system, reaching extraction coverage of 100% of failed workloads.Multi‑Ratio Video Resizing for Mobile Feed and Marketingurn:rss-system-design:3a95764b112026-10-26T11:59:52ZNetflixNetflix describes an automated system that crops 16:9 video into 9:16, 4:5 and 1:1 variants using shot detection, active speaker detection, saliency and temporal smoothing, integrated into its Clips tool.How NVIDIA Minerva Turns Telemetry Chaos Into Peak GPU Yieldurn:rss-system-design:b9987900d62026-10-26T11:59:51ZNVIDIANVIDIA Minerva is an AIOps platform that maps AI factory infrastructure to jobs in a knowledge graph, detects gray failures with machine learning, correlates alert storms into incidents and offers the Sherlock agent.How NVIDIA Mission Control 2.5 Brings Autonomous Recovery to Kubernetesurn:rss-system-design:8e25e92e122026-10-26T11:59:50ZNVIDIANVIDIA Mission Control 2.5 extends its autonomous recovery engine from Slurm to Kubernetes workloads under NVIDIA Run:ai, combining scheduler placement, job recovery and hardware recovery to handle hardware faults and transient errors.Agentic harnesses (2/2): eight reflexes to unlearn for a frugal harness (translated from French)urn:rss-system-design:0610bf3dad2026-10-26T11:59:49ZOCTO TechnologyA French guide to eight habits to unlearn when building agent harnesses for Claude Code, covering concise instructions, Tool Search, effort levels, total task cost, prompt caching, high-level tools, restarts and context compaction.Advancing Generative Recommendation Inference: our collaboration with NVIDIA on TensorRT LLMurn:rss-system-design:435adb508e2026-10-26T11:59:48ZSpotifySpotify describes moving generative recommendation inference from vLLM and SGLang to TensorRT LLM with NVIDIA, adding beam search, scoring and soft-token features, with up to 54% cost reduction for beam search.AI Changed How Spotify Builds. What We Learned (and Fixed) About Quality at Higher Velocityurn:rss-system-design:2b91345b562026-10-26T11:59:47ZSpotifySpotify describes quality problems during its shift to AI-assisted development, including a June 24 publishing delay and failover capacity issues, the fixes it made, and incident and pull request data on AI impact.Indexing the Data Lake for Online Point Queriesurn:rss-system-design:ae736d3c242026-10-26T11:59:46ZSpotifySpotify describes Random Access Parquet, an external index mapping keys to Parquet files and rows so online services and AI agents can run fast point queries over its GCS data lake.Portal by Spotify cut my Claude Code token usage by 90%urn:rss-system-design:44a6c9acd92026-10-26T11:59:45ZSpotifyA Spotify engineer describes the shunt Claude Code plugin and two AiKA Modes in Portal that route bulk file reads and boilerplate code to Gemini 2.5 Flash, saving around 90% of bulk-read tokens.When Can LLMs Replace Humans in A/B Tests?urn:rss-system-design:3d6c1ee8702026-10-26T11:59:44ZSpotifyA paper summary on when LLM predictions can replace human outcomes in A/B tests, defining surrogacy and comparability conditions and testing gpt-4o-mini calibration methods on the Upworthy Research Archive.Why Spotify Is Not Using Bayesian A/B Testingurn:rss-system-design:6cb336b8582026-10-26T11:59:43ZSpotifySpotify examined common claims about Bayesian A/B testing, compared Bayesian configurations with frequentist ones, and explained why its experimentation program kept only frequentist inference rather than adding a Bayesian mode.GHARP: Real-Time Gaussian Head Animation from Large-Scale Reconstruction Priorurn:rss-system-design:954e49682b2026-10-25T12:00:00ZAppleA research abstract presenting GHARP, a two-stage method that animates 3D Gaussian heads in real time from a few images, reporting state-of-the-art quality on Ava-256 and up to 13× faster runtime on an A100.Genie One MCP: Give any AI Agent the Right Business Contexturn:rss-system-design:84da03a25e2026-10-25T11:59:59ZDatabricksDatabricks introduces Genie One MCP, which exposes Genie One and the Genie Ontology context layer to MCP-compatible assistants through five tools, with on-behalf-of user authentication and Unity Catalog permission enforcement.The challenge of building an exploit-strategy AI for EDGE POKER (translated from Japanese)urn:rss-system-design:cae7d46fe52026-10-25T11:59:58ZDeNAA DeNA intern builds an exploit-strategy AI for six-player poker in EDGE POKER by adding an MLP to the frozen EDGE AI model, raising simulated profit from +27.1 to +52.9 bb/100.Background Coding Agents: Supercharging Downstream Consumer Dataset Migrations (Honk, Part 4)urn:rss-system-design:8fddb1ff582026-10-25T11:59:57ZSpotifySpotify used its Honk background coding agent with Backstage and Fleetshift to migrate consumers of two deprecated datasets, rolling out 240 automated PRs across dbt and BigQuery Runner pipelines.Better Experiments with LLM Evals — A funnel, not a forkurn:rss-system-design:c25f08e5042026-10-25T11:59:56ZSpotifySpotify argues LLM evals belong before A/B tests in an evaluation funnel, citing that about 12% of its tests ship positive results and about 42% of launched experiments are rolled back.Building a Natural Language Interface to the Spotify Ads API with Claude Code Pluginsurn:rss-system-design:7f479a29922026-10-25T11:59:55ZSpotifySpotify describes an open-source Claude Code plugin that turns natural language into Spotify Ads API v3 calls using Markdown skills, an agent, a PreToolUse hook and an approximately 8,600-line OpenAPI spec instead of MCP.Coding Is No Longer the Constraint: Scaling Developer Experience to Teams and Agents at Spotifyurn:rss-system-design:efe4d0122a2026-10-25T11:59:54ZSpotifySpotify describes its Fleet Management system Fleetshift, its Claude-based background coding agent Honk, and Backstage standardization, citing more than 2.5 million automated maintenance PRs and a 76% increase in pull request frequency.Content Ingestion & Podcast Video Incident Reporturn:rss-system-design:f33ed4bb312026-10-25T11:59:53ZSpotifySpotify reports on the June 24 video podcast publishing delay caused by transcoding capacity limits, a batch job and a scheduling bug, and the fixes including about 67% more transcoding capacity.Encoding Your Domain Expert: The Context Layer Behind Spotify's Data Assistanturn:rss-system-design:3aae9c642f2026-10-25T11:59:52ZSpotifySpotify describes the cluster-based context layer behind its Vedder data assistant, where domain experts curate datasets, question-SQL pairs and docs over 70,000+ datasets, with curators accepting only 12.5% of mined pairs.Bringing post-quantum cryptography to geographically fenced key managementurn:rss-system-design:d39f172dbe2026-10-23T12:00:00ZCloudflareCloudflare launched a post-quantum GeoKey Manager in closed beta that wraps customer decryption keys so only servers in a chosen region can unwrap them, using hybrid ML-KEM-768 and X25519.Origin Range Requests: Serving large files without range-request tuningurn:rss-system-design:3b10b170bb2026-10-23T11:59:59ZCloudflareCloudflare rebuilt range-request caching on its Pingora-based cache to align origin fetches to disk chunks, roughly halving p99 time to first byte for R2 and cutting tiered cache traffic about 6%.Built for Agents: The Case for Spark Declarative Pipelinesurn:rss-system-design:ef77d0b3742026-10-23T11:59:58ZDatabricksDatabricks benchmarked 528 agent-built data pipelines and found declarative Spark Declarative Pipelines cheaper, faster and leaner than procedural PySpark for coding agents.Beyond API Compatibility: Building a Tool to Measure the Blast Radius of Library Changesurn:rss-system-design:43a840a3952026-10-23T11:59:57ZNetflixNetflix's Java Platform team built Quasar, which indexes code from CI builds into Parquet on S3 to show which repositories use API symbols that a library change affects.Scaling NVIDIA TensorRT Performance Optimization Beyond a Single Build Route with the Global Performance Tuner, Part 1urn:rss-system-design:314759a41f2026-10-23T11:59:56ZNVIDIAPart 1 introduces the Global Performance Tuner in TensorRT 11.1, describing heuristic sprawl in default scheduling and a mathematical model of builder and search algorithms over accuracy-feasible solutions.Scaling NVIDIA TensorRT Performance Optimization Beyond a Single Build Route with the Global Performance Tuner, Part 2urn:rss-system-design:c5f800f6652026-10-23T11:59:55ZNVIDIANVIDIA describes the design of the TensorRT 11.1 Global Performance Tuner, which searches build routes of over 200 compiler knobs under an accuracy tolerance and gave 1.14x to 1.63x speedups on ten workloads.Using AI Agents to Teach Robots Like Microduck to Walk in Simulationurn:rss-system-design:3981bbd8682026-10-23T11:59:54ZNVIDIANVIDIA describes a workflow in which Hermes, Nemotron and Astra agents plan and build a Microduck obstacle course, Isaac Lab trains a navigation policy on 4,096 parallel environments, and Isaac Sim validates it.Agentic harnesses (1/2): why yesterday's good practices are expensive today (translated from French)urn:rss-system-design:1d5bab4c4a2026-10-23T11:59:53ZOCTO TechnologyThe article argues that coding-agent harness practices built for weaker models, such as long AGENTS.md files and maximum reasoning, now cost more, using token prices and prompt-cache behaviour as worked examples.My agents were reading too much: what ten weeks of token hunting taught me (translated from French)urn:rss-system-design:5e3ced5ac72026-10-23T11:59:52ZOCTO TechnologyA product coach measured the billable tokens of his AI agents and applied seven measures, such as a 150-line CLAUDE.md cap, after which median tokens per launched agent fell 55%.Rapidly scaling online storage to serve over 1 billion ChatGPT users (Part II)urn:rss-system-design:298c8de07d2026-10-23T11:59:51ZOpenAIOpenAI describes scaling its Habitat storage platform on Azure Cosmos DB through tenancy changes, hot-key throttling, CDC-invalidated Valkey caching and a Rust rewrite, cutting read p50 to roughly 10 ms.How Retention Policies Silently Break Continuous Aggregatesurn:rss-system-design:0ddec0b5cd2026-10-23T11:59:50ZTimescaleA TimescaleDB retention job that dropped raw chunks inside a continuous aggregate's refresh window made the next refresh silently delete buckets; the fix is one retention-versus-refresh invariant plus an audit query.Localizing compliance at scale: achieving South Korea’s ISMS certificationurn:rss-system-design:e0bd59849e2026-10-22T12:00:00ZCloudflareCloudflare obtained South Korea's ISMS certification and, to enforce itemized consent, removed social sign-up in South Korea rather than build a separate consent gate after authentication.The AI Bill is comingurn:rss-system-design:41b6b147542026-10-22T11:59:59ZDatabricksDatabricks co-founder Patrick Wendell explains how AI coding agent costs spiraled at Databricks and how a central gateway, model routing and spend visibility cut unit costs by up to 90%.The Fraud Analyst AI Coworker: Detecting Telecom Fraud in Real-Time on One Platformurn:rss-system-design:c110e13f202026-10-22T11:59:58ZDatabricksDatabricks argues that telecom fraud signals are split across six or seven systems synced by batch jobs, and outlines a unified real-time architecture and AI analyst assistant.Dark mode was never on our roadmap. Here’s how it got built.urn:rss-system-design:dd3410d13d2026-10-22T11:59:57ZMediumOne Medium web engineer built dark mode using CSS light-dark() tokens instead of caching separate light and dark copies, with a Statsig flag splitting the Cloudflare Workers edge cache during rollout.Single-tenant Titus: VM-level isolation on a multi-tenant container platformurn:rss-system-design:8e4ffc76622026-10-22T11:59:56ZNetflixNetflix added an allowlisted single-tenant mode to its Titus container platform that gives a task its own right-sized EC2 instance, after weighing shared-kernel tuning, microVMs, gVisor and plain EC2.How to Read Data into GPU Memory in Python with NVIDIA GPUDirect Storageurn:rss-system-design:af796ac2c62026-10-22T11:59:55ZNVIDIAA tutorial on reading data into GPU memory with the NVIDIA KvikIO Python API over GPUDirect Storage and cuFile, verifying the nvidia-fs path with gds_stats and explaining how a 1 GiB read is split.Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodesurn:rss-system-design:085abd1e412026-10-22T11:59:54ZNVIDIANVIDIA announces eight-node DGX Spark clusters with 1 TB unified memory and benchmarks inference, the RoCEnante collective backend, ring versus switch topologies, Switchyard model routing and nanochat pre-training scaling.We gave AI agents a mobile dev environmenturn:rss-system-design:ce29ad327e2026-10-22T11:59:53ZShopifyShopify built Remote Dev, a pool of warm remote Macs that AI agents claim over SSH to build and test its mobile apps, with Android emulators on separate Linux VMs.We built three memory systems. The most expressive one wasn’t the most useful.urn:rss-system-design:b4f2c6ca052026-10-22T11:59:52ZTimescaleTiger Data built three memory systems for its Slack agent Eon, a fact list, a PostgreSQL knowledge graph and an ltree tree with hybrid search, and kept the tree.Automate medical coding with Step Functions and Amazon Bedrock AgentCoreurn:rss-system-design:b8951545d92026-10-21T12:00:00ZAWSAWS shows a medical coding pipeline using Step Functions, Lambda, an Amazon Bedrock AgentCore harness and Aurora pgvector, which autocoded confident terms and sent cases below a 0.90 confidence gate to human coders.Cloudflare deploys protections against php2shell vulnerabilityurn:rss-system-design:da92d86b612026-10-21T11:59:59ZCloudflareCloudflare deployed WAF protections against the php2shell PHP vulnerabilities about 9 hours before public disclosure, using full transaction detection that combines request and origin-response signals.How Commerce.com and Cloudflare are getting storefronts ready for AI agentsurn:rss-system-design:c89f2704352026-10-21T11:59:58ZCloudflareCommerce.com and Cloudflare enabled WebMCP agent tools across almost 100,000 storefronts with a per-merchant opt-out built from Configuration Rules and custom hostname metadata.Introducing full transaction detection for Cloudflare WAFurn:rss-system-design:5e1de757df2026-10-21T11:59:57ZCloudflareCloudflare added full transaction detection to its WAF, evaluating each request together with the origin's response, and response rules that can block a matching response.Aligning Dispatch to Reality: Using Reinforcement Learning to Adapt Real-Time Objective Weighturn:rss-system-design:29b1a3108a2026-10-21T11:59:56ZDoorDashDescribes an offline-trained reinforcement learning policy that adjusts a dispatch objective-weight multiplier per store, tested in a global switchback experiment against the production optimizer.Achieve Productive AI Capacity Faster and More Predictably with NVIDIA DSX PODurn:rss-system-design:0acaf72d592026-10-21T11:59:55ZNVIDIANVIDIA describes the DSX POD reference design for Vera Rubin rack-scale systems, which pre-integrates and validates compute, networking, storage, 800 VDC power and liquid cooling in factory-built containers before site delivery.Design Proteins with RFD4-Proteina Using Reward-Guided Inference-Time Searchurn:rss-system-design:b43bb9d51c2026-10-21T11:59:54ZNVIDIAA tutorial on designing PD-L1 protein binders with NVIDIA RFD4-Proteina through the open model and a NIM microservice, using reward-guided inference-time search scored by OpenFold3 pLDDT.Developing an AI Factory Digital Twin with an Agent-Assisted Workflowurn:rss-system-design:b7422fdbc42026-10-21T11:59:53ZNVIDIANVIDIA describes an agent-assisted workflow that turns DSX Blueprint assets, site plans and USGS geographic data into an editable OpenUSD AI factory campus model and proposes a fixed-power capacity study.How NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scaleurn:rss-system-design:a304b6a0502026-10-21T11:59:52ZNVIDIANVIDIA describes the OpenShell reference architecture, with gateway, supervisor, sandbox and policy prover components, running on Vera CPUs with optional NVIDIA Sentry on BlueField-4 DPUs for hardware-enforced agent governance.Inside NVIDIA Sentry: Secure AI Agents at Runtimeurn:rss-system-design:15dc543e012026-10-21T11:59:51ZNVIDIANVIDIA describes Sentry, a BlueField-4 based framework that uses DOCA Flow, the Envoy-based DOCA Tower proxy with Open Policy Agent, and DOCA Argus memory introspection to monitor and contain AI agents.How we built ChatGPT’s novel ad stack for user trust and relevanceurn:rss-system-design:edef63b51a2026-10-21T11:59:50ZOpenAIOpenAI describes the ChatGPT ads stack: stateful ads sessions, a post-trained LLM orchestrator calling retrieval, filtering, scoring and auction tools, and an ads service separate from answer generation.Designing IIoT Ingest That Survives a Cut Uplinkurn:rss-system-design:2184dc86d62026-10-21T11:59:49ZTimescaleA guide to an IIoT ingest path in PostgreSQL that replays buffered rows exactly once after an outage, using a unique constraint, a SQLite buffer and a rate-capped COPY replay.8 eMudhra certificates are compromised. Here's what you need to knowurn:rss-system-design:f1bf501de52026-10-20T12:00:00ZCloudflareCloudflare describes its response after private keys of eight eMudhra intermediate certificates were compromised, including trust-store updates and CDN and Zero Trust mitigations deployed on October 20.Introducing tsrs: A port of TypeScript to Rusturn:rss-system-design:a3f6bdb00a2026-10-20T11:59:59ZCloudflareCloudflare, VoidZero and Owner.com are building tsrs, a faithful Rust port of the TypeScript compiler that is up to 8x faster and uses about 70% less memory than TypeScript 7.0.How DoorDash Built a Next-Generation Real-Time Dispatch Platformurn:rss-system-design:6d5fd506332026-10-20T11:59:58ZDoorDashDescribes how DoorDash replaced a monolithic dispatch engine with a coordinator and distributed compute tier, reporting ten-fold throughput and more than 99.96% of runs finishing in under 20 seconds.The Agents on Rails Files: designing black-box evals for coding agentsurn:rss-system-design:78192128a72026-10-20T11:59:57ZEvil MartiansEvil Martians replaced hidden Rails tests with browser-based black-box checks for stage 3 of their Agents on Rails coding-agent benchmark and calibrated the grader with mutants.The Quiet Tax on Machine Learning Teamsurn:rss-system-design:fe8770f1fd2026-10-20T11:59:56ZExpedia GroupExpedia Group built a CLI that scaffolds ML projects per business domain and copies shared infrastructure code into each project, to stop repeating CI/CD, permission and cluster setup for every model.How 3 Java Champions are using GitHub Copilot for increased productivityurn:rss-system-design:fbec33bd282026-10-20T11:59:55ZGitHubThree Java Champions describe using GitHub Copilot, including migrating about 900 slides from Reveal.js to Marp and diagnosing a Pi4J virtual-thread pinning bug on a Raspberry Pi 4.App stability at scale with a layered defenseurn:rss-system-design:6d1bafd6332026-10-20T11:59:54ZGrabGrab describes layered release safety for its Android app: custom lint rules, tests, an early-access gate, staged rollouts that auto-pause on crashes, feature-flag kill switches and alerting.Local LLM on iOS 27: what Apple doesn't say in the docs (translated from French)urn:rss-system-design:8b567325552026-10-20T11:59:53ZOCTO TechnologyThe author embedded Qwen2.5-1.5B-Instruct in an iOS 27 app via Core AI, measured launch time and schema compliance, and found Apple's system model faster on compatible iPhones.Cross-store product clustering across millions of storesurn:rss-system-design:2a8881dbb82026-10-20T11:59:52ZShopifyShopify groups the same product across millions of stores under one global ID using embedding retrieval, a three-stage scorer cascade trained on LLM-judge labels and graph-based clustering.How Shopify turns open models into task expertsurn:rss-system-design:497306ca802026-10-20T11:59:51ZShopifyShopify post-trains open models such as Nemotron 3.5 Lightning as task-specific subagents for Sidekick, using an LLM judge and nightly self-healing data, reporting faster responses and about 90% lower serving cost.How We Cut the Coupling Tax in a Post-Booking APIurn:rss-system-design:bd26f067d12026-10-19T12:00:00ZAgodaAgoda describes the Platform framework, a modular monolith with strict layers enforced by ArchUnit in CI for its post-booking service BAPI, and reports faster code review and fewer AI prompts per changed line.Beyond the patch: Protecting your NetScaler gateway against future vulnerabilitiesurn:rss-system-design:e356517d092026-10-19T11:59:59ZCloudflareCloudflare explains how to protect Citrix NetScaler and other public login gateways after actively exploited vulnerabilities, using access policies, WAF inspection and restricted origin connections.Cutting our inference bill by 68% by being obsessive about optimizationurn:rss-system-design:bddfe014682026-10-19T11:59:58ZCloudflareCloudflare cut its AI code reviewer's average cost per review from $2.43 to $0.77 with a shared prompt-cache prefix, fewer wasted LLM calls and model and effort experiments.Introducing Clairic, a new System One class of extraction modelsurn:rss-system-design:52d55b205d2026-10-19T11:59:57ZCloudflareCloudflare previewed Clairic, an extraction model on Workers AI that returns typed values, estimates and forecasts with confidence scores in a single pass instead of generating tokens.GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabularyurn:rss-system-design:219bbb66f62026-10-19T11:59:56ZMetaMeta built GenRecs Tokenization, which quantizes multimodal item embeddings into shared codebook tokens for ads recommendation models, reporting a 0.35% conversion lift with GEM and the Adaptive Ranking Model.Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Superclusterurn:rss-system-design:bdf2dd774c2026-10-19T11:59:55ZMetaMeta describes Prometheus, its Ohio GPU supercluster on NVL72 racks, and the in-rack spares, per-replica fault tolerance and asynchronous checkpointing that cut training interruptions by 99%.The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Dourn:rss-system-design:c854e99d012026-10-19T11:59:54ZMetaMeta replaced a single all-knowing agent for data column onboarding with a 1,300-line deterministic orchestrator that routes workflows to 14 specialist subagents, cutting preparation from days to under an hour.Accelerating Mamba-2 with High-Performance Forward Pass Kernels in NVIDIA CUDA Tileurn:rss-system-design:e2b31666c72026-10-19T11:59:53ZNVIDIAJane Street and NVIDIA implemented the Mamba-2 forward pass as three cuTile kernels, explaining the chunked log-space math and optimizations, and report 1.14x to 1.82x speedups over a Triton implementation on a B300.Introducing ModuleLens: Zooming into PyTorch Modules for Human and Agentic Engineeringurn:rss-system-design:156f20b2882026-10-19T11:59:52ZPinterestPinterest built ModuleLens, which captures inputs, weights and init arguments of PyTorch modules during training so any module can be reconstructed and debugged locally without multi-GPU setup.Building Overexposed: understanding application exposureurn:rss-system-design:a8bfe677072026-10-19T11:59:51ZShopifyShopify built Overexposed, a scanner that lists GKE Ingresses and Cloud Run services from Cloud Asset Inventory and probes each anonymously to find unintended public exposure.Spin-Up, Test, Tear-Down: How trivago Runs 600+ Preview Environments on Kubernetesurn:rss-system-design:55c045d7ab2026-10-19T11:59:50Ztrivagotrivago runs 600+ per-PR preview environments on Kubernetes by deploying only the changed service with ArgoCD and scaling previews to zero with Kedify's KEDA HTTP add-on.A pragmatic guide to AI-assisted developmenturn:rss-system-design:f1c56c4fdb2026-10-18T12:00:00ZthoughtbotThe author used Claude Code with Opus 5.5 to trace and fix a bug in an internal scheduling tool, with separate sessions for implementation, review and verification, ending in a one-line change.Continuous accountability for Verified botsurn:rss-system-design:34e98f18862026-10-16T12:00:00ZCloudflareCloudflare built a system that scores whether Verified bots keep following robots.txt and slow down after HTTP 429 responses, turning repeated behaviour into an explainable trust band.Ontology-Driven Observability: Building an End-to-End Knowledge Graph at Netflix Scaleurn:rss-system-design:82a33711582026-10-16T11:59:59ZNetflixNetflix describes building an ontology-driven knowledge graph that links metrics, alerts, incidents, deployments and on-call data so incident investigations stop taking hours to reconstruct context.Agent rooms: the only place where you are happy to be contradicted (translated from French)urn:rss-system-design:c9dd80811d2026-10-16T11:59:58ZOCTO TechnologyA product coach used twelve agent rooms in which AI agents with written roles gave positions and debated before he decided, and reports what the rooms caught and where they failed.Towards a model platform (translated from French)urn:rss-system-design:48d0baca732026-10-16T11:59:57ZOCTO TechnologyOCTO describes a model platform for versioning, deploying, serving, monitoring and governing models, motivated by EU AI Act obligations, with the architecture options it compared.Open Source Profiler on an Internet Scale Fleet: Pinterest and Intel Partnership on gProfilerurn:rss-system-design:0a44faa1372026-10-16T11:59:56ZPinterestPinterest and Intel describe adopting and extending the open-source gProfiler for continuous profiling across hundreds of thousands of instances, including dynamic profiling and hardware counter flame graphs.Advancing Generative Recommendation Inference: our collaboration with NVIDIA on TensorRT LLMurn:rss-system-design:e42f5608ab2026-10-16T11:59:55ZSpotifySpotify and NVIDIA added beam search, embedding inputs and multi-item scoring to TensorRT LLM for Spotify's LLM podcast recommendations, reporting up to 54% lower beam search cost.How Dynata scaled AI across the development lifecycle with AI-DLC and Kiro IDEurn:rss-system-design:8c1fd1b8fd2026-10-15T12:00:00ZAWSDynata adopted the AI-DLC methodology with Kiro IDE through a pilot and three-day workshops, estimating about a 70 percent reduction in development phase time for full-stack features.A fully programmable secure web gatewayurn:rss-system-design:ac1f797b552026-10-15T11:59:59ZCloudflareCloudflare lets customers run Worker code