Wire Formats Compared: Protobuf, Cap'n Proto & FlatBuffers
Choosing a wire format at scale: JSON, Protocol Buffers, Cap'n Proto, FlatBuffers, Avro and MessagePack compared on payload size, encode/decode cost, and schema evolution.
Mechanical Snail
Methodically crawling through caches, concurrency, and chaos...at a pace slower than drying paint
Browse by topic
Posts
AI Trading System Architecture for Financial Markets
The architecture of production AI trading systems: data pipelines, feature engineering, prediction models, risk management and backtesting under adversarial market dynamics.
Columnar Storage in Go: Fast Financial Aggregation
Building a columnar storage engine in Go: cache-friendly memory layout, fixed-point decimal arithmetic for financial precision, SIMD-friendly aggregations, and honest benchmarks.
Building an MCP Server in Go: Graph Tools & Access Control
Building a Model Context Protocol server in Go: JSON-RPC tool definitions, social graph traversal, role-based access control with inheritance, and idempotent moderation operations.
Neural Networks from Scratch in Rust
Neural networks from first principles in Rust: manual backpropagation, cache-friendly matrix operations, SIMD utilization, and where inference time is actually spent.
Caching Strategies: Cache-Aside, Write-Through & Eviction
A survey of production caching strategies: cache-aside, read-through, write-through and write-back, LRU and friends as eviction policies, and the consistency trade-offs each one …
ML Risk Models for Nordic Power Futures & GoO Portfolios
Machine learning risk models for Nordic power futures and Guarantees of Origin portfolios: why historical VaR underestimates fat tails and how ML captures non-linear weather …
Low-Latency Trading Systems in Rust
Engineering microsecond-latency trading systems in Rust: lock-free data structures, cache-line optimization, kernel bypass networking, and hot paths that never allocate.
LLMs & Foundation Models in Recommenders (Part 6 of 6)
Part 6 of 6: foundation models, LLMs and multimodal understanding in recommendation — what changes once the ranker can actually read the content it is ranking.
Feed Ranking Architecture & Operations (Part 5 of 6)
Part 5 of 6: a reference feed-ranking architecture — request flow, latency budgets, ads blending, two-tower retrieval and ranking implementations, and production monitoring.
Recommender Ethics, Fairness & Governance (Part 4 of 6)
Part 4 of 6: amplification harms, filter bubbles, fairness metrics and the governance controls that keep large-scale recommendation systems accountable.
Recommendation Systems in Production (Part 3 of 6)
Part 3 of 6: running recommenders in production — offline ranking metrics, A/B testing statistics, distributed training, model serving and compression, cold start, and drift.
Ranking & Re-ranking Recommendations (Part 2 of 6)
Part 2 of 6: ranking and re-ranking — pointwise, pairwise and listwise losses, Deep & Cross and Wide & Deep architectures, multi-task learning, diversity via MMR and DPP, and …
Recommendation System Architecture & Features (Part 1 of 6)
Part 1 of 6: the multi-stage architecture of a production recommendation system — candidate retrieval, embedding-based nearest neighbour search, and feature engineering at scale.
Social Media Platform Architecture at Scale
Systems architecture for an Instagram-scale social platform: fan-out on write versus read, graph sharding, feed caching, and absorbing celebrity-tier traffic spikes.
Building a Time Series Database in Go
Inside a high-performance time series database in Go: LSM-style storage, delta-of-delta and Gorilla compression, inverted label indexes, and range query execution.
The Actor Model: Concurrency Through Message Passing
How the Actor Model replaces shared memory with message passing: its origins in Erlang, supervision trees, mailbox semantics, and where actors fit in modern distributed systems.
Vector Optimization: SIMD, Cache Lines & Memory Bandwidth
Why vector code is memory-bound, not compute-bound: cache-line behaviour, array-of-structs versus struct-of-arrays layout, AVX vectorization, and the bandwidth ceiling.
Building a Derivatives Pricing DSL in Rust
Designing a type-safe DSL in Rust for valuing futures and forwards: the pricing mathematics, a typed expression language, and an evaluation engine for portfolios of thousands of …