---
title: "Architecture"
description: "834 posts about Architecture, summarised, each linking to the original."
---

# Architecture
> 834 posts about Architecture, summarised, each linking to the original.

## Articles

### [Announcing chDB 4: write Pandas, run ClickHouse, now on Hex](https://yomu.fyi/post/announcing-chdb-4-write-pandas-run-clickhouse-now-on-hex.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: Ryadh Dahimene, Auxten Wang
- Published: Mar 11, 2026
- Type: Announcement

chDB 4 introduces Data Store, a native component that allows users to write Pandas-like code that executes on the ClickHouse OLAP SQL engine. Standard Pandas workflows suffer from eager execution, excessive memory materialization, and lack of automatic query optimizations. To solve this, Data Store implements a four-layer architecture that lazily records operations as a chain of LazyOp objects and compiles them into optimized ClickHouse SQL plans upon evaluation triggers. When encountering operations unsupported by ClickHouse, the QueryPlanner splits the pipeline into segments, routing tasks between ClickHouse and Pandas while using Python's memoryview to minimize data copying overhead. This release also integrates natively into Hex notebooks.


### [Keep the Tokens Flowing: Lessons from 16 Open-Source RL Libraries](https://yomu.fyi/post/keep-the-tokens-flowing-lessons-from-16-open-source-rl-libraries.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Amine Dirhoussi, Quentin Gallouédec, Kashif Rasul, Lewis Tunstall, Edward Beeching, Albert Villanova del Moral, Nouamane Tazi, Leandro von Werra, Sergio Paniego
- Published: Mar 10, 2026
- Type: Benchmark

At-scale reinforcement-learning training is constrained by long, variable data-generation workloads that leave training GPUs idle, especially for reasoning and agentic models. The surveyed architecture separates inference and training onto distinct GPU pools, connects them through a rollout buffer, and synchronizes weights asynchronously so both stages can proceed concurrently. Sixteen open-source libraries are compared across orchestration, buffering, weight synchronization, staleness, partial rollouts, LoRA, and distributed backends; Ray appears in 8 of 16 orchestration implementations, while NCCL broadcast is the default weight-transfer method. The proposed TRL async trainer favors a lightweight bounded queue with per-token model\_version metadata, packed NCCL transfers, and experimental support for prefix-resume or abort-and-retry handling of rollouts interrupted by weight updates.


### [What Replo learned optimizing 100+ billion events in ClickHouse](https://yomu.fyi/post/what-replo-learned-optimizing-100-billion-events-in-clickhouse.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: ClickHouse
- Published: Mar 9, 2026
- Type: Problem & solution

Replo built an analytics pipeline on ClickHouse to serve more than 4,000 Shopify merchants processing 3,000 to 5,000 events per second. The initial single-table architecture recalculated session metrics on the fly, leading to poor query efficiency as volume scaled. To speed up dashboard queries, the team implemented a mark-and-unmark pattern with SummingMergeTree and refreshable materialized views, but real-time deduplication and fractional attribution caused runaway query backlog loops. Replo resolved the issue by limiting real-time recomputations to a specialized table holding only the last 40 minutes of purchase-related events. This architecture allowed processing over 100 billion records while keeping query times fast and ingestion lag to approximately one minute.


### [Ulysses Sequence Parallelism: Training with Million-Token Contexts](https://yomu.fyi/post/ulysses-sequence-parallelism-training-with-million-token-contexts.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Kashif Rasul, Stas Bekman
- Published: Mar 9, 2026
- Type: Explainer

Ulysses Sequence Parallelism addresses the memory and compute pressure of training language models on sequences extending beyond tens of thousands of tokens, where attention remains quadratic in sequence length even with FlashAttention. It shards tokens across GPUs, projects local queries, keys, and values, then uses two all-to-all operations to exchange data so each GPU computes a subset of attention heads before restoring sequence-sharded outputs. The post describes integrations with Accelerate, the Transformers Trainer, and TRL's SFTTrainer, including variable-length handling and weighted loss aggregation for uneven valid-token counts. In reported four-GPU tests, SP=4 reduced per-GPU memory sufficiently for 96K-token training on H100 80GB GPUs, reached 13,396 tokens per second at 64K, and matched non-SP canonical token-normalized loss within 0.000005 maximum absolute difference.