inline in Cloudflare Gateway and DLP policy evaluation through a new Forward action, Custom Functions and DLP custom validation functions.Cloudflare and OpenAI: Making AI discovery better for everyoneurn:rss-system-design:58433230c32026-10-15T11:59:58ZCloudflareCloudflare and OpenAI ran a research pilot in which signals from Cloudflare's network helped OpenAI's search crawler discover and index new web pages for ChatGPT.Deploy with Voidurn:rss-system-design:10217d62a92026-10-15T11:59:57ZCloudflareCloudflare and VoidZero open-sourced Void, a Vite-based platform, SDK and CLI that let companies host internal and agent-built apps on Cloudflare without giving every employee cloud access.How our AI Research team uses AI Runtimeurn:rss-system-design:eb6210d05e2026-10-15T11:59:56ZDatabricksDatabricks' AI Research team explains how it trains models up to nearly a trillion parameters on AI Runtime, including GPU node health checks, container reuse and experiment tracking.2026 World Cup: Breaking Records, Spiky Charts and Reliability at Scaleurn:rss-system-design:40f3aca09c2026-10-15T11:59:55ZMetaMeta describes how it prepared its infrastructure for record World Cup 2026 traffic with capacity buffers, graceful degradation, per-region forecasting and continued deployments.How Meta Built Forum With AI-Native Engineeringurn:rss-system-design:d57855fe702026-10-15T11:59:54ZMetaMeta describes how a small team built its Forum app with AI agents running the plan-build-test loop under engineer supervision, gated by automated verification.Designing A Secure Voice AI System for Regulated Industriesurn:rss-system-design:0be5c974462026-10-15T11:59:53ZHaptikHaptik lists ten design principles for secure voice AI in banking, insurance and healthcare, covering data flow, authentication, least privilege, tenant isolation, RAG, tools and failure handling.NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoningurn:rss-system-design:f4d6574a762026-10-15T11:59:52ZNVIDIANVIDIA announces Nemotron 3.5 Super VL, a 120B-parameter MoE multimodal model, and describes its distillation, NVFP4 quantization, speculative decoding, prefix caching and Dynamo serving optimizations.Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72urn:rss-system-design:d6c9c7105f2026-10-15T11:59:51ZNVIDIANVIDIA describes Meta's Muse Spark Open, an 816B mixture-of-experts coding model with 42B active parameters and 1M token context, and its serving on GB300 NVL72 with NVFP4, Dynamo and NeMo post-training.Systemic Digital Twin: collecting field activities (translated from French)urn:rss-system-design:55ce7c5c652026-10-15T11:59:50ZOCTO TechnologyAn experience report on building the data model of an industrial digital twin from activities filmed and timed on the shop floor instead of from design-office specifications."I'm Sorry, Have We Met?" — Frictionless Memory for AIurn:rss-system-design:2bf9c7d1422026-10-15T11:59:49ZtrivagoTrivago compares four approaches to AI chatbot memory, from resending the whole conversation to RAG over memories and a three-stage SimpleMem pipeline, using a concierge demo.How UTS established a modern curriculum data services foundation on AWSurn:rss-system-design:54e8cb63d92026-10-14T12:00:00ZAWSUniversity of Technology Sydney built serverless curriculum data services on AWS using MSK Serverless, Neptune and OpenSearch, reporting zero production incidents since November 2024 and 1.6 second average change propagation.Multi-region disaster recovery for Amazon API Gateway WebSocket APIsurn:rss-system-design:af01e7dd6d2026-10-14T11:59:59ZAWSAWS describes an active-passive multi-Region disaster recovery pattern for API Gateway WebSocket APIs using Route 53 failover, DynamoDB Global Tables, a canary health probe and client reconnection with backoff and jitter.Inside Hyperdrive: How we solved a rare Hyperdrive MySQL caching bugurn:rss-system-design:d5ae0799382026-10-14T11:59:58ZCloudflareCloudflare describes how it fixed encoding bugs and a rare cache-format bug in Hyperdrive's MySQL support on the way from beta to general availability.Yandex outage: Data shows that instead of going dark, malicious bots went on the moveurn:rss-system-design:ee3005fe322026-10-14T11:59:57ZCloudflareCloudflare analyzed bot traffic from Yandex's networks during an October outage and found that malicious automation moved to other hosting providers instead of stopping.Reducing Qwen training time by 6x on AI Runtimeurn:rss-system-design:1c3ae698f32026-10-14T11:59:56ZDatabricksDatabricks cut fine-tuning time for Qwen 3.8 27B from 7 hours 44 minutes to 1 hour 17 minutes using fused kernels and changes to batching, tokenization and padding.Behind the Scenes: Evolving Netflix’s Ads Event Pipeline for Live — Part IIurn:rss-system-design:c628cb242f2026-10-14T11:59:55ZNetflixNetflix replaced a key-value metadata lookup in its Ads Event Publisher with a real-time Apache Flink streaming join to support ads on live streams, migrating without downtime.DataJunction as Netflix’s answer to the missing piece of the modern data stackurn:rss-system-design:cafc6aac0f2026-10-14T11:59:54ZNetflixNetflix explains why it chose the open-source DataJunction semantic layer for experimentation metrics and how a small team scoped its architecture.Integrating Netflix’s Foundation Model into Personalization applicationsurn:rss-system-design:6c2c424ddc2026-10-14T11:59:53ZNetflixNetflix describes three ways it integrated its personalization foundation model into production applications, through embeddings, a subgraph and fine-tuning, and the trade-offs of each.Measuring the Impact of Personalized Recommendationsurn:rss-system-design:9a9192d40f2026-10-14T11:59:52ZNetflixNetflix researchers built and validated a behavioral choice model to measure the causal effect of personalized recommendations on viewing, comparing the current recommender with simpler baselines.Netflix Conductor: The Next Chapterurn:rss-system-design:f58f3f82b92026-10-14T11:59:51ZNetflixNetflix rebuilt its Conductor workflow orchestration engine as usage grew, covering the data plane, evaluation engine, task polling, execution control and authoring SDK.Netflix Stratum Media Processing: Automated Container Right-Sizingurn:rss-system-design:9735a418922026-10-14T11:59:50ZNetflixNetflix built the Stratum Resource Tuner to right-size container resource limits automatically on its serverless media processing platform, with verification experiments and rollback safeguards added after an incident.Reliable GPU compute for large training & inference jobsurn:rss-system-design:e4f4b681412026-10-14T11:59:49ZNetflixNetflix detects failing GPUs and network fabric, remediates machines automatically and lets long training jobs survive the loss of a machine on its Kubernetes-native compute platform.Running Apache Spark experiments in my sleep (and on a plane)urn:rss-system-design:c6c59089922026-10-14T11:59:48ZNetflixA Netflix engineer ran parallel full-scale Spark experiments to find and fix out-of-memory failures in the member-sessionizer pipeline, and describes the scaling limit that remained.The Lifecycle of LLM-as-a-Judge: Building, Aligning, and Monitoring at scaleurn:rss-system-design:44f624f4242026-10-14T11:59:47ZNetflixNetflix describes the lifecycle of an LLM judge for recommendation explanations, from human-labeled benchmarks and reasoning-aligned rubric tuning to deployment and weekly human monitoring.TimeSeries Tiered Storage Journey: Kafka/Flink Streams to Native Cassandra Cold Readsurn:rss-system-design:1aa653cec92026-10-14T11:59:46ZNetflixNetflix moved older TimeSeries data from hot Cassandra storage to S3, first through a Kafka and Flink streaming cold tier and then by reading Cassandra backups in S3 directly.Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gymurn:rss-system-design:df0826a0d82026-10-14T11:59:45ZNVIDIASiemens used NVIDIA NeMo Gym and a composite verifier to evaluate its Solido Characterization Suite agent, raising the benchmark score from about 55% to 95% while cutting turns and output tokens.Uber’s AI infra, from the insideurn:rss-system-design:bc6f7d6d132026-10-14T11:59:44ZThe Pragmatic EngineerNotes from a podcast with Uber engineer Uday Kiran Medisetty on Uber's developer tools and AI infrastructure, including its model gateway, spend controls, code review and CI changes.CASTLE: training natural language search before the first real queryurn:rss-system-design:23358fac722026-10-13T12:00:00ZAirbnbAirbnb built CASTLE, which generates synthetic queries and relevance labels from de-identified booking sessions to train natural language search before any real queries exist.Bulk Executor for Amazon DynamoDB, Part 2: Lessons we learned the hard wayurn:rss-system-design:8b501063e92026-10-13T11:59:59ZAWSThe authors describe building the open-source Bulk Executor for DynamoDB on AWS Glue, covering cost estimation, job termination, Live Tail output limits, error handling, rate limiting and Glue upgrades.Stopping business email compromise with AI-driven email security: How VIPRE protects modern enterprises on AWSurn:rss-system-design:a65842349d2026-10-13T11:59:58ZAWSVIPRE describes its Integrated Email Security architecture on AWS using SQS, S3, ECS, GPU-hosted BERT models, MSK, RDS and OpenSearch, which scanned 517,726 messages in a sample 24 hours.Rewinding the exponential: How we keep hardware usage downurn:rss-system-design:6f6e1a44862026-10-13T11:59:57ZCloudflareCloudflare describes five data pipeline changes, from splitting a monolithic schema to zstd compression, replication, sampling and aggregation, that deferred about two and a half years of hardware growth.Under the hood, off the hot path: How D1 scales to millions of databasesurn:rss-system-design:201b2521fb2026-10-13T11:59:56ZCloudflareCloudflare rebuilt the control plane of D1, its serverless SQLite database, to support millions of databases while keeping query requests away from central single-region components.Don’t get burned, make your Stripe MRR boring and correcturn:rss-system-design:90e2b859582026-10-13T11:59:55ZEvil MartiansA guide to calculating monthly recurring revenue from Stripe subscription state and choosing subscription-change strategies so that each MRR movement is classified correctly.Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AIurn:rss-system-design:2064494a552026-10-13T11:59:54ZMetaMeta moved its GEM ads recommendation foundation model from per-surface optimization to one central model, with a new architecture, knowledge transfer, cross-domain training and more efficient GPU training.GitHub’s model for secret detection: Finding more secrets with less noiseurn:rss-system-design:e5e7270c9b2026-10-13T11:59:53ZGitHubGitHub replaced an LLM-based generic secret detection pipeline with regex candidate generation plus a fine-tuned ModernBERT classifier, and describes its training, evaluation and rollout.How to Rewrite Apache Spark UDFs for GPU Acceleration with AI Agentsurn:rss-system-design:069074fd4b2026-10-13T11:59:52ZNVIDIANVIDIA describes how Spark UDFs force CPU fallback in the cuDF plugin, three ways to rewrite them, and agent skills used to convert 30+ production UDFs, one with a 22x speedup.NVIDIA Run:ai Is Now Open Source as Part of NVIDIA DSX AI Factory Platformurn:rss-system-design:028efe51602026-10-13T11:59:51ZNVIDIANVIDIA open-sources the Run:ai platform with four components, KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map, plus related projects Model Streamer and Grove, and lists partner integrations.Building resilient AI-powered software, and planning for when things go wrongurn:rss-system-design:837e840d6f2026-10-13T11:59:50ZThe Pragmatic EngineerExcerpts from Sam Newman's book Building Resilient Distributed Systems on disaster recovery objectives and patterns, infrastructure availability, and the resilience of AI-powered software, including multi-model and multi-vendor approaches.Announcing pg_vectorsearch: Scalable Open-Source Vector Search in PostgreSQLurn:rss-system-design:20f71877682026-10-13T11:59:49ZTimescaleTiger Data introduces pg_vectorsearch, an open-source PostgreSQL extension whose partition-based PRISM index replaces graph-based vector indexing, with benchmarks against pgvector's HNSW.Agentic antimicrobial resistance surveillance with AWS HealthOmics and Amazon Bedrock AgentCoreurn:rss-system-design:078860a3922026-10-12T12:00:00ZAWSAWS presents a CDK reference architecture that ingests public SRA sequencing runs, detects resistance genes with HealthOmics, stores results in Iceberg tables and answers questions through a Strands agent on AgentCore.Filter, reshape, and redact logs before export with Transformersurn:rss-system-design:5fd5ac5f152026-10-12T11:59:59ZCloudflareCloudflare embedded Apache DataFusion in its Go-based Logpush service so customers can filter, reshape and redact logs before export, after rejecting a Workers-based prototype.OCP Summit 2026: Powering AI with Open Innovations in Scale-Up, Scale-Out, and Scale-Across Networksurn:rss-system-design:4b416d68142026-10-12T11:59:58ZMetaMeta describes changes to its AI network architecture, including a multiplane two-tier fabric, a cross-region