Etsy engineering blog

Etsy explains how it measures and reduces the environmental impact of cloud computing and AI. The post covers Cloud Jewels energy estimates, Google Cloud emissions data, model right-sizing, internal sustainability tooling, emissions targets, and token-level impact accounting.

Etsy describes how Kafka Skills use coding agents, deterministic scaffolding, validation tools, and local test orchestration to accelerate machine-learning streaming pipeline development. The post also covers discoverability telemetry and least-privilege controls for safely allowing agents to generate code, open pull requests, and run workflows.

Etsy describes a contrastive reinforcement-learning approach that fine-tunes Qwen3-VL-8B to generate product summaries emphasizing details that distinguish engaged listings from alternatives. Using triplet engagement data, a frozen two-tower retriever, and GRPO, the method improved a semantic-relevance model’s macro F1 score by 8.7%.

Etsy explains how it improved Ads Search ranking with a Multigate Mixture of Experts architecture and add-to-cart auxiliary task. The post details expert utilization, temperature scaling, negative transfer from noisy signals, calibration, and measurable offline and serving gains.

Etsy describes its five-year migration from custom MySQL shard routing to Vitess vindexes across roughly 1,000 shards and 425 TB of data. The post explains the hybrid SQLite/hash vindex design, incremental rollout strategy, transaction constraints, scatter-query safeguards, and resulting improvements in scalability and reliability.

Etsy describes a human-grounded LLM framework for evaluating and improving search relevance across millions of query-listing pairs. A cascaded pipeline of an LLM annotator, fine-tuned teacher, and lightweight two-tower student model supports offline measurement and real-time ranking improvements with under 10 ms of added latency.

Etsy explains how it evolved CUPED into CUPAC, using machine-learning predictions from more than 100 pre-experiment features to reduce metric variance in online experiments. The LightGBM-based pipeline, deployed with BigQuery, Dataflow, Spark, and Airflow, achieved 27% average variance reduction and shortened experiments by nearly three days.

Etsy explains how it used the Speculation Rules API to prefetch product pages on search-result hover, improving TTFB, FCP, LCP, and other metrics by roughly 20–24%. The post details caching limits, eagerness settings, redirects, cookies, shadow DOM interactions, and analytics changes required for a reliable rollout.

Etsy describes an LLM-powered pipeline that extracts structured product attributes from listing text and images across millions of unique items. The post covers silver-label evaluation, context engineering, parallel inference, validation, monitoring, and the resulting increase in complete attribute coverage from 31% to 91%.

Etsy explains how it replaced hard-coded experiment segmentations with a self-service, configuration-driven system. The design uses validated SQL definitions, automated deployment, and dynamically generated parallel pipeline tasks, tripling supported segmentations while improving failure isolation and speeding experimentation insights.

Etsy describes how it uses LLMs to turn buyer activity into structured, privacy-conscious interest profiles for personalized search. The post details scaling and cost optimizations that cut profile generation from 21 days to 3 days for 10 million users and reduced estimated costs by 94% per million users.

The post examines two Etsy question-answering pilots using prompt engineering, embedding-based retrieval, and LLMs for employee onboarding and seller forums. It reports accuracy results, analyzes hallucination failure modes, and shows how requesting uncertainty, reasoning, and source snippets can improve answer verification while revealing important limitations.

Etsy details the architecture behind its tax calculation infrastructure, including Vertex’s Quotation and DistributeTaxRequest APIs, tax-category mapping, and reporting workflows. The post explains how separating workloads across instances, load balancing by identifiers, and shadow traffic testing helped scale checkout traffic without degrading performance.

Etsy details its gradual adoption of Jetpack Compose, from training and design-system components to migrating production screens alongside XML Views. The team reports faster rendering, improved developer experience, stronger testability, and lessons around previews, unidirectional data flow, lazy lists, and interoperability.

Etsy explains how it uses supervised multimodal learning to detect marketplace policy violations at scale. The system combines BERT/ALBERT text encoders, EfficientNet image features, hard-negative sampling, focal loss, time-based evaluation, and champion-challenger A/B testing to improve detection while limiting false positives.

Etsy explains how its FinOps team forecasts, monitors, and optimizes Google Cloud spending using Cost Per Visit, regression analysis, dashboards, alerts, and cross-functional reviews. The post details practical savings initiatives including GCS lifecycle management, CPU platform migration, automated ML deployment configuration, and network compression.

Etsy explains how it builds efficient visual representations for image search and recommendations using EfficientNet, EfficientFormer, multitask learning, and generative evaluation data. The post details fine-tuning, retrieval-based evaluation, latency and memory optimizations, and measured improvements from production experiments.

Etsy describes Macramé, a reactive Android architecture that replaced a 4,000-line listing screen with immutable state, StateFlow, event handlers, and RecyclerView-based UI models. The rewrite reduced time to first content by 18%, increased business metrics, and raised business-logic test coverage to 76%, while revealing trade-offs such as boilerplate and harder debugging.

Etsy engineers explain how they built the Deals tab in SwiftUI and Tuist within a legacy Objective-C and UIKit codebase. The post covers modular architecture, typed API responses, preview-driven development, and interoperability patterns that decoupled delivery from backend work while meeting tight product deadlines.

Etsy explains how it expanded its AR shopping experience by moving dimension parsing from its iOS app to a cached backend service. The post details robust multilingual regex parsing, 3D measurement boxes built with SceneKit, and LiDAR-based environment occlusion for realistic item visualization.