When Worlds Collide: Productionalizing a Data Science Model
summary
Built a production reliability score for Specialists with two components (retention and completion), persisted histories and a daily gauge visible in the app; implementation uses Celery chords to parallelize per-specialist jobs and a yaml config file so data science can change weights without code. Initial Jupyter notebook handoff hid research and led engineers to change weights which lowered the mean score unexpectedly. Outcome: the pipeline works at scale; lessons are document decisions and keep data science in the loop in person.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:24Shift gig marketplace
- 2:47Reliability score for Specialists
- 4:55Data science to engineering handoff
- 7:44Bolero architecture design
- 13:27Implementing score with conflicts
- 16:35Conflict resolution breakthrough
- 19:01Testing production model
- 21:11Key takeaways
- 24:02Cross-functional teams
- 25:26Q&A
related talks
17mProductionizing Data Science Workloads
38mData Secrets from a Platform Engineer
29mData Science Best Practices : From Proof of Concepts to Production
32mTips and tricks: data science prototype into production
34mData Products, From 0 to data science pipeline
26mWorking with Data Scientists and why I left my job (not)
18mBeyond Inference: Bringing ML into Production
05mProductionalizing Data Science Models