aggregation layer with distance-aware buffering and workloads spanning its data centers and public cloud.Shipping Media Over QUIC at Meta Scale for Instagram Live Ingesturn:rss-system-design:05317d29c22026-10-12T11:59:57ZMetaMeta replaced its in-house RUSH protocol with Media over QUIC for Instagram Live ingest, using an edge translation layer and server-side rollout routing until reaching parity.Turning Opportunistic AI Inference Capacity Into Better Ads Recommendationsurn:rss-system-design:2d2a49a48e2026-10-12T11:59:56ZMetaMeta built a system that runs optional ads ranking work on preemptible idle inference capacity, with elasticity levers, isolation of baseline capacity, feedback loops and value-aware load shedding.GrabMart data for faster disease outbreak detectionurn:rss-system-design:adfb9360e32026-10-12T11:59:55ZGrabGrab describes an early-stage, undeployed concept for detecting disease outbreaks from aggregated GrabMart over-the-counter product demand, with privacy and governance controls.Flex Model: an event-driven feature engineering systemurn:rss-system-design:c43be8ebb52026-10-12T11:59:54ZJane StreetJane Street replaced a monolithic feature system with Flex Model, an event-driven engine that turns market events into feature vectors using a fixed per-event plan built at startup.ReplaySSM: Rethinking State Storage for Faster Hybrid-Model Decodingurn:rss-system-design:03ba9e80df2026-10-12T11:59:53ZNVIDIAReplaySSM, from NVIDIA and the Dao AI Lab, stores SSM state as a checkpoint plus a buffer of recent inputs, cutting state traffic, simplifying speculative rollback and limiting quantization error in hybrid-model decoding.I don't code, BMAD codes for me with its agents: take it on trust or on proof? (translated from French)urn:rss-system-design:1012b63fca2026-10-12T11:59:52ZOCTO TechnologyA non-coding product owner describes validating eight agent-built projects through readable evidence and a four-question product quality evaluation that left half its cells unmeasured.Deno is joining Cloudflareurn:rss-system-design:1aab300ff02026-10-09T12:00:00ZCloudflareCloudflare announces that the Deno team is joining it to make self-hosting the open-source Workers runtime workerd practical, and explains its reasoning on platform lock-in.Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flashurn:rss-system-design:f0f76cf0382026-10-09T11:59:59ZCloudflareCloudflare released Clef-omni, a decision model that takes audio, video, image and text in one call, and made Clef faster and Clef-flash cheaper.Introducing on-demand CPU and memory profiling with flamegraphs for Workers and Durable Objectsurn:rss-system-design:c966b1370f2026-10-09T11:59:58ZCloudflareCloudflare added on-demand CPU and memory profiling for Workers and Durable Objects and explains how profiling requests reach production isolates and two fixes it led to internally.From Activity to Intent: Generating User Journeys with LLMsurn:rss-system-design:4d7e7fa4e02026-10-09T11:59:57ZPinterestPinterest replaced a multi-stage keyword-clustering pipeline with a single fine-tuned 4B LLM to infer user journeys, covering distillation trade-offs, serving architecture, safety controls and online results.Bridging technical depth and usability: The story behind Radar’s redesignurn:rss-system-design:21a756e1682026-10-08T12:00:00ZCloudflareCloudflare redesigned the Radar landing page for non-expert audiences, replacing a bento layout with a tabbed flow and moving to shared Kumo components.One Signal, Many Uses: Identity Intelligence for DoorDash Integrityurn:rss-system-design:12459b33fa2026-10-08T11:59:59ZDoorDashDescribes DoorDash's identity intelligence platform, which layers an entity-resolution graph, relationship signals and decisioning models on one shared signal store to serve support, promotions, account security and bot mitigation.Explainable AI: If You Can’t Evaluate It, Can You Trust It?urn:rss-system-design:f4079ace9c2026-10-08T11:59:58ZFeedzaiFeedzai compared eight SHAP formulations with offline metrics and a study of 37 analysts, finding that explanations changed analyst confidence without improving decision accuracy.Building Reliable Data Analytics Agents: Lessons from the KDD Cupurn:rss-system-design:333c65dc3b2026-10-08T11:59:57ZNVIDIADescribes the harness behind the NVIDIA KGMON team's second place in the KDD Cup 2026 Data Agents competition, including a unified SQL surface, small tool set, trace inspection and repeated attempts.Building an evidence-grounded agentic security operations harness on Cloudflareurn:rss-system-design:d59fe0acf62026-10-07T12:00:00ZCloudflareCloudflare rebuilt an agentic security operations harness after a single-agent prototype failed, moving recon and scope enforcement into deterministic code with specialist agents and citation validation.Secret protection must scale with softwareurn:rss-system-design:3902c431972026-10-07T11:59:59ZGitHubGitHub's secret protection lead shares nine quarters of push data and introduces a fine-tuned classifier, built with Microsoft, that extends push protection to unstructured secrets.Faster Scientific Image Analysis with NVIDIA cuPhotonurn:rss-system-design:bf70c89ad02026-10-07T11:59:58ZNVIDIANVIDIA introduces cuPhoton, an open-source CUDA-X toolkit that keeps astronomical image data on the GPU from FITS loading through alignment, PSF-matched subtraction, dipole fitting and candidate review, scaling across MPI workers.Scaling Decision Optimization to 100 Million Variables and Beyond with mPDLP in NVIDIA cuOpturn:rss-system-design:446f80fce92026-10-07T11:59:57ZNVIDIANVIDIA introduces mPDLP in cuOpt, which distributes the PDLP linear programming solver across NVLink-connected GPUs with min-cut partitioning, benchmarked on more than 100 LP instances against single-GPU PDLP and D-PDLP.The Machines that Make the Machinesurn:rss-system-design:733e0d20922026-10-07T11:59:56ZNVIDIANVIDIA Seattle Robotics Lab describes teaching robots busbar assembly and multi-connector insertion for GB300 tester trays, using classical pipelines, the DOPER pose estimator, sim-to-real RL and the TALOS orchestration system.Validate AI Factory Changes with Digital Twins and AI Agentsurn:rss-system-design:7f53a9edce2026-10-07T11:59:55ZNVIDIANVIDIA describes using DSX Air node-based digital twins with AI agents and NVIDIA Brev GPUs to validate AI factory configuration and software changes in CI/CD, illustrated by a VSS video intelligence workflow.Building resilient systems with Sam Newmanurn:rss-system-design:c2e92180362026-10-07T11:59:54ZThe Pragmatic EngineerNotes from a podcast with Sam Newman on microservices, three rules of distributed systems, idempotency, resilience concepts and how AI changes software development and architecture.Workbench: Building an AI Content Generation and Quality Assurance System at Scaleurn:rss-system-design:488c7c38a12026-10-07T11:59:53ZThumbtackThumbtack built Workbench to generate and automatically review 800,000 landing-page descriptions with LLMs, choosing between deterministic validators and LLM judges and between serial and batch pipelines.Continuous Aggregate Refresh Policies for Solar and Wind SCADA Data: Choosing the Right Policy Shapeurn:rss-system-design:54eb9bf0222026-10-07T11:59:52ZTimescaleTimescale's guide explains how to read continuous aggregate refresh policies and choose among five policy shapes for solar and wind SCADA data, including retention and late data.Beyond synthetic testing: Capturing and replaying real database workloads at Airbnburn:rss-system-design:b15d69d9be2026-10-06T12:00:00ZAirbnbAirbnb built a ProxySQL-based system that captures and replays real MySQL workloads, used for its MySQL 5.7 to 8.0 upgrade and capacity testing.The keys to the Internet change on October 11. Are you ready?urn:rss-system-design:c34e297b4a2026-10-06T11:59:59ZCloudflareCloudflare explains the October 11, 2026 DNS root key-signing key rollover, how resolvers learn the new key, and the readiness test it built using RFC 8509.Scaling and Operating a Large dbt Project on Databricks: IFCO's Data Team on Performance, Visibility, and Debuggingurn:rss-system-design:45312a386b2026-10-06T11:59:58ZDatabricksIFCO and Databricks cut a dbt job's runtime by about 66% by reading the executed query plan, bounding incremental windows, clustering on the query grain and running one task per model.Don't bet on SELECT's order: shuffle your rows and see what foldsurn:rss-system-design:23378d3e182026-10-06T11:59:57ZEvil MartiansEvil Martians shuffled unordered SELECT results to expose order-dependence bugs in Gitea, Rails and Django, and explains the trade-offs in how their shuffling tools work.NTS: Authenticated Time at Metaurn:rss-system-design:982e05c64d2026-10-06T11:59:56ZMetaMeta built an authenticated time service with Network Time Security, using stateless cookies from a derived daily key and separate key-exchange and NTP servers.From a Trace to Money Flow: How we built an interactive visualization to uncover money launderingurn:rss-system-design:69f9338d4c2026-10-06T11:59:55ZFeedzaiFeedzai explains how it redesigned a fraud analysis graph into a Money Flow visualization that combines a network, a table and a sankey diagram to trace laundering networks.AICR v1.0: Open, stable, and verifiable GPU cluster configurationurn:rss-system-design:8b9eb794762026-10-06T11:59:54ZNVIDIANVIDIA announces AICR v1.0, which provides version-locked, validated GPU Kubernetes cluster recipes with signed evidence, deployer-neutral bundles and a stable compatibility contract for its CLI, REST API, Go SDK and schemas.Control How Your GPU Shares Work with Green Contextsurn:rss-system-design:269dd578832026-10-06T11:59:53ZNVIDIAExplains CUDA green contexts, available in the Runtime API since CUDA 13.1, which partition GPU SMs and workqueues, with a Blackwell benchmark cutting critical-kernel latency to 0.007 ms.How DOCA GPUNetIO Unifies GPU-Initiated Networking Across the NVIDIA Software Stackurn:rss-system-design:0ab304e5c72026-10-06T11:59:52ZNVIDIANVIDIA describes DOCA GPUNetIO, which lets CUDA kernels drive RDMA, Ethernet and DMA operations, its open-source version, and its use as a shared GDA-KI layer in NCCL, NVSHMEM and NVQLink.Scale Bitwise-Deterministic Pretraining with NVIDIA Megatron Coreurn:rss-system-design:23edddcf342026-10-06T11:59:51ZNVIDIADescribes how NVIDIA Megatron Core verifies, debugs and optimizes bitwise-deterministic pretraining, using a trillion-parameter Nemotron case study where determinism overhead fell to approximately 2% at 2,432 GPUs.Metrics Board: Building an Agent-ready Metrics Layerurn:rss-system-design:197d134e722026-10-06T11:59:50ZPinterestPinterest built Metrics Board, an agent-ready metrics layer that accepts any SQL with modelling metadata, runs one Airflow DAG per metric and publishes in two phases.The state of the tech industry in 2026urn:rss-system-design:659ba4f1512026-10-06T11:59:49ZThe Pragmatic EngineerWritten version of a keynote on the state of the tech industry in 2026, covering what changed with AI agents, what broke, what stayed the same and what comes next.NEAREST BY Join: Scaling Vector Search in Databricks Runtimeurn:rss-system-design:ae0435475e2026-10-05T12:00:00ZDatabricksDatabricks moved batch vector search from a per-row endpoint to a native top-k NEAREST BY join, with a fused GEMM operator and an IVF index stored as a Delta table.ReviewBench: An open benchmark for AI code reviewurn:rss-system-design:ee3fc94a6f2026-10-05T11:59:59ZGitHubGitHub built ReviewBench, an open benchmark for AI code review, with a multi-source ground truth and metrics whose offline results matched an online A/B test.Smarter personalization: How property data helps us understand user price sensitivityurn:rss-system-design:dec4559e732026-10-05T11:59:58ZGrabGrab describes an unimplemented proposal to add residential property values to user profiles, using LLM-structured property data, building-level price estimates and clustered affluence tiers.Building and testing a config management orchestrator for Windowsurn:rss-system-design:4e8b809e292026-10-05T11:59:57ZJane StreetJane Street replaced a per-host PowerShell configuration system for Windows with a central OCaml orchestrator and a stateless F# agent, migrating in stages with an audit mode.Unsupervised Topic Discovery with Text Embeddingsurn:rss-system-design:20cd67e6ce2026-10-05T11:59:56ZWealthfrontWealthfront compared bag-of-words, LLM labelling and embeddings with UMAP and HDBSCAN to discover topics in call notes, then added a KNN classifier for new data.Powering AI-led research through simulationurn:rss-system-design:9c11b5f9912026-10-04T12:00:00ZGrabGrab built the sim-rs marketplace simulator with pluggable stages, reproducible runs and safeguards against agents gaming metrics, reporting 10x and 24x speedups for AI-led research.Scaling and benchmarking a critical message bus using a new indexing strategyurn:rss-system-design:39b1c476592026-10-04T11:59:59ZJane StreetJane Street rebuilt the indexing of its Aria message bus with per-partition indexes and a shared block pool, reporting a 30% CPU reduction after a 13-minute recovery incident.Architect