A Case Study in API Cost of Running Analytics in the Cloud at Scale with an...
summary
Measurement shows 1–10 PB/day Presto workloads produce >10 billion small I/O requests (50% under 10 KB), yielding cloud GET costs that can exceed $1M/year. Adding a Presto local cache and SSD tiers on HDFS cuts cross-network reads by ~64% and reduces processes blocked by I/O by ~86%. An Alluxio distributed cache prototype delivers ~8–10 GB/s per worker, linear scale, and drives P95 read latency from ~100 ms down toward sub‑millisecond for small objects, proving caching both lowers cost and improves latency.
AI-written summary · about
chapters (13)
- 0:00API cost in cloud analytics
- 3:12Modern data stack challenges
- 7:57Cost of small fragmented requests
- 10:30Allux cache for Presto
- 11:32Uber case study overview
- 13:38HDFS workload analysis findings
- 15:48Local cache experiments
- 17:18Soft affinity scheduler
- 20:07HDFS SSD caching
- 22:35Caching for AI workloads
- 26:28PyArrow caching for AI
- 27:55Cost risks and caching benefits
- 29:01Q&A
related talks
20mData Caching for Enterprise-Grade Petabyte-Scale OLAP
28mOptimizing Latency-Sensitive Queries for Presto at Facebook
22mPresto for Your Fast, Open SQL Lakehouse
22mIntroduction to Presto: The SQL Engine for Data Platform Teams
24mCracking the Cost Code: A Cost Optimized Approach to Distributed...
46mPresto - Today and Beyond - The Open Source SQL Engine for Querying all Data Lakes
17mCloud Compute Cost Showdown
17mBreaking Through AI Workload Bottlenecks How OceanBase’s AI Native Capabilities Empower Enterprise