Spotify logo

Spotify

Our R&D teams tell the stories behind how we build at Spotify. AI, mobile, web, data science, experimentation, developer tools, design, open source, and more.

63 postsVisit website
Try:
Spotify logo
Spotify

Beyond Winning: Spotify’s Experiments with Learning Framework

Spotify describes how its Confidence experimentation platform evolved from focusing on velocity to emphasizing experiment quality via the Experiments with Learning (EwL) metric. EwL defines a successful experiment as one that is valid and decision-ready (detects wins, regressions, or is a well-powered neutral), and Spotify reports a ~64% learning rate versus a ~12% win rate. The framework helps prioritize testing capacity, improve platform tooling (sample-size calculators, health checks), and guide organizational practices and investments in experimentation.

Spotify logo
Spotify

Incident Report: Spotify Outage on April 16, 2025

Spotify experienced a global outage on April 16, 2025 after a change in Envoy filter ordering triggered a bug that crashed Envoy instances worldwide. Rapid restarts combined with client retry traffic exposed a configuration mismatch (Envoy heap > Kubernetes memory limit), causing continuous pod restarts; mitigation involved increasing perimeter capacity and fixes to the bug and configuration. Spotify plans to improve rollout practices and monitoring.

Spotify logo
Spotify

A Behind-the-Scenes Look at How We Release the Spotify App (Part 1)

Spotify describes its weekly mobile-app release process: trunk-based development, nightly alpha builds, a release branch, manual and automated testing, use of feature flags, staged rollouts via app stores, and active monitoring (crash rates/ANRs) through a Release Manager Dashboard to balance speed and quality.

Spotify logo
Spotify

A Behind-the-Scenes Look at How We Release the Spotify App (Part 1)

Spotify describes its weekly mobile release process for the main iOS and Android apps: trunk-based development with nightly builds and alpha/beta programs, branching a release for final stabilization, manual and automated testing, a Release Manager Dashboard that tracks blocking bugs/crashes/distribution status, and a phased rollout (1% then 100%) controlled via backend feature flags. The piece highlights balancing speed and quality, criteria for submission to app stores, and monitoring/mitigation steps during rollout (crash rates, ANRs, pause and fix workflows).

Spotify logo
Spotify

An Insider’s Tips for Taking the Certified Backstage Associate (CBA) Exam

A Spotify Backstage engineer describes their experience taking the Certified Backstage Associate (CBA) exam and gives practical preparation tips. The post explains exam logistics and proctoring, outlines four main subject areas to study (Backstage development workflow, infrastructure, catalog, and customization), lists what is out of scope, and recommends hands-on practice with Backstage, Yarn, Node, Docker, TypeScript, React, and the Catalog (YAML/Kinds).

Spotify logo
Spotify

An Insider’s Tips for Taking the Certified Backstage Associate (CBA) Exam

A Spotify engineer describes his experience taking the Certified Backstage Associate (CBA) exam, explains why Backstage certification is valuable, gives proctored-exam tips, and outlines the four CBA subject areas (development workflow, infrastructure, the Backstage Catalog, and customization). The post highlights what to study (Backstage CLI, TypeScript, Yarn, Docker, configuration, catalog YAML/Kinds, frontend/backend plugins, React and Material UI) and clarifies what the exam does not cover.

Spotify logo
Spotify

Building Confidence: A Case Study in How to Create Confidence Scores for GenAI Applications

Spotify's Financial Engineering team evaluated three approaches to generate confidence scores for GenAI outputs in an invoice-parsing automation: a separate calibrator model, token-level logprobs, and majority-voting ensembles. Majority voting (with weighted votes and calibration via Platt scaling) showed the strongest correlation with accuracy. The post covers implementation details (number of models, weighting strategies, calibration), limitations (long-text fields, granularity), and operational tradeoffs (cost, latency, permutations of prompts to increase granularity).

Spotify logo
Spotify

How We Generated Millions of Content Annotations

Spotify built a scalable annotation platform (three pillars: scaling human expertise, annotation tooling, and foundational infrastructure/integration) to generate millions of high-quality annotations for ML and GenAI use cases. The platform combines large annotator workforces and LLM-based augmentation, custom interfaces for audio/video/NLP tasks, project metrics and agreement handling, and flexible integrations (data models, APIs, CLIs/UIs, batch orchestration) to accelerate model training and evaluation (10x annotation corpus, ~3x annotator productivity improvements).