highly available Oracle Database on Amazon EVS and FSx for ONTAPurn:rss-system-design:367a9821e32026-10-02T12:00:00ZAWSReference design for highly available Oracle Database on Amazon EVS with FSx for NetApp ONTAP NFS datastores, i7i bare metal hosts, SnapMirror cross-region DR, security controls and cost options.Deploy open source Regional availability tools in your VPCurn:rss-system-design:e2b31f9c492026-10-02T11:59:59ZAWSAWS introduces Capability Insights for AWS, an in-VPC Regional availability dashboard refreshed every 24 hours, and Workload Analysis, which uses CloudTrail and CloudFormation data to narrow 200+ services to those an account runs.2026 Birthday week: network performance updateurn:rss-system-design:d78d05baa02026-10-02T11:59:58ZCloudflareCloudflare reports it is fastest in 74% of the 1,000 largest networks, up from 60%, and adds Challenge Page measurements to its error-page based methodology.Announcing Cloudflare OHTTP Gateway – expanding access to Cloudflare’s privacy-preserving infrastructureurn:rss-system-design:1f522f410f2026-10-02T11:59:57ZCloudflareCloudflare launched a closed beta OHTTP Gateway that decapsulates Oblivious HTTP requests for apps behind Cloudflare, complementing its renamed OHTTP Relay.Protected Quick Tunnels: simple accountless authentication for your next dev projecturn:rss-system-design:86aa5ffa972026-10-02T11:59:56ZCloudflareCloudflare built Protected Quick Tunnels, accountless authentication that splits work between a stateless auth broker and local authorization in cloudflared, after rejecting two other designs.Streamline: custom video pipelines with Cloudflare Stream and Workersurn:rss-system-design:e733c3de832026-10-02T11:59:55ZCloudflareCloudflare describes an architecture for long-running custom video pipelines with Containers, a Durable Object orchestrator and a Worker control plane on Cloudflare Stream.Agentic on the coach's couch: 24 questions, eight projects (translated from French)urn:rss-system-design:fff4ba5c8f2026-10-02T11:59:54ZOCTO TechnologyOCTO built a 24-question assessment grid, scored from evidence in each repository, for eight AI-agent-assisted projects and proposes it as an extra input for team coaching diagnostics.Accelerating airline retailing innovation: how Datalex modernized with AWS Experience-Based Acceleration and agentic AIurn:rss-system-design:3503217de02026-10-01T12:00:00ZAWSDatalex used a three-day AWS Experience-Based Acceleration workshop to move a Reservation component slice to Spring Boot on Java 21, build CI/CD on Amazon ECS and prototype Bedrock AgentCore agents.AI Search is now generally availableurn:rss-system-design:468c7037a62026-10-01T11:59:59ZCloudflareCloudflare made AI Search generally available, adding native image embeddings alongside captions, OCR for scanned PDFs, 10 MiB files and usage-based pricing from November 1, 2026.Announcing Cloudflare K2: serverless event streamsurn:rss-system-design:34410eab952026-10-01T11:59:58ZCloudflareCloudflare K2 is a serverless durable event log built on R2 object storage, accepting about one second of p99 produce latency for simpler replication.Introducing Clef: our open-source decision models, and new RL fine-tuning platformurn:rss-system-design:b82838909f2026-10-01T11:59:57ZCloudflareCloudflare released Clef and Clef-flash, open-source decision models on Workers AI that return typed classifications with probabilities, and a reinforcement learning fine-tuning service.Introducing Cloudflare Basin: an open, serverless data platform, now generally availableurn:rss-system-design:8b31ee03802026-10-01T11:59:56ZCloudflareCloudflare renamed its data platform Basin and made it generally available: Pipelines, Catalog and SQL on Apache Iceberg and R2, with ingestion up to 3GB/s per stream.Introducing Workers KV Instant — powered by Quicksilverurn:rss-system-design:dfea90186a2026-10-01T11:59:55ZCloudflareCloudflare introduced Workers KV Instant, a Quicksilver-backed KV mode with faster replication but limits of 1 MB per namespace and one write per second.Support for modern cryptographic algorithms in Workersurn:rss-system-design:277ffec4332026-10-01T11:59:54ZCloudflareCloudflare Workers added opt-in post-quantum ML-KEM and ML-DSA algorithms to Web Crypto, backed by BoringSSL in workerd, behind a compatibility flag while the specification moves.Lakebase Postgres branch-based restores for fast recovery at scaleurn:rss-system-design:e79eb7ef5d2026-10-01T11:59:53ZDatabricksDatabricks explains how Lakebase Postgres restores a database as a metadata-only branch on decoupled storage instead of the multi-step point-in-time recovery process.How recurring network maintenance exposed 6 bugsurn:rss-system-design:24c5143edc2026-10-01T11:59:52ZJane StreetJane Street traced how recurring network maintenance exposed six bugs in Kafka relays, including DNS hangs, retry storms and file-descriptor exhaustion across glibc, Async and its Kafka client.Build Local AI Apps with C++ and NVIDIA TensorRT RTX Samplesurn:rss-system-design:182f0cde602026-10-01T11:59:51ZNVIDIANVIDIA introduces DIN Deploy, open-source C++ samples using ONNX Runtime with the TensorRT RTX execution provider for local ASR, SAM 2.1 segmentation and FLUX.2 image generation, with DGX Spark benchmarks.Experience report: AI agents and skills on eight projects, we keep the wheel and the ditch can wait (translated from French)urn:rss-system-design:6bfc1904782026-10-01T11:59:50ZOCTO TechnologyAn OCTO consultant describes a hub that measures, proposes and applies changes across eight agent-assisted projects, with rules that came from four incidents.ShopGym: Realistic, reproducible sandboxes for shopping agentsurn:rss-system-design:56ca8429bd2026-10-01T11:59:49ZShopifyShopify built ShopGym, which turns live storefronts into resettable sandbox shops for testing shopping agents, and measured where the copies diverge from live stores.How MHK built a HIPAA-eligible agentic AI solution on Amazon Bedrockurn:rss-system-design:5624b84fc62026-09-30T12:00:00ZAWSMHK built the SmartProminence AI Orchestrator, a HIPAA-eligible controller-agent framework on AWS using Bedrock, ECS Fargate and SQS, reducing manual review effort by 90% and cutting feature deployment to 2 weeks.Running multi-day AZ evacuation drills with ARC Zonal Shifturn:rss-system-design:67dbc8cb482026-09-30T11:59:59ZAWSAWS explains how to run a 48 to 72 hour Availability Zone evacuation drill with ARC Zonal Shift across ECS, EKS, RDS for PostgreSQL and Aurora PostgreSQL, with monitoring metrics and restore steps.Cloudflare Containers, rebuilt to scale agent sandboxesurn:rss-system-design:0f52892baf2026-09-30T11:59:58ZCloudflareCloudflare moved Containers from deploy-time configuration to per-sandbox control in a Durable Object to scale agent sandboxes, with the migration plan and measured startup results.Cut your AI spend with AI Gateway's Auto Routerurn:rss-system-design:9f522b01c32026-09-30T11:59:57ZCloudflareCloudflare built the AI Gateway Auto Router, in which a classifier feeds a scoring matrix that balances model quality against cost, with a cache-dependent penalty for switching models.Identify AI model overuse with User Insightsurn:rss-system-design:02c0c173df2026-09-30T11:59:56ZCloudflareCloudflare added task context to AI Gateway User Insights, including a model overkill view that flags requests sent to models more capable than the task needs.Monetization Gateway beta: charge AI agents for consumption with HTTP 402urn:rss-system-design:e008d3bb3e2026-09-30T11:59:55ZCloudflareCloudflare's Monetization Gateway closed beta lets site owners charge AI agents per request for websites, APIs, MCP tools or datasets using HTTP 402 Payment Required.Pay Per Use: when AI uses your work, you should get paidurn:rss-system-design:b5fd7ca54b2026-09-30T11:59:54ZCloudflareCloudflare's Pay Per Use beta lets AI companies offer a price for each use of a publisher's content, report each use through an API, and have Cloudflare bill buyers and pay publishers.Simplifying domains for people and agentsurn:rss-system-design:30863997b72026-09-30T11:59:53ZCloudflareCloudflare describes domain search across 420+ extensions that combines precomputed datasets, DNS, caches and live registry checks, with one Durable Object coordinating each search.The Internet has a second audienceurn:rss-system-design:7a8efd08c72026-09-30T11:59:52ZCloudflareCloudflare argues that agents now form a second web audience, citing request growth from 63 to 115 million per second, and describes controls, payments and cost reductions for serving them.A FUSE server for our in-house storage backendurn:rss-system-design:da2df5102d2026-09-30T11:59:51ZJane StreetJane Street built a FUSE server in OCaml that maps its in-house object store to POSIX, raising read throughput from about 30 MiB/s to 1.3 GiB/s by removing copies.A resilient kernel log reporterurn:rss-system-design:e51b7c45ac2026-09-30T11:59:50ZJane StreetJane Street built a kernel log reporter that avoids disk and network dependencies using mlockall, /dev/kmsg, cached DNS, UDP and pstore, and used it to diagnose a real crash.Deploying an HSTU Generative Recommender with NVIDIA Dynamo-Tritonurn:rss-system-design:3d3417f2d52026-09-30T11:59:49ZNVIDIANVIDIA describes serving HSTU generative recommender models with PyTorch AOTI, NV Embedding Cache, FlexKV-backed KV caching and Dynamo-Triton, reporting up to 5.93x lower latency with GPU KV-cache hits.Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languagesurn:rss-system-design:64e9abde362026-09-30T11:59:48ZNVIDIANVIDIA fine-tunes Nemotron 3.5 ASR on Najdi and Hijazi speech with minimal curation, weighted FLEURS replay and duration bucketing, cutting target WER from 55.05% to 29.96% while retaining English.Tracing Agent Harness Behavior with NVIDIA NeMo Relayurn:rss-system-design:4c2efc6a872026-09-30T11:59:47ZNVIDIAAn NVIDIA tutorial on tracing Hermes Agent runs with NeMo Relay using ATOF, ATIF and OpenTelemetry traces, plus a 108-run Hermes ToolPerf A/B study of tool-layer fixes on two models.Distributed databases with Peter Mattisurn:rss-system-design:3ae27470c62026-09-30T11:59:46ZThe Pragmatic EngineerNotes from a podcast with Peter Mattis of Cockroach Labs on distributed databases, Colossus erasure coding, B-trees versus LSM trees, replica counts, and how AI changes code writing and review.Normalizing a Wide Well and Pipeline Telemetry Table Without Re-Ingesting Historyurn:rss-system-design:23f79c61e02026-09-30T11:59:45ZTimescaleA guide to migrating a live wide time-series table in TimescaleDB to a narrow table plus metadata table without re-ingesting history, using sliced transactions and a compatibility view.Build adaptive AI interfaces with the AG-UI protocol, agent swarms, and Nova Act on AWSurn:rss-system-design:6d44f447e92026-09-29T12:00:00ZAWSAWS reference design for a radiology assistant using the AG-UI protocol over SSE, a Strands swarm of three debating agents with an 80% consensus threshold, and Amazon Nova Act for legacy RIS and EMR submission.Can you use autoregressive diffusion to generate market data?urn:rss-system-design:8dd66e73032026-09-29T11:59:59ZJane StreetJane Street describes designing an autoregressive diffusion model to generate market data, including which configurations failed or did not scale and what the results showed.Mitigating memorization in LLMsurn:rss-system-design:fb6d4207562026-09-29T11:59:58ZJane StreetA Jane Street intern measured when LLMs succeed by recalling memorized outcomes and tested mitigations such as divergence decoding, distillation and synthetic rewrites on a cricket dataset.Trading off compute for memory with activation checkpointingurn:rss-system-design:54e57c5e172026-09-29T11:59:57ZJane StreetJane Street designed a planner that chooses which activations to checkpoint, trading compute for memory with a greedy algorithm after finding gaps in PyTorch's built-in options.What the SP interns have wrought, 2026 editionurn:rss-system-design:799fea18cc2026-09-29T11:59:56ZJane StreetJane Street describes Strategy and Product intern projects, including modeling loan fee expense and allocating financing costs to desks, and managing risk-limit overrides on special dates.What the TDOE interns have wrought, 2026 editionurn:rss-system-design:0edf0d705c2026-09-29T11:59:55ZJane StreetJane Street describes three Trading Desk Operations intern projects, including a system that flags discrepancies between internal voice-trade records and the public tape.AI Native by Design: Lessons Learned from Building NVIDIA TensorRT Model Connecturn:rss-system-design:cf46223ea72026-09-29T11:59:54ZNVIDIANVIDIA describes lessons from building TensorRT Model Connect, an open source C++ reference implementation collection covering 128 model families, around coding agents with isolation, reversibility and automated validation.Lower the Cost of Building and Running Visual AI Agents with NVIDIA VSS Blueprint 3.3urn:rss-system-design:b9ecbf3ecc2026-09-29T11:59:53ZNVIDIANVIDIA VSS Blueprint 3.3 adds a Build Vision Agent skill that composes video agent deployments from four profiles and Adaptive EVS, which cut VLM input tokens 80% and