### [LeRobot v0.5.0: Scaling Every Dimension](https://yomu.fyi/post/lerobot-v0-5-0-scaling-every-dimension.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Steven Palma, Pepijn Kooijmans, Jade Choghari, Caroline Pascal, Khalil Meftah, Martino Russi, Nicolas Rabault, Michel Aractingi, Virgile BATTO, Thomas Wolf
- Published: Mar 9, 2026
- Type: Announcement

LeRobot v0.5.0 is a major release that expands supported hardware, policies, datasets, simulation, and its software foundation, following more than 200 merged PRs and more than 50 new contributors since v0.4.0. It adds full Unitree G1 humanoid support for locomotion, manipulation, teleoperation, and whole-body control, alongside OpenArm, Earth Rover, CAN-bus motor controllers, and other integrations. The policy additions include Pi0-FAST's autoregressive VLAs with FAST action tokenization, Real-Time Chunking for continuously blended predictions, Wall-X, X-VLA, SARM, and PEFT fine-tuning. For data and simulation workflows, streaming video encoding removes waits between recording episodes, while EnvHub loads environments from the Hugging Face Hub and IsaacLab-Arena supplies GPU-accelerated simulation. The codebase now requires Python 3.12, uses Transformers v5, supports installable third-party policy plugins, and adds remote Rerun visualization plus installation and documentation improvements.


### [Designing flexible payment flows with Checkout Session](https://yomu.fyi/post/designing-flexible-payment-flows-with-checkout-session.md)
- Company: [Stripe](https://yomu.fyi/company/stripe.md)
- Author: Allison Farris
- Published: Mar 6, 2026
- Type: Tutorial

The post addresses the architectural challenge of taking a digital product business global while supporting localized tax, currencies, one-time purchases, subscriptions, and a branded embedded checkout. It recommends Checkout Sessions as the orchestration layer, with Stripe creating and managing the underlying Payment Intent while declarative parameters handle features such as Stripe Tax, Adaptive Pricing, subscriptions, and promotional codes. The implementation creates a session with ui\_mode, mode, automatic\_tax, adaptive\_pricing, line items, and optional payment\_intent\_data, then passes the client secret to Stripe Elements for the Payment Element and Currency Selector. This approach supports localized payment methods and currency conversion without a backend rewrite, while direct Payment Intents remain appropriate when an application needs full control over carts, discounts, taxes, or its own checkout abstraction.


### [Bringing Robotics AI to Embedded Platforms: Dataset Recording, VLA Fine‑Tuning, and On‑Device Optimizations](https://yomu.fyi/post/bringing-robotics-ai-to-embedded-platforms-dataset-recording-vla-fine.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Gaetan Bahl, Enzo Ruedas, Tess Boivin
- Published: Mar 5, 2026
- Type: Tutorial

Deploying Vision–Language–Action policies on embedded robots is constrained by compute, memory, power, and real-time control requirements; synchronous inference can leave the arm idle and cause oscillatory corrections. NXP presents dataset-recording practices for a tea-bag placement task, including fixed cameras, controlled lighting, a gripper-mounted view, calibration backups, and diverse training and validation splits with recovery episodes. The example uses 120 episodes, three 640×480 cameras at 30 fps, batch size 8, and checkpoint selection based on training and validation success rather than loss alone. For deployment, the VLA is decomposed into vision, LLM backbone, and action-expert stages, with selective quantization and asynchronous scheduling constrained by inference time being shorter than action execution. On the i.MX 95, optimized ACT latency reached 0.32 seconds with 0.89 global accuracy, while ONNX FP32 ACT reached 0.96 global accuracy at 2.86 seconds.


### [Introducing Modular Diffusers - Composable Building Blocks for Diffusion Pipelines](https://yomu.fyi/post/introducing-modular-diffusers-composable-building-blocks-for-diffusion.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: YiYi Xu, Alvaro Somoza, Dhruv Nair, Sayak Paul
- Published: Mar 5, 2026
- Type: Announcement

Modular Diffusers introduces a composable alternative to Diffusers’ existing DiffusionPipeline for building diffusion workflows from reusable blocks. A ModularPipeline can load a workflow definition, load model components separately with settings such as bfloat16, and preserve the familiar inference API, while blocks for text encoding, image encoding, denoising, and decoding can be inspected, run independently, or rearranged. Custom Python blocks declare components, inputs, outputs, and computation logic; the example adds a Depth Anything V2 depth processor to a Qwen Image ControlNet workflow, with data flowing between compatible blocks. Blocks and complete pipelines can be published to the Hub, and Mellon can generate node interfaces from their consistent definitions. The post presents the system as an early, flexible direction and says Mellon is not ready for production use.