Spotify logo
Spotify

Are You a Dalia? How We Created Data Science Personas for Spotify’s Analytics Platform

Spotify’s Analytics Platform team created six personas for internal data practitioners using a mixed-methods process (interviews + quantitative rules), built a daily-updated dataset to classify employees into personas, and drove adoption through stakeholder engagement, visual assets, and practical recommendations.

Spotify logo
Spotify

Unlocking Insights with High-Quality Dashboards at Scale

Spotify describes its dashboard platform and practices: supporting thousands of Tableau and Looker Studio dashboards via a SQL scheduling tool that produces Tableau .hyper extracts from BigQuery, a Dashboard Quality Framework (Vital Signs + Spicy Design Checklist) that labels dashboards (Low/High/Golden), and a Dashboard Portal for discovery, curation, embedding, and usage metrics. The post focuses on data-platform work to scale dashboard production, governance, and consumption.

Spotify logo
Spotify

Technical Decision-Making in a Fragmented Space: Spotify In-Car Case Study

Spotify describes engineering challenges of delivering consistent in‑car experiences across Bluetooth, projected (Apple CarPlay, Android Auto) and embedded platforms (Tesla, Android Automotive), and explains how they use RFCs (with diagrams and code examples) to break down problems, align stakeholders, and make architecture and product decisions (automatic playback, APIs, logging) across teams.

Spotify logo
Spotify

Data Platform Explained Part II

Part II of Spotify's Data Platform series describes the platform's building blocks (data collection, management, processing), the event delivery infrastructure and scale (≈1 trillion events/day, 38k pipelines), the tooling/runtime choices (BigQuery, Flink, Dataflow, Scio/Beam, Kubernetes operators, PubSub), the move from on-prem Hadoop to GCP, and operational practices around monitoring, lineage, access controls and developer community.

Spotify logo
Spotify

Fixed-Power Designs: It’s Not IF You Peek, It’s WHAT You Peek at

A Spotify devblog post summarizing Nordin and Schultzberg (2024) which introduces the "fixed-power design": estimate the required sample size from outcome data while the experiment is running and stop when current sample size exceeds that estimate. The paper shows the difference-in-means estimator is consistent under this design and that standard fixed-sample confidence intervals have asymptotically correct coverage, so no sequential-correction is required. The post compares fixed-sample, sequential, and fixed-power designs, discusses practical implications for running experiments at scale, and mentions Spotify's experimentation engine, Confidence.

Spotify logo
Spotify

Data Platform Explained

Spotify engineers describe the motivation, evolution, and core components of their internal data platform—covering data collection, processing, and management. The post explains triggers for building a platform (searchable/democratized data, reporting, data quality, experimentation, and ML enablement), recalls starting from a large Hadoop cluster, and points to experimentation and ML as key platform use cases. Future posts in the series will dive deeper into implementation details and lessons learned.

Spotify logo
Spotify

Risk-Aware Product Decisions in A/B Tests with Multiple Metrics

This post summarizes a 2024 research paper introducing a risk-aware framework for evaluating A/B tests with multiple metrics. The approach quantifies the probability of making incorrect decisions across various success criteria and provides guidelines for balancing trade-offs between metrics.

Spotify logo
Spotify

Applying the Facade Pattern on Spotify for Artists

Spotify for Artists applied the facade design pattern by creating S4X‑Masthead (a package that exports MastheadHeader, MastheadFooter, and MastheadProvider React components) to keep masthead UI and integrations consistent across local design systems (Encore) and a generalized support platform (Turnkey). Data is loaded dynamically from a CMS and integrations (translations, authentication) are exposed via the provider to prevent implementation drift across teams. The article motivates the approach with Conway’s Law and the Inverse Conway Maneuver and outlines benefits and future work.

Spotify logo
Spotify

Exploring the Animation Landscape of 2023 Wrapped

Spotify describes the animation strategy for 2023 Wrapped: adopting a Lottie-first approach for generic brand visuals while retaining native animations for parameterized, data-driven visuals to achieve web parity, reduce engineering duplication, and balance payload/operational trade-offs and designer/engineer responsibilities.