Scaling Validation and Quality of Streaming Data Products at Twitter
summary
Claims concrete results: QE at Twitter builds a staging environment mirroring prod with fake clients and runs end-to-end stress tests, measures pipeline data loss against SLAs and enforces system tests as blocking deploy step so interaction bugs must be fixed before production. A custom load injection tool that duplicates real events lets teams hit tweets-per-second high-water marks. Example: a soccer-ball emoji operator failure is traced by cross-datacenter rule-comparison tests; teams adopt the tools and many issues get fixed earlier.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 0:24Good nip at Twitter
- 0:40Quality engineering
- 3:03Staging environment
- 3:48Operating constraints
- 4:57DDI DNS adage
- 6:12Testing for data products
- 8:03PowerTrack
- 8:41Two problems to solve
- 10:07Puli as hell to solve
- 14:50Experimenting with embedding
- 16:03Testing as a service
- 18:11Five factor testing
- 20:06Better testing tools
- 21:24Overhead
- 23:27What Were We Missing
- 24:08Stress Testing
- 26:17Load Testing
- 27:09Load Tool
- 29:03Reinventing Tools
related talks
42mDeciphering the way data is tested
37mRethinking How We Test Our Async Architecture
01h 02mThe Art & Joy of Testing in Production
19mTesting things that don't want to be tested: A game streaming story
50mData-Driven Engineering
25mContinuous Delivery and Testing in Production
56mTesting at scale
25mTesting Data Pipelines