Rethinking the Language Runtime for Scale
summary
Claim: Lasp runtime targets clusters of 10k–20k nodes by combining CRDT-based language semantics with a reworked runtime. Implementation uses Erlang, Docker, Mesos, EPMD (initially), mochi_global, and a