### [Introducing ClickStack embedded in ClickHouse](https://yomu.fyi/post/introducing-clickstack-embedded-in-clickhouse.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Author: The ClickStack Team
- Published: Mar 5, 2026
- Type: Announcement

ClickHouse 26.2 introduces an embedded distribution of the ClickStack observability UI packaged directly inside the ClickHouse binary. Designed for local exploration and instance inspection, this distribution allows developers to visualize internal query logs, traces, and system metrics without deploying external services. To preserve build integrity and keep the added binary footprint around 4.1 to 4.2 MB, the team avoided a Node.js dependency by embedding a pre-built static Next.js application using CMake-generated C++ structs. Supporting components such as MongoDB, Express, and the Pyodide WebAssembly runtime were omitted, disabling alerting and persistent dashboards in favor of simplicity. Users can access the interface directly over HTTP to monitor local databases and query performance out of the box.


### [Towards Model-based Verification of a Key-Value Storage Engine](https://yomu.fyi/post/towards-model-based-verification-of-a-key-value-storage-engine.md)
- Company: [MongoDB](https://yomu.fyi/company/mongodb.md)
- Author: Will Schultz, Murat Demirbas
- Published: Feb 27, 2026
- Type: Problem & solution

MongoDB developers formalized the interface boundary between their distributed transactions protocol and the underlying WiredTiger key-value storage engine using a compositional TLA+ specification. This modular model established an abstract contract to verify that the storage engine's concrete timestamp-based operations matched the semantics expected by the distributed transaction layer. To check conformance, a modified version of the TLC model checker generates a complete graph of reachable states for finite parameters and computes path coverings. Each identified path is translated into an automated test case consisting of a sequence of storage engine API calls. For a finite model with two keys and two transactions, this workflow generated 87,143 test cases that executed against WiredTiger in approximately 40 minutes.


### [Mixture of Experts (MoEs) in Transformers](https://yomu.fyi/post/mixture-of-experts-moes-in-transformers.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Aritra Roy Gosthipaty, Pedro Cuenca, merve, Ilyas Moutawwakil, Arthur Zucker, Sergio Paniego, Pablo Montalvo
- Published: Feb 26, 2026
- Type: Explainer

Mixture of Experts (MoEs) in Transformers let a Transformer retain its backbone while replacing selected feed-forward layers with routed expert sub-networks, allowing total model capacity to exceed the active parameters used per token. The post explains why sparse models can reduce inference cost and memory pressure, then focuses on making them first-class in Transformers. Its weight-loading refactor converts independently serialized expert tensors into packed runtime layouts through composable WeightConverter operations and lazy materialization. The Experts Backend supplies eager, batched\_mm, and grouped\_mm execution strategies, while expert parallelism shards experts across devices and aggregates partial outputs. Training support uses grouped GEMM and custom Triton plus LoRA kernels, with reported improvements of approximately 12× speed, over 35% VRAM reduction, and roughly 6× longer context.


### [Innovating with MongoDB | Customer Successes, February 2026](https://yomu.fyi/post/innovating-with-mongodb-customer-successes-february-2026.md)
- Company: [MongoDB](https://yomu.fyi/company/mongodb.md)
- Author: Katie Palmer
- Published: Feb 25, 2026
- Type: Problem & solution

Modern artificial intelligence startups frequently face operational bottlenecks when deploying autonomous agents and dynamic interfaces on rigid relational schemas or restricted database engines. To address these limitations, platforms including Modelence, Thesys, Emergent Labs, and Heidi integrated MongoDB Atlas as their operational data foundation. Emergent Labs eliminated recursive schema migration failures previously encountered with PostgreSQL by adopting document structures that match native JSON generation, supporting millions of deployed applications. Meanwhile, Heidi migrated from Amazon DocumentDB to utilize MongoDB Vector Search for retrieval-augmented generation pipelines under a single unified API without requiring external search databases. These architectural transitions streamlined continuous agent iterations, eliminated operational scaling downtime, and accelerated deployment cycles across diverse production environments.


