Remote execution, the DIY edition

on YouTube ↗31m

" summary

Stripe runs 80 million actions weekly across ~65,000 vCPUs and replaces gVisor with Firecracker to regain near‑native filesystem speed; content uses S3/DynamoDB and artifactory‑backed remote downloader with checksum validation. LVM snapshots attach input trees in a few hundred ms, SquashFS images cut node_modules to one file and give ~10% faster builds and ~50% memory drop, action merging saves ~20% executor load, and get_tree caching + parallel walk yields 20x latency improvement to ~500ms.

AI-written summary · about

" chapters (17)

1. 0:00Remote execution intro
2. 1:49Bazel at Stripe
3. 3:13Security challenges
4. 5:18Firecracker sandboxing
5. 6:42Controlled internet access
6. 8:10Cache tags for isolation
7. 9:34Time-traveling tests
8. 11:21Asset API for downloads
9. 13:07Tracing for debugging
10. 15:35Action merging
11. 17:40Hedging flaky tests
12. 18:43Large input roots
13. 20:50LVM snapshots
14. 21:54Optimizing get tree RPC
15. 22:55SquashFS for directories
16. 25:01Trade-offs and future
17. 26:03Q&A