raised concurrent streams from 13 to 19.How to roll out Genie One: A step-by-step enterprise playbookurn:rss-system-design:26f8d7081d2026-09-28T12:00:00ZDatabricksDatabricks lays out a four-phase rollout of Genie One, from one governed sales ops domain and a 5 to 10 user pilot to org-wide access with federated ownership and Unity Catalog governance.Add Runtime Controls to AI Agents with NVIDIA OpenShellurn:rss-system-design:87dd0c72c82026-09-28T11:59:59ZNVIDIANVIDIA OpenShell 0.1.0 is an open-source agent runtime with a gateway, per-sandbox supervisor and kernel-isolated sandbox, YAML policies compiled to OPA/Rego, credential protection and a formal policy prover.NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoringurn:rss-system-design:7ecae0e8872026-09-28T11:59:58ZNVIDIANVIDIA introduces the Open Agent Safety Platform, a three-layer stack in which OpenShell sandboxes agents and optional NVIDIA Sentry enforces policy on BlueField-4 DPUs in Vera Rubin POD compute trays.How NVIDIA DSX MaxLPS Maximizes AI Factory Throughput and Efficiencyurn:rss-system-design:f66c13c5f02026-09-27T12:00:00ZNVIDIANVIDIA and Nscale evaluate DSX MaxLPS power sharing on GB300 NVL72 systems running Kimi K2.5, raising managed GPUs from 140 to 192 and aggregate throughput by 49.2% within a 264.4 kW budget.Building Ask DoorDash (Part 6): A Grocery Reorder Case Studyurn:rss-system-design:056c165fee2026-09-25T12:00:00ZDoorDashExplains how DoorDash moved reorder reasoning offline, precomputing consumers' recurring items daily and resolving them against the live catalog, which cut one reorder path from about 16 seconds to about 2 seconds.Automated merchant review summary system with integrated feedbackurn:rss-system-design:daa573a9712026-09-24T12:00:00ZGrabGrab built a merchant review summary pipeline with stratified sampling, separate LLM generation and validation stages, and regeneration triggered by new reviews or negative feedback.Efficient MoE Training for Biological Foundation Modelsurn:rss-system-design:440bdec9772026-09-24T11:59:59ZNVIDIAA tutorial on the NVIDIA BioNeMo MoE recipe using Transformer Engine GroupedLinear, MXFP8 and a fused GroupedMLP kernel, reaching up to 2.21x the Hugging Face baseline throughput on eight B200 GPUs.How SWE-Serve Exposes the Gap Between Local Tests and Live Servingurn:rss-system-design:5a58fad97a2026-09-23T12:00:00ZNVIDIASWE-Serve turns 83 merged SGLang pull requests into 53 tasks for coding agents; live serving tests cut the pass rate from 69.4% to 45.9%, and 11 models were evaluated.Introducing NV-Reason-CT Open 3D CT VLM for Radiologist Chain-of-Thought Reasoningurn:rss-system-design:690ec484952026-09-23T11:59:59ZNVIDIANVIDIA introduces NV-Reason-CT, an open 3D CT vision language model combining Qwen3.5-4B with a Primus 3D ViT, trained with supervised fine-tuning and GRPO, reporting 0.614 Macro-F1 on CT-RATE.Manage Kubernetes Node Fleets with NodeWrighturn:rss-system-design:5cbd0f3cb32026-09-23T11:59:58ZNVIDIANVIDIA introduces NodeWright, formerly Skyhook, an open-source Kubernetes-native package manager that configures and updates GPU node operating systems with cordon, drain and progressive rollout without disrupting workloads.Validate GPU Cluster Readiness Before AI Workloads Landurn:rss-system-design:d96e7f34c42026-09-23T11:59:57ZNVIDIANVIDIA describes NVCRE, an open source Kubernetes controller that runs NCCL, DCGM and NeMo workloads across topology-aware node groups to certify GPU cluster readiness and name failing nodes.Accelerating a ROS 2 Node with an AI Agent and NVIDIA Isaac ROSurn:rss-system-design:6d4e4f6e902026-09-22T12:00:00ZNVIDIANVIDIA shows how an AI agent skill migrates the Depth Anything 3 TensorRT ROS 2 node to the rosidl::Buffer CUDA buffer backend in ROS 2 Lyrical for zero-copy GPU message transport with CPU fallback.Enabling Private High-Performance Production AI Inference with NVIDIA Confidential Computingurn:rss-system-design:62573cb0c72026-09-22T11:59:59ZNVIDIANVIDIA describes TensorRT LLM adaptations for confidential computing on Blackwell GPUs and a CC-on versus CC-off benchmark where DeepSeek-R1 retained 96.1 to 98.2 percent of output-token throughput on eight B200 GPUs.Topology-Aware Workload Scheduling with NVIDIA Topographurn:rss-system-design:f4365050502026-09-22T11:59:58ZNVIDIANVIDIA Topograph discovers cluster network topology from cloud APIs or on-premises fabrics and publishes it as Kubernetes labels, NFD resources, Slurm configuration or Slinky ConfigMaps for topology-aware scheduling.Triton Auto Tuner: Finding the Right Triton Configuration for Hundreds of Modelsurn:rss-system-design:16c8962dd32026-09-21T12:00:00ZDoorDashDescribes DoorDash's Triton auto tuner, which benchmarks hundreds of CPU models on a fixed pool and picks serving configurations, with mean throughput gains of 1.75x to 6.78x depending on CPU tier.How to Evaluate AI Agents From Tool Calls to Task Completionurn:rss-system-design:7efd5bb08b2026-09-21T11:59:59ZNVIDIANVIDIA explains how AI agent evaluation moved from scoring single tool calls to scoring task completion in stateful environments, defining step-level and end-to-end metrics and a SWE-bench Verified trace example.Simplifying Model Serving Across Multiple GPUs with NVIDIA TensorRT Multi-Device Integration in NVIDIA Dynamo-Tritonurn:rss-system-design:ed226b851c2026-09-21T11:59:58ZNVIDIADynamo-Triton 26.07 serves TensorRT multi-device plans across GPUs; Cosmos 3 Nano generation with Ulysses context parallelism dropped from 156.595 to 34.183 seconds on eight GPUs.Turn Your Latest Observations Into Timely Weather Decisions With NVIDIA Earth-2urn:rss-system-design:6919dca5472026-09-21T11:59:57ZNVIDIAA tutorial on AI data assimilation in NVIDIA Earth2Studio, using Score-Based Data Assimilation with CorrDiff and StormCast and the HealDA model to turn observations into weather states for forecasting.How CSIRO built scalable, cost-optimized genomic variant querying on AWSurn:rss-system-design:2c4baab63f2026-09-18T12:00:00ZAWSCSIRO describes sBeacon, a serverless GA4GH Beacon implementation on S3, Lambda, DynamoDB and Athena, benchmarked on 1000 Genomes chromosome 1 with 18 second ingestion and about 1.5 second queries.ReadyOn’s Four Walls of tenant isolation on Amazon EKSurn:rss-system-design:1e1e57fecb2026-09-18T11:59:59ZAWSReadyOn describes its Four Walls tenant isolation on Amazon EKS: GitOps-managed namespaces, dedicated Karpenter node pools, per-tenant VPC security groups and dedicated Aurora clusters, validated against a 10-stage threat model.Building cloud-native PACS on AWSurn:rss-system-design:0885acafac2026-09-17T12:00:00ZAWSAWS presents a hub-and-spoke hybrid architecture for medical imaging PACS, with local PACS at each hospital replicating asynchronously to Amazon S3, S3 storage tiers by study age, and cloud-only versus hybrid guidance.How Equinix cut operational overhead with a shared services architecture on Amazon EKSurn:rss-system-design:b109f4a4d82026-09-17T11:59:59ZAWSEquinix moved from self-managed Kubernetes on EC2 with team-owned clusters to a shared services North Star architecture on Amazon EKS, reporting 40% less operational overhead and 4x deployment frequency.AI Skills with Matt Pocockurn:rss-system-design:f5a472e0bb2026-09-17T11:59:58ZThe Pragmatic EngineerNotes from a podcast with Matt Pocock on using skills such as grill-me, tracer bullets, cloud agents and codebase fundamentals to plan, delegate and course-correct with coding agents.Inside Vera, DoorDash’s Data Agenturn:rss-system-design:ec9089ff0b2026-09-16T12:00:00ZDoorDashDescribes Vera, DoorDash's internal data agent, with evaluation results across 19 domains, a retrieval and ranking layer over modeled data, and a pass rate that rose from 43% to 90%.How to Use AI Agents to Prepare 3D Scenes for Simulationurn:rss-system-design:74574a99c82026-09-16T11:59:59ZNVIDIANVIDIA describes an agent workflow in which Codex with GPT-6 Astra coordinates NemoClaw subagents using Blender MCP and Omniverse Libraries to turn a Blender scene into a SimReady OpenUSD world for Isaac Sim.TensorRT Edge-LLM Completes the MLPerf Edge Agentic Benchmark 6.4x Faster on Jetson AGX Thorurn:rss-system-design:eef51a90bf2026-09-16T11:59:58ZNVIDIATensorRT Edge-LLM ran Qwen3.6-27B on Jetson AGX Thor in the MLPerf v6.1 Edge Agentic benchmark at 52.33 tokens per second, 6.4x faster than llama.cpp, using NVFP4, KV cache reuse and tree-based MTP.Translating CUDA Tile Operations from Python to Rust Using Agentic AIurn:rss-system-design:29379bc9762026-09-16T11:59:57ZNVIDIANVIDIA describes a multi-agent skill that translated all 24 public TileGym operators from cuTile Python and Triton-TileIR to cuTile Rust, using Tile IR diffs and validators, reaching a 0.995 geomean.AI Changed How Spotify Builds. What We Learned (and Fixed) About Quality at Higher Velocityurn:rss-system-design:7b76faa3362026-09-16T11:59:56ZSpotifySpotify reviews quality issues during its AI ramp-up, including a June 24 content-processing incident, failovers under compute shortages and mobile regressions, and its data showing merged changes rising from roughly 8,100 to 17,000.Dense vs. MoE Models: Active Parameters, Throughput, and When to Choose Eachurn:rss-system-design:033fe5de072026-09-15T12:00:00ZNVIDIANVIDIA compares dense and Mixture-of-Experts models, explaining routing, active versus total parameters and serving trade-offs, using Nemotron 3.5 Lightning with 30B total and 3B active parameters as the main example.How NVIDIA Groq 3 LPX Deterministic Execution Drives Power-Efficient High-Interactivity Inference on NVIDIA Vera Rubinurn:rss-system-design:7c29d21d0e2026-09-15T11:59:59ZNVIDIANVIDIA explains how Groq 3 LPX deterministic execution lets the LPU compiler predict per-cycle current and use Preemptive Power and Clock Period Synthesis to reduce voltage droop and power for inference.How NVIDIA NVLink 6 Delivers Multi-Layer Resiliency for AI Factoriesurn:rss-system-design:38adb4ac912026-09-15T11:59:58ZNVIDIANVIDIA describes the NVLink 6 resiliency stack for Vera Rubin NVL72, from FEC, Physical Layer Retry and credit-based flow control to Dynamo Shadow Engine Recovery cutting inference downtime from 283 to 7.3 seconds.Scaling Federated Learning Across Docker, Kubernetes, and Slurm with NVIDIA FLAREurn:rss-system-design:88ab290a222026-09-15T11:59:57ZNVIDIANVIDIA FLARE separates persistent federation services from per-job workers launched on Docker, Kubernetes or Slurm, and adds studies that map jobs to site-owned datasets, images, secrets and scheduling policy.KCP: How to Migrate to Confluent Cloud in Days, Not Weeksurn:rss-system-design:5e8d80fa9d2026-09-14T12:00:00ZConfluentConfluent describes KCP, an open source tool that discovers Amazon MSK clusters into a kcp-state.json file and generates Terraform to migrate topics, ACLs, schemas and connectors to Confluent Cloud using Cluster Linking.Accelerating Dropless MoE Training in JAX with NVIDIA Transformer Engineurn:rss-system-design:0f819a17ee2026-09-14T11:59:59ZNVIDIANVIDIA describes Transformer Engine grouped GEMM, MXFP8 quantization and NCCL EP dispatch and combine for dropless MoE in JAX, raising DeepSeek-V3 training from 103 to 1,068 TFLOPS/GPU with 97% efficiency at 1,024 GPUs.Rapidly scaling online storage to serve over 1 billion ChatGPT usersurn:rss-system-design:c7d3c8aeae2026-09-11T12:00:00ZOpenAIOpenAI describes how Habitat, its online storage platform serving over 70 million requests per second and 500 PB, evolved from a Python library to a service and was rewritten in Rust.Building an Agentic SOC on a Streamurn:rss-system-design:280f36c23e2026-09-10T12:00:00ZConfluentConfluent describes an agentic SOC with a triage coordinator, evidence-domain agents, an adversarial evaluator and Flink priority lanes that investigated about 4,700 alerts in 30 days, escalating roughly 5%.From Wafer-Out to First Token: Codifying Supply Chain Expertise with Nemotron and Palantir Foundryurn:rss-system-design:68e50065a32026-09-10T11:59:59ZNVIDIANVIDIA and Palantir describe a supply chain allocation workflow using Palantir Foundry, NVIDIA cuOpt and a LoRA post-trained Nemotron 3.5 Lightning model that reached 86.7% allocation-decision accuracy.High-Throughput Structure Prediction with BioNeMo Inference Runtimeurn:rss-system-design:de969d49852026-09-10T11:59:58ZNVIDIAA tutorial on NVIDIA BioNeMo Inference Runtime for Boltz-2 structure prediction, covering serial and Ray backends and a 1,000-target benchmark showing 2.90x residue throughput on 8xH100 GPUs.How Full-Stack NIM Optimizations Deliver 2.5x More Users on Nemotron 3 Ultraurn:rss-system-design:eca8346ffb2026-09-10T11:59:57ZNVIDIANVIDIA describes how NIM 2.0.12 serving optimizations, including autotuned kernels, prefix reuse, scheduler tuning and MTP speculative decoding, raised Nemotron 3 Ultra throughput on 4xB200 from 718 to 1,997 tok/s at 50 TPS per user.When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Servingurn:rss-system-design:2751a48f602026-09-09T12:00:00ZNVIDIANVIDIA explains encode-prefill-decode disaggregation in Dynamo for multimodal models, benchmarking aggregated, colocated and disaggregated encoder topologies