Lifting the veil on Meta’s microservice architecture: Analyses of topology and...
summary
Analysis of Meta's microservice estate uses Canopy traces and 22 months of deployment records to show topology is heterogeneous and in constant flux: the inference platform alone creates ~100 new service IDs per day, regular services ~10, 89% of new service IDs get deprecated, and total deployed instances grow to ~12M. Request workflows are wide and shallow, but call depth is unreliable; sampling and multi-tenancy drive variation. Grouping by children set explains much of the concurrency and reduces workflow variance.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:44Assumptions
- 3:52Testbeds
- 4:52Findings
- 6:48Methodology
- 7:32Unique Service IDs
- 9:17Daily Churn Deployed Services
- 10:31Long Term Growth
- 11:27Distributed Tracing
- 12:36Trace Profiles
- 13:31Properties
- 16:46Concurrency Rates
- 18:35Implications
related talks
23mMicroservices – Diptanu Choudhury
31mHundreds of Microservices without breaking your APIs
26mFighting the Good Fight at the Hot Gates of Micro-Services
18mBuilding Resilient Monitoring at Meta
37mMicroservice Gardening
46mModelling Microservices at Spotify
24mHow Deep Systems Broke Observability
18mMaking Deployments Safe at Meta