Cracking the Cost Code: A Cost Optimized Approach to Distributed...
summary
Intuit replaces stateful trace stores and indexing with columnar Parquet files on S3+Kafka ingestion, per-service Parquet schemas generated by a metadata manager, and stateless query engines (Trino/ClickHouse/Athena) to serve trace-by-ID and exemplar searches. Peak ingest reaches ~8M spans/sec; query load stays under ~100 QPS, yielding ~1–2s searches. Compactors reduce small files, partitioning by time and service, and this architecture delivers a 65% cost reduction while scaling from 500k to 8M TPS without prescaling.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 2:55User Journey
- 5:32Architecture
- 9:33Criteria
- 11:43Row vs Column
- 13:40Park Injection
- 17:13Final Architecture
- 18:49Data Injection
- 21:51Summary
related talks
32mDiscover the Modern Way to Build Real-Time Analytics Solutions at Any Scale
29mIntroduction to Distributed Tracing
29mWide Event Analytics
26mColumnar Storage: Redefining Data Management for the Modern Era
42mColumnar Storage: Redefining Data Management for the Modern Era
25mRemove Slow, Expensive ETL & Bring Data To The Data Lake For Lower Cost, More Performant Analytics
16mFrom Sampling To Full Visibility: Scaling Tracing To Trillions of Spans
37mKEYNOTE: Data Architecture Turned Upside Down