on GB200 and RTX 6000D GPUs with Qwen3.5 models.Building Codex with Tibo Sottiauxurn:rss-system-design:bd9f5944e42026-09-09T11:59:59ZThe Pragmatic EngineerNotes from a podcast with Tibo Sottiaux on how OpenAI built Codex, covering the choice of Rust, open source, multi-model support, the harness, and AI code review and maintenance.Introducing CUDA Rust: Two Tracks for Writing GPU Kernelsurn:rss-system-design:e3813994de2026-09-08T12:00:00ZNVIDIANVIDIA introduces CUDA Rust with two tracks, the cuda-oxide SIMT codegen backend and the cutile-rs Tile track, showing a vector addition kernel in each and how Rust ownership catches aliasing.Why Spotify Is Not Using Bayesian A/B Testingurn:rss-system-design:d857e04e6b2026-09-08T11:59:59ZSpotifySpotify explains why its experimentation platform keeps frequentist analysis only, arguing Bayesian A/B testing is a family of configurations and that empirical Bayes priors are hard to maintain.Building a Memory-Driven Agent with NVIDIA NemoClawurn:rss-system-design:fb4b499a382026-09-04T12:00:00ZNVIDIANVIDIA describes a NemoClaw Chief of Staff agent using a Markdown self model, a SQLite obligation ledger and OpenShell sandboxing, scoring 90.9% versus 82.8% for agentic RAG on 186 questions.Frontier Reasoning Reaches the Edge: How to Deploy and Optimize Models on NVIDIA Jetsonurn:rss-system-design:76e0669fae2026-09-04T11:59:59ZNVIDIAGuide to running Nemotron 3.5 Lightning and Qwen3.8-27B on NVIDIA Jetson with vLLM, combining NVFP4 quantization and speculative decoding methods MTP, DFlash and DSpark, with SpeedBench throughput results.How to Carry User Identity Across Federated Kubernetes and AI Platformsurn:rss-system-design:972797ae1a2026-09-03T12:00:00ZNVIDIANVIDIA describes a central identity gateway using OIDC, a Redis session store and a /gateway/userinfo validation endpoint for federated Kubernetes platforms, which reduced repeated login events by 55% across AWS and OCI clusters.NVIDIA PAIR Virtual Inference Router Expands Available Compute on Your Local Networkurn:rss-system-design:87b23dabf12026-09-03T11:59:59ZNVIDIANVIDIA PAIR beta routes independent Ollama and LM Studio inference requests from multi-agent workloads to paired machines on a home network, cutting a five-subagent Hermes demo from 18 minutes to 8 minutes 48 seconds.Portal by Spotify cut my Claude Code token usage by 90%urn:rss-system-design:a38ea74ec02026-09-03T11:59:58ZSpotifyA Claude Code plugin called shunt uses hooks to route large file reads and boilerplate code generation to Gemini 2.5 Flash AiKA Modes in Portal, cutting bulk-read token use by around 90 percent.Co-Designing AI Models Using Speculative Decoding for Faster LLM Inferenceurn:rss-system-design:eb9b5454702026-09-02T12:00:00ZNVIDIANVIDIA presents five guidelines for choosing speculative decoding draft length and draft mechanism across the Pareto frontier, comparing MTP, EAGLE-3, DFlash, DSpark, external drafts and n-gram on SPEED-Bench.The Modern CUDA Toolbox in Practice: A Step-by-Step Optimization Walkthroughurn:rss-system-design:747fd24caa2026-09-02T11:59:59ZNVIDIAAn example CUDA image pipeline is improved in six steps using Compute Sanitizer, NVTX and Nsight Systems, CUB, pooled and pinned memory and per-thread streams, cutting runtime from 6.8 seconds to 23 milliseconds.Building an Adaptive Agentic Cybersecurity System with NVIDIA Nemotronurn:rss-system-design:f5c3936fbb2026-09-01T12:00:00ZNVIDIANVIDIA and CrowdStrike evaluated red and blue agents built on Nemotron 3 Ultra and a fine-tuned Nemotron 3 Super, raising backtest detection pass rates from 16.5% to 41.9%.How to Size GPUs for AI Inference and TCO Without Overspendingurn:rss-system-design:f77fde1f552026-09-01T11:59:59ZNVIDIANVIDIA guide to sizing GPUs for LLM inference by token patterns, concurrency, latency and contract length, with core-and-flex capacity, FP8 quantization of Llama-3.1-8B and pruning Qwen3-8B to about 6B.Scale AV Perception Across Vehicle Platforms with NVIDIA Omniverse NuRecurn:rss-system-design:f66af0c63c2026-08-31T12:00:00ZNVIDIANVIDIA shows how Omniverse NuRec reconstructs recorded drives with 3D Gaussian splatting, renders them through a new carline camera rig, refines them with Harmonizer and prepares data for perception training.Deploy an Open Model from Checkpoint to Inference in Two Commands with NVIDIA TensorRT Model Connecturn:rss-system-design:36ea24db0b2026-08-28T12:00:00ZNVIDIANVIDIA introduces TensorRT Model Connect, reference implementations that build a TensorRT bundle from a Hugging Face model ID with trtmc build and run it from native C++ without PyTorch.Hemut: Building the Internet of Freight on Real-Time Dataurn:rss-system-design:a92b49fccb2026-08-27T12:00:00ZConfluentHemut, an AI-native operating system for trucking carriers, built a real-time data foundation on Confluent Cloud that has been in production since March 2026 and has carried more than 286 million events.Building Ask DoorDash (Part 5): A Grounded Interface For Shopping Agentsurn:rss-system-design:ecf170f54d2026-08-27T11:59:59ZDoorDashDescribes how Ask DoorDash keeps each shopping list as one stored artifact with separate agent and consumer views, sends deterministic edits straight to it, and passes app context to the agent as scope.Experiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Codingurn:rss-system-design:31dc16f83f2026-08-26T12:00:00ZNVIDIANVIDIA announces Day 0 support for Alibaba's Qwen3.8-Flash-Next, a hybrid Gated DeltaNet and Qwen Sparse Attention MoE model, reporting over 16K tokens per second per GPU on GB300 NVL72.NVIDIA NVLink Fusion Brings NVHBM to Next-Generation AI Infrastructureurn:rss-system-design:6e4e2ae6ba2026-08-26T11:59:59ZNVIDIANVIDIA describes NVHBM, a custom HBM base die offered via NVLink Fusion, claiming up to 30% more bandwidth, up to 67% less PHY area and 15% lower HBM power than HBM4e.CUDA Python 1.0: Stable APIs, One Foundation, Full Platform Accessurn:rss-system-design:37d720a2102026-08-25T12:00:00ZNVIDIANVIDIA releases CUDA Python 1.0 with CUDA 13.3, a semantically versioned set of libraries including cuda.core, cuda.compute, cuda.bindings and nvmath-python built on one shared foundation.Restore LLM Inference Capacity in Seconds with Shadow Engine Recovery in NVIDIA Dynamourn:rss-system-design:fa67efb4332026-08-25T11:59:59ZNVIDIANVIDIA Dynamo shadow engine recovery keeps a pre-initialized standby engine sharing weights via the GPU Memory Service, restoring a GLM-5.2 worker in 7.3 seconds versus a 283-second cold restart.Jalapeño’s first results show industry-leading speed and efficiency in AI inferenceurn:rss-system-design:9ad99bb5f42026-08-25T11:59:58ZOpenAIOpenAI reports first results for Jalapeño, its first custom inference chip, showing 1.5 to 1.9 times more work per watt and 1.7 to 3.6 times lower latency on InferenceX across three models.Automating Feature-Flag Cleanup at Scale with a Multi-Agent LLM Systemurn:rss-system-design:5d0e0389302026-08-24T12:00:00ZDoorDashA multi-agent LLM system that turns stale feature-flag Jira tickets into merge-ready pull requests, with 45 of 50 recent flags cleaned up at an average cost of $4.79 each.Giga-Scale AI and the Ethernet Evolution: How Spectrum-X Ethernet Rewrites the Rulesurn:rss-system-design:9116fdfc752026-08-24T11:59:59ZNVIDIANVIDIA explains why traditional Ethernet struggles with AI collectives and how Spectrum-X uses adaptive routing, targeted congestion control and a NIC plane load balancer, reporting 2.68 ms failover.How NVIDIA Groq 3 LPX Unlocks Ultrafast Interactivity at Long Context on NVIDIA Vera Rubinurn:rss-system-design:be7c2e9c052026-08-24T11:59:58ZNVIDIANVIDIA describes how Groq 3 LPX uses compiler-scheduled chip-to-chip communication and fine-grained compute overlap to serve Gemma 4 31B at 3,431 output tokens per second at 100K context.Maximizing AI Factory Performance per Watt with NVIDIA DSX MaxLPSurn:rss-system-design:0c5b454e1e2026-08-24T11:59:57ZNVIDIANVIDIA describes DSX MaxLPS, which uses Dynamic Power Software to reallocate stranded rack power, workload power profiles and 45° C liquid cooling, projecting up to 40% more Rubin GPU capacity within the same power budget.NVIDIA BlueField-4 Powers New Scale-In Network Infrastructure for Agentic AI Factoriesurn:rss-system-design:9cae77aa752026-08-24T11:59:56ZNVIDIANVIDIA introduces Scale-In infrastructure using BlueField-4 DPUs, DOCA and Spectrum-X Ethernet to offload security, tenant isolation, storage access and provisioning from host CPUs in AI factories.NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watturn:rss-system-design:213bcc052e2026-08-24T11:59:55ZNVIDIANVIDIA presents SemiAnalysis AgentX results, which replay Claude Code sessions, showing GB300 NVL72 up to 15x the throughput per megawatt of H200 NVL8 and Vera Rubin NVL72 up to 30x over GB300.Solving Agentic AI Fleet Challenges with NVIDIA Vera CPUurn:rss-system-design:514a3ff5ea2026-08-24T11:59:54ZNVIDIANVIDIA uses telemetry from 163,594 agentic sessions to argue that agent fleets need one balanced CPU design point and claims Vera delivers up to 1.5x the agentic performance of competitors.GPU-Accelerated Clustering for Financial Instruments at Scaleurn:rss-system-design:afbc0226932026-08-21T12:00:00ZNVIDIANVIDIA describes a GPU SymNMF clustering pipeline with the AdaptGrow solver that groups financial instruments from correlation and tail-dependence matrices, scaling to 100,000 instruments on one GB200 and 1 million across 16 nodes.NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agentsurn:rss-system-design:49d124445f2026-08-21T11:59:59ZNVIDIANVIDIA describes AVO, a long-horizon agent architecture with persistent memory and a supervisor, which optimized attention kernels over seven days and scored 100.00 RHAE on the ARC-AGI-3 public set.Where Security Fits in an AI Agent Stackurn:rss-system-design:a047b7204d2026-08-21T11:59:58ZNVIDIANVIDIA maps the AI agent stack into layers from models and harnesses to secure runtimes such as OpenShell and argues enforcement belongs in the runtime below the agent, with four security profiles.How Generative Recommenders Are Redefining RecSys at Scaleurn:rss-system-design:cdef83f6df2026-08-20T12:00:00ZNVIDIANVIDIA describes generative recommenders based on HSTU and Semantic IDs and its recsys-examples and nv-embedding-cache libraries for dynamic embeddings, KV caching and tiered embedding lookup, with H100 benchmark results.Building Federated Multimodal AI Workflows with NVIDIA FLAREurn:rss-system-design:4c03a8029a2026-08-19T12:00:00ZNVIDIAThe post shows how NVIDIA FLARE federates vision-language model training using object externalization, tensor streaming and disk-backed aggregation, and how FedUMM cut per-client communication from 28.6 GB to 0.094 GB.Developing NVIDIA Holoscan Applications with CLI, Skills, and AI Coding Agentsurn:rss-system-design:f58673775d2026-08-19T11:59:59ZNVIDIANVIDIA builds a real-time endoscopic tool segmentation app in HoloHub with a coding agent over three iterations and compares agent time and token usage with and without the CLI, skills and documentation.Evaluating AI Agent Skill Performance with NVIDIA SkillEvaluatorurn:rss-system-design:fcada65fd82026-08-19T11:59:58ZNVIDIANVIDIA introduces SkillEvaluator, an open source tool that evaluates agent skills through static checks, distinctiveness checks and live Harbor sandbox runs, reporting results for more than 300 skills across over 30 products.Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Controlurn:rss-system-design:7519b4bb122026-08-19T11:59:57ZNVIDIAThe tutorial post-trains the 4B Cosmos 3 Edge model on the Cosmos3-DROID dataset into a robot policy that runs on Jetson AGX Thor and reaches 22.9% success on RoboLab.How Agoda Scaled Its Price Cache with DragonflyDBurn:rss-system-design:cb0fad365e2026-08-18T12:00:00ZAgodaAgoda describes migrating its Price Cache from a 72-shard SQL Server setup to DragonflyDB, using load tests, dual-read parity checks, an A/B rollout and automatic per-pod failover, improving P99 read latency eight times.How AI Coding Agents Can Unlock Materials Simulation with NVIDIA ALCHEMI Toolkiturn:rss-system-design:782bc9685a2026-08-18T11:59:59ZNVIDIANVIDIA benchmarks AI coding agents using ALCHEMI Toolkit agent skills to generate 45 GPU-accelerated atomistic simulation pipelines across three workflows and five prompt levels, validated on NVIDIA H200 GPUs.Run Massive-Scale UMAP in Minutes Using Multiple GPUs—Without Losing Accuracyurn:rss-system-design:d0ca6ad0052026-08-18T11:59:58ZNVIDIAcuML and cuVS 25.06 distribute UMAP all-neighbors kNN graph construction across multiple GPUs using overlapping balanced clusters, processing the 106M-vector MIRACL dataset in 8 minutes on eight H100 GPUs.Developing Nemotron 3.5 Lightning NVFP4 with QAD Using NVIDIA Model Optimizerurn:rss-system-design:14f6b5d39d2026-08-17T12:00:00ZNVIDIANVIDIA describes using quantization-aware distillation in Model Optimizer to build a 22 GB NVFP4 checkpoint of Nemotron 3.5 Lightning, comparing PTQ recipes and QAD results across intermediate and final checkpoints.When Can LLMs Replace Humans in A/B Tests?urn:rss-system-design:96b6f867292026-08-13T12:00:00ZSpotifySpotify authors use the Upworthy archive and gpt-4o-mini to show raw LLM predictions recover 39% of human A/B treatment effects, and that tree-based calibration works only under surrogacy and comparability assumptions.Serve Qwen3.8-2.4T-A95B, a 2.4T-Parameter Model, with Configurable Reasoning on NVIDIA GB300 NVL72urn:rss-system-design:ee0ca67a3d2026-08-12T12:00:00ZNVIDIADescribes serving Alibaba's open-weight Qwen3.8-2.4T-A95B, a 2.4T-parameter MoE model with 95B active parameters, on NVIDIA GB300 NVL72 at over 4K tokens per second per GPU in FP8.How we built a realtime system for responsive voice AI in six monthsurn:rss-system-design:d04cc35bd32026-08-03T12:00:00ZOpenAIOpenAI describes the GPT-Live voice architecture, including a Go media path, model instance handoff for context compaction, delegation to GPT-5.5, the WARP and Instant Connect protocols and a silent production test.How GPT-5.6 fuses frontier intelligence with frontier efficiencyurn:rss-system-design:6d964f10c32026-07-28T12:00:00ZOpenAIOpenAI explains how GPT-5.6 Sol in Codex optimized load balancing, production kernels and speculative decoding, cutting serving costs by 20%, and how its Rust agentic harness limits context and preserves prompt caching.Indexing the Data Lake for Online Point Queriesurn:rss-system-design:617692e6352026-07-27T12:00:00ZSpotifySpotify describes Random Access Parquet, an external key-to-file-and-row index plus Parquet write-time layouts that let online services and AI agents run fast point queries directly on files in its GCS data lake.How We Eliminated Configuration-Related Outages and Scaled Consul Management at Agodaurn:rss-system-design:44685783122026-07-08T12:00:00ZAgodaAgoda moved over 34,000 Consul key-value configurations across four data centers from UI edits to a GitOps workflow with JSON Schema validation, CODEOWNERS review and a sync service, onboarding 100+ teams.Deep Dive into SASL PLAIN and SCRAM in Kafka: Login Modules and Config Hot-Reloadurn:rss-system-design:c9fe13a7032026-07-06T12:00:00ZConfluentExplains how Kafka's PlainLoginModule, ScramLoginModule and FileBasedLoginModule store SASL credentials and whether they support hot-reload, with a three-listener broker example and recommendations for zero-downtime credential rotation.Core dump epidemiology: fixing an 18-year-old bugurn:rss-system-design:4c4d91d3ce2026-06-29T12:00:00ZOpenAIOpenAI traced crashes in its C++ Rockset service to a bad Azure host and an 18-year-old race condition in GNU libunwind, found by automatically labeling a year of production core dumps.How Agoda Simulates Booking Flows to Test Flight Integrationsurn:rss-system-design:04c75578992026-05-28T12:00:00ZAgodaAgoda describes its Workflow Simulator, which tests flight supplier connectors by simulating upstream callers through a DAG-based executor, scenario builder and endpoint and workflow assertions instead of a full end-to-end stack.Building self-improving tax agents with Codexurn:rss-system-design:9da64821c62026-05-27T12:00:00ZOpenAIOpenAI and Thrive Holdings describe Tax AI, which processed 7,000 returns for Crete firms and uses practitioner corrections, production traces and targeted evals to drive Codex-generated fixes.How Agoda Powers Multiple Travel Brands with One Design Systemurn:rss-system-design:d0975e48342026-05-13T12:00:00ZAgodaAgoda describes its drone-js design system, where design tokens and deep-merged theme JSON generate per-brand CSS packages applied at runtime to support multiple white-label travel partners.Building a safe, effective sandbox to enable Codex on Windowsurn:rss-system-design:5502a50a462026-05-13T11:59:59ZOpenAIOpenAI explains how it built a Windows sandbox for Codex using synthetic SIDs, write-restricted tokens, dedicated sandbox users, firewall rules and separate setup and command-runner binaries.Unlocking large scale AI training networks with MRC (Multipath Reliable Connection)urn:rss-system-design:a654fb66812026-05-05T12:00:00ZOpenAIOpenAI describes MRC, a protocol developed with AMD, Broadcom, Intel, Microsoft and NVIDIA that uses multi-plane networks, packet spraying and SRv6 source routing to keep large GPU training jobs running through failures.How OpenAI delivers low-latency voice AI at scaleurn:rss-system-design:d6581b92c82026-05-03T12:00:00ZOpenAIOpenAI describes splitting its WebRTC stack into a lightweight Go UDP relay and a stateful transceiver, routing packets via ICE ufrag hints and a global relay fleet for voice AI.House of Coroutines: How we Debugged 3M Unbounded Coroutines in Agoda’s Kotlin Serviceurn:rss-system-design:6612b7fbc02026-04-29T12:00:00ZAgodaAgoda describes debugging review-api, a Kotlin service, with Pyroscope, finding hot-path deserialization, about 3 million unbounded coroutines and a Koin thread-safety bug, cutting memory from about 12 GB to under 2 GB.An open-source spec for orchestration: Symphonyurn:rss-system-design:29e3d132682026-04-26T12:00:00ZOpenAIOpenAI describes Symphony, an open-source spec and Elixir reference implementation that turns a Linear board into a control plane for Codex agents, which raised landed PRs by 500% on some teams.Speeding up agentic workflows with WebSockets in the Responses APIurn:rss-system-design:5278705df02026-04-22T12:00:00ZOpenAIOpenAI explains how a WebSocket mode with a connection-scoped in-memory cache made Responses API agent loops 40% faster end-to-end and let GPT‑5.3‑Codex‑Spark reach its 1,000 TPS target.Say What You See: Building Multimodal Content Signals at Scaleurn:rss-system-design:431ac8a8fd2026-03-31T12:00:00ZAgodaAgoda describes a topic-based system mapping 69 image tags and 45 review tags to shared topics, built as PySpark jobs on Kubeflow and served from Couchbase across 700M+ images.From model to agent: Equipping the Responses API with a computer environmenturn:rss-system-design:d4732a41e52026-03-11T12:00:00ZOpenAIOpenAI describes how the Responses API, a shell tool, hosted containers, an egress proxy with secret injection, versioned skills and native compaction give agents a computer environment for long-running workflows.Mobile Release Best Practices: What We Learned While Scaling Across Appsurn:rss-system-design:9dc55c5eb62026-03-03T12:00:00ZGojekGojek describes how it made weekly mobile releases for its Driver, Consumer and Merchant apps more predictable through a release portal, release managers, release trains, automation and feature flags.Beyond rate limits: scaling access to Codex and Soraurn:rss-system-design:956f105c952026-02-13T12:00:00ZOpenAIOpenAI describes the in-house access engine behind Codex and Sora that combines rate limits and credits per request, using idempotent asynchronous debits and atomic balance updates for auditable billing.Harness engineering: leveraging Codex in an agent-first worldurn:rss-system-design:989e332f832026-02-11T12:00:00ZOpenAIAn OpenAI team built an internal product of about a million lines written entirely by Codex over five months, describing repository knowledge structure, agent tooling, enforced architecture and automated cleanup.Unlocking the Codex harness: how we built the App Serverurn:rss-system-design:a45c7ee6fc2026-02-04T12:00:00ZOpenAIOpenAI describes the Codex App Server, a bidirectional JSON-RPC protocol and process that exposes the Codex harness to the CLI, VS Code, Codex Web, the macOS app and partners like Xcode and JetBrains.Inside OpenAI’s in-house data agenturn:rss-system-design:efcfc2c5312026-01-29T12:00:00ZOpenAIOpenAI describes its internal GPT-5.2 data agent for a platform with 3.5k users, 600 petabytes and 70k datasets, covering layered context, RAG over embedded metadata, memory, evals and pass-through permissions.Unrolling the Codex agent loopurn:rss-system-design:5d4a0e56b42026-01-23T12:00:00ZOpenAIOpenAI explains how the Codex CLI agent loop builds prompts for the Responses API, keeps requests stateless for Zero Data Retention, preserves prompt-cache prefixes and compacts conversations via the /responses/compact endpoint.Scaling PostgreSQL to power 800 million ChatGPT usersurn:rss-system-design:0e9e0157f32026-01-22T12:00:00ZOpenAIOpenAI explains how one Azure PostgreSQL primary with nearly 50 read replicas serves millions of queries per second for ChatGPT, using read offload, PgBouncer, cache locking and rate limiting.How We Used Codex to Ship Sora for Android in 28 Daysurn:rss-system-design:6009bfe89b2025-12-11T12:00:00ZOpenAIFour OpenAI engineers used Codex to build the Sora Android app in 28 days, with hand-built architecture, planning before coding, parallel sessions and translation from the iOS code.How we built OWL, the new architecture behind our ChatGPT-based browser, Atlasurn:rss-system-design:b5d0db5c032025-10-29T12:00:00ZOpenAIOpenAI describes OWL, its layer that runs the Chromium browser process outside the Atlas app over Mojo IPC, covering compositing, input forwarding and isolated agent sessions.Part II: Challenges in Building a Real-Time Feature Engineering Systemurn:rss-system-design:3db4af4c582025-09-07T12:00:00ZGojekGojek outlines six challenges in building a real-time feature engineering system for ML models, including misgauged features, OLAP versus OLTP feature disparity, non-reusable features, dynamic Kafka data and feature monitoring.Part-I: Proactive Risk Detection in Fintech: From Data to Real-Time Model Deploymenturn:rss-system-design:772f840e6f2025-09-02T12:00:00ZGojekGojek explains moving from rule-based to model-based proactive risk detection and the steps for deploying real-time risk models, including closest-event triggering, a roughly 2-second latency target and feature categorization.Introducing xkafka — Kafka, but Simpler (for Go)urn:rss-system-design:ff20f2eb442025-06-30T12:00:00ZGojekGoTo introduces xkafka, a Go library giving Kafka HTTP-like Message, Handler and Middleware abstractions, streaming and batch consumers, configurable concurrency, offset handling for delivery guarantees and layered error handling.GoSage: How We Detect Fraud Syndicates at Gojek with Graph Neural Networksurn:rss-system-design:b73d004f1b2024-12-04T12:00:00ZGojekGojek describes GoSage, a graph neural network with hierarchical attention built on PyTorch Geometric that classifies entities in a customer and merchant interaction graph to detect collusion fraud.How Gojek Allocates Personalised Vouchers At Scaleurn:rss-system-design:b7b3c2c0f52024-07-17T12:00:00ZGojekGojek describes its voucher allocation engine, which uses deep-learning causal inference to predict uplift and cost per customer and a knapsack optimiser under budget, with dbt, elementary and Hydra.Evolution of Bid Notifications to Courierurn:rss-system-design:527efb52b02023-12-08T12:00:00ZGojekGojek describes moving driver bid notifications from a vendor to Pusher with Courier, FCM and APNS, rolled out by city tier, reaching a 99.9%+ delivery rate and about 22% lower P99 bid acceptance latency.Courier: Reimagining How We Send Push Notificationsurn:rss-system-design:8823b6149d2023-10-23T12:00:00ZGojekGojek describes moving its Pusher service, which sends about 100 million push notifications daily, onto its MQTT-based Courier with QoS 1, deduplication and FCM/APNs fallback, raising delivery rates to 95%+ and 99%+.Modularizing Our Driver App For Android: Benefits & way ahead (Part-3)urn:rss-system-design:d4d2d1151a2023-09-27T12:00:00ZGojekGojek describes the results of modularizing its Android driver app, with CI builds dropping from about 30 to a little over 20 minutes and app-module classes falling from 78% to 68%.Modularizing Our Driver App For Android: The Process (Part-2)urn:rss-system-design:7f3c0095032023-06-22T12:00:00ZGojekGojek describes how it split its Android driver app into core, utilities, features and app groups with api and implementation modules, and scripts that extracted 14 modules in 2 sprints.Modularizing Our Driver App For Android: The Need (Part-1)urn:rss-system-design:6f797937972023-05-08T12:00:00ZGojekGojek explains why it modularized its single-module Android driver app after team growth led to 30-minute CI builds, merge conflicts and a need for per-feature ownership, crash and binary size metrics.Building a multi-pipeline CI setup for running automated tests in GitLaburn:rss-system-design:cc29871b422023-02-23T12:00:00ZGojekGojek explains how it combined an application's GitLab development pipeline with a functional test pipeline from a separate test project, using reusable CI templates and the needs, allow_failure and resource_group keywords.How We Estimate Food Debarkation Time With ‘Tensoba’urn:rss-system-design:38383761d62022-01-31T12:00:00ZGojekExplains how GoFood built Tensoba, XGBoost models predicting T1, T2 and T3 delivery phases for Discovery and Checkout ETAs, cutting MAPE from 31% to 23%, served via Merlin, Clockwork, Redis and Feast.How Gojek Uses NLP to Name Pickup Locations at Scaleurn:rss-system-design:b6af53d1432020-04-30T12:00:00ZGojekGojek describes CartoBERT, an ALBERT-based model pre-trained on about 200 million booking text sentences and fine-tuned for token classification, which raised pickup point name accuracy by more than 25% to about 93%.