### [We proactively fixed ~100 security issues in 6 days with 0 humans](https://yomu.fyi/post/we-proactively-fixed-100-security-issues-in-6-days-with-0-humans.md)
- Company: [Ramp](https://yomu.fyi/company/ramp.md)
- Author: Eli Block
- Published: Feb 20, 2026
- Type: Tutorial

Ramp Security Engineering describes a six-day experiment in which home-grown agents found, validated, and fixed nearly 100 latent backend security issues, including some high-severity findings, without human involvement until pull-request review. The issues had not been uncovered by penetration testing, bug bounty testing, static analysis, or trials of 10+ code-scanning vendors, and all were patched within a week of discovery. The pipeline used specialized vulnerability detectors, adversarial manager agents that rejected 40% of initial proposals in sample testing, integration tests for validation, and an internal coding agent that generated patches. Live interactive validation struggled with complex preconditions, so test-based validation became the main approach; a human team member still reviewed and landed each pull request.


### [Scaling Localization with AI at Lyft](https://yomu.fyi/post/scaling-localization-with-ai-at-lyft.md)
- Company: [Lyft](https://yomu.fyi/company/lyft.md)
- Author: Stefan Zier
- Published: Feb 19, 2026
- Type: Problem & solution

Lyft re-architected its batch localization pipeline to combine artificial intelligence with professional linguist oversight. The previous workflow relied exclusively on human translation, which resulted in multi-day turnarounds and linear cost scaling that hindered rapid market expansions. Under the new architecture, a dual-path system generates three candidate translations using a fast generation model, evaluates them against defined rubrics using a reasoning-focused model, and immediately releases approved drafts. Linguists review the drafts asynchronously in the Smartling translation management system to establish the final system of record. This pipeline reduced delivery latency from days to minutes while achieving a state where 95% of machine-generated translations require no significant changes during human review.


### [Building a mental model for Stripe payments](https://yomu.fyi/post/building-a-mental-model-for-stripe-payments.md)
- Company: [Stripe](https://yomu.fyi/company/stripe.md)
- Author: Allison Farris
- Published: Feb 19, 2026
- Type: Tutorial

An online bookshop checkout must coordinate card networks, issuing banks, fraud checks, regulatory requirements, currencies, and varied payment methods while completing in roughly 2–3 seconds. Stripe abstracts these concerns through a unified API, with the PaymentIntent representing the payment lifecycle from amount and currency collection through tokenization, authorization, capture, and settlement. The Payment Element and client-side tokenization keep sensitive card details away from the shop’s servers, while Stripe may perform risk checks and additional 3D Secure authentication. Because payments can change state minutes, hours, or days later through refunds and disputes, webhook events should be treated as the source of truth for synchronizing order and downstream systems without polling.


### [Minions: Stripe’s one-shot, end-to-end coding agents—Part 2](https://yomu.fyi/post/minions-stripe-s-one-shot-end-to-end-coding-agents-part-2.md)
- Company: [Stripe](https://yomu.fyi/company/stripe.md)
- Author: Alistair Gray
- Published: Feb 19, 2026
- Type: Explainer

Stripe’s minions are an unattended agentic coding flow whose pull requests are human-reviewed but contain no human-written code; more than 1,300 merge weekly. The system runs agents in AWS EC2-based devboxes that are prewarmed to become usable within about 10 seconds, providing parallel, predictable, and isolated environments. Its custom harness uses code-defined blueprints that combine deterministic nodes, such as linting and pushing changes, with agent loops for tasks such as implementation and fixing CI failures. Minions receive automated feedback through local linting, one full CI iteration, and a possible second repair attempt, while devbox isolation and security controls limit destructive actions; Stripe presents the approach as an extension of long-running investments in human developer productivity.


### [Custom Kernels for All from Codex and Claude](https://yomu.fyi/post/custom-kernels-for-all-from-codex-and-claude.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: ben burtenshaw, Sayak Paul, Aritra Roy Gosthipaty, shaun smith
- Published: Feb 13, 2026
- Type: Announcement

Writing CUDA kernels that integrate with diffusers and transformers involves GPU-specific optimization, library integration, PyTorch registration, and environment compatibility. It presents a cuda-kernels agent skill that packages architecture guides, integration patterns, kernel templates, troubleshooting material, benchmarks, and Kernel Hub examples for coding agents. Claude and Codex used it to generate RMSNorm, RoPE 3D, GEGLU, and AdaLN kernels for LTX-Video in diffusers, and an RMSNorm kernel for Qwen3-8B in transformers, including CUDA source, PyTorch bindings, build configuration, and benchmark scripts. On H100 80GB HBM3, the LTX-Video RMSNorm tests averaged 1.88x speedup and the optimized pipeline with torch.compile reached 1.43x the listed baseline, while Qwen3-8B RMSNorm averaged 1.94x and reached 2.47x at 8192 tokens. The skill supports packaging kernels for precompiled distribution through the Hub, where get\_kernel selects a matching binary across Python, PyTorch, and CUDA variants.


### [Transformers.js v4: Now Available on NPM!](https://yomu.fyi/post/transformers-js-v4-now-available-on-npm.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Joshua, Nico Martin
- Published: Feb 9, 2026
- Type: Announcement

Transformers.js v4 is now available on NPM, marking a major release after a year of development. It introduces a C++-rewritten WebGPU runtime tested across about 200 supported model architectures, enabling the same JavaScript code to run WebGPU-accelerated models in browsers, Node, Bun, Deno, and desktop applications. Specialized ONNX Runtime Contrib Operators support new export strategies for large language models, and the post reports a roughly fourfold speedup for BERT-based embedding models using MultiHeadAttention. The release also restructures the repository as a pnpm-workspaces monorepo, modularizes model definitions, moves examples to a separate repository, and switches from Webpack to esbuild, reducing build times from two seconds to 200 milliseconds and shrinking transformers.web.js by 53%. New APIs cover model asset inspection, caching, environment fetch controls, logging, and standalone tokenization, while added architectures include Mamba, MLA, and MoE models compatible with WebGPU.


### [Towards self-driving codebases](https://yomu.fyi/post/towards-self-driving-codebases.md)
- Company: [Cursor](https://yomu.fyi/company/cursor.md)
- Author: Wilson Lin
- Published: Feb 5, 2026
- Type: Explainer

The post describes research into scaling long-running autonomous coding agents, using a web browser as a complex benchmark and running thousands of agents continuously for one week. An initial single-agent plan failed because the model lost track, stopped prematurely, and stalled on complex details; shared-state coordination then introduced locking contention, while more structured planner, executor, worker, and judge roles improved ownership and throughput. The team built a Rust-based harness on one large Linux VM, logged agent messages and tool activity for replay and analysis, and iterated through designs that eventually favored a more dynamic system with freshness mechanisms and fewer overloaded roles. The research concludes that architecture, prompts, observability, explicit throughput trade-offs, and anti-fragile recovery matter, reporting linear scaling of token throughput with minimal harness overhead while humans retained taste, judgment, and direction.


### [Edge AI Made Easy: MongoDB and ObjectBox Data Synchronization](https://yomu.fyi/post/edge-ai-made-easy-mongodb-and-objectbox-data-synchronization.md)
- Company: [MongoDB](https://yomu.fyi/company/mongodb.md)
- Author: Puja Roy
- Published: Feb 3, 2026
- Type: Announcement

Modern edge AI deployments require immediate real-time decision-making, local privacy, and offline operation while retaining the analytical scale of central cloud infrastructure. To address these hybrid demands, MongoDB and ObjectBox partnered to release the ObjectBox MongoDB Sync Connector as part of ObjectBox 5.0. ObjectBox operates as a lightweight on-device database with local vector search, multi-language bindings, and high-throughput I/O suited for resource-constrained hardware. In environments such as industrial factory floors and restaurant point-of-sale terminals, devices capture and evaluate operational data onsite without depending on active network connections. The native connector subsequently synchronizes critical operational datasets to MongoDB Atlas for long-term storage, fleet-wide monitoring, and machine learning model retraining.


[Newer posts](https://yomu.fyi/topic/architecture/page/23.md) · [Older posts](https://yomu.fyi/topic/architecture/page/25.md)