Optimizing Cost and Performance with arm64
summary
Honeycomb.io evaluates and then migrates core services to arm64 over a year, moving 92% of compute workloads to Graviton2 m6g instances; ingestion service Shepherd runs as the first experiment in a dogfood environment, built with Go cross-compilation and arm docker images, yielding a 40% lower compute cost, higher RAM per instance, lower median and much lower 99th‑percentile tail latency. Migration exposes software incompatibilities and a Kafka rollback, but ultimately four of five largest services run on graviton2 handling ~90% of traffic.
AI-written summary · about
chapters (29)
- 0:00Intro
- 0:34WTF is architecture? Why multiarch?
- 2:09History: 80s, 90s, 00s, 10s, and beyond
- 6:52If it ain't broke...
- 8:52ARM is more efficient.
- 12:05Data storage engine and analytics tool
- 12:44Service Level Objectives (SLO)
- 13:02SLOs are user flows
- 13:52Same reliability, lower costs with ARM6
- 15:18Complexity stayed manageable
- 16:24Prod: customers observe data
- 17:30Kibble observes dogfood
- 18:02Dogfood observes prod
- 18:19Service Architecture
- 18:48Shepherd: ingest API service
- 19:26Is it feasible to migrate?
- 20:27Producing artifacts for Arm64
- 21:19Initial findings
- 22:01A/B testing
- 24:41Dogfood Shepherd cost reduction
- 25:03Migrated prod Shepherd
- 26:22AWS ran out of m6gd spot instances
- 27:13Kafka + the long tail
- 27:23Graviton2 going strong
- 27:44Have a measurable goal in mind
- 28:00Acknowledge hidden risks
- 28:16Take care of your people
- 28:41Optimize for safety
- 29:11Graviton2 blog posts
related talks
44mWhat to know before adopting Arm: Lessons learned at Datadog
22mGraviton
01h 14mModernizing and Optimizing Applications for the Cloud
25mAdopt AWS Graviton2 instances for best price-performance in EC2
36mWhy and How to Choose Arm64 for your Cloud Applications
34mARM Migration Made Practical
30mHome Lab to the Moon and Back
29mThe Progress for Cloud Computing on Arm architecture