Building MLOps Infrastructure at Japan's Largest C2C E-Commerce Site
summary
Mercari adds learning-to-rank as a re-ranker over Elasticsearch, serving a small DNN (trained with TF‑ranking / PyTorch) via a separate RPC microservice with timeouts and fallbacks to preserve latency (tens of ms) under thousands QPS; offline pipelines move from BigQuery SQL to DBT, online features go Bigtable → Memorystore Redis, and re-ranking depth is 120 candidates. Packaging the model as packaged model in container images speeds scaleup; introducing Seldon + Istio for routing and shadow traffic cuts A/B test setup ~70% and drops skips from ~60% → ~12% on peak.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:59About Mercari
- 2:53Problems
- 7:22MLOps
- 14:50Data Pipeline
- 17:46ML Model Serving
- 25:21Selden Istio
- 28:39MLOps Monitoring
- 30:37Conclusion
- 34:41Questions
related talks
44mRobust AI Search Ranking for Radical C2C Marketplace Growth
25mBetter search relevance using Learning to Rank at mobile.de
29mBuilding ML Platform Capabilities at Global Scale
20mAuto Content Moderation in C2C e-Commerce
40mBuilding highload ML powered service
30mOperationalizing Machine Learning
29mTeam Teardown: MLOps in the Cloud for IoT
26mManaging ML Models @ Scale