This week, Cloudflare saves 100TB of RAM with math and Rust, while pgrust pushes Postgres analytics into a faster gear. We explore how to coordinate agents without losing track of their work—and why better planning doesn’t always need a plan. Plus, local decision models and an unexpected language model built with gzip.
— Mahdi Yusuf (@myusuf3) or LinkedIn
👋🏾 You are reading Architecture Notes - Your Sunday newsletter, which curates best system design and architecture news from around the web. We would appreciate you sharing it with like-minded people.
Saving Another 100TB of RAM with Math (and Rust)
Cloudflare reclaimed more than 100TB of RAM by rethinking the consistent-hashing structures in its Pingora Backend Router. The team combined a more compact Rust representation with a mathematical analysis of how many hash points it actually needed, showing how small data-structure decisions can become enormous infrastructure costs at scale.
13 Multi-Agent Orchestration Patterns with a Reactive Reducer
Brad Murry maps 13 multi-agent orchestration patterns—from pipelines and fan-out to supervisors and review loops—onto a reactive reducer with typed shared state and explicit transitions. Each pattern comes with use cases and failure modes, showing how declared control flow, durable state, and retry budgets can make agent workflows easier to inspect and recover when a run goes wrong.
Level Up Your DevOps Game with iximiuz Labs
Preparing for a DevOps interview? iximiuz Labs is the ultimate resource to help you succeed. Dive into in-depth courses on Linux, networking, containers, and Kubernetes, paired with hands-on challenges designed to sharpen your skills.
Whether you’re just starting out or looking to refine your expertise, iximiuz Labs gives you the tools and confidence to stand out in any DevOps interview. Ready to take the next step? Start practicing today at iximiuz Labs.
Rebuilding Postgres for Faster Analytics: Batching, Operator Fusion, and SIMD
The pgrust team reports a 300× speedup over Postgres on ClickBench, with roughly 10× coming from query-engine changes. This walkthrough explains how batching, operator fusion, and SIMD cut CPU and memory overhead, making it a useful guide to the costs hidden inside a database execution engine; the headline result applies to the reported benchmark.
Fatih Arslan organizes agent work as Markdown plans in a shared Git repository, with folders for drafts, queued work, active tasks, and completed projects. Long-running coordinators maintain the index, giving humans and agents a common view of what needs doing and a history of decisions—a practical answer to the coordination problems that appear when several agents work at once.
Ayman Nadeem reflects on building Nuanced around persistent plans, only to find that long AI-generated specs added friction without improving understanding. The argument isn’t against planning, but against separating it from building: better workflows interleave action, inspection, and adjustment. The harder problem remains helping humans follow architectural decisions as multiple agents change a system faster than anyone can review every diff.
Projects
Google’s ax is a declarative orchestrator for agent workloads, with Kubernetes-style manifests for tasks, workspaces, network gateways, and models. It runs on Agent Substrate and supports sandboxed execution, outbound traffic controls, and task suspension and resumption, offering a look at how infrastructure is adapting to stateful agents; the project is still evolving and warns of breaking changes.
Agent Substrate runs stateful agents in isolated sandboxes and multiplexes them across a smaller pool of workers, preserving memory and filesystem state through suspend and resume. Built on Kubernetes with support for sandbox technologies including gVisor and microVMs, it targets the cost of keeping mostly idle agents ready to work; the project is in early development and is not yet production-ready.
Ollaya is an open-source runtime for running decision models locally, turning text or JSON into typed choices, scores, and yes/no answers with probabilities instead of generated prose. It offers a TypeSafe-compatible API, a CLI, and desktop apps, making it useful for classification, routing, and filtering without sending application data to a hosted model. The project is currently in beta.
Nathan Barry turns a familiar compressor into a text generator by searching for continuations that compress best. Using DEFLATE through Python’s zlib and beam search, the experiment produces recognizably Shakespeare-like fragments without neural networks or learned weights. The output is far from coherent, but it makes the connection between compression and prediction tangible—and shows why looking several bytes ahead beats choosing one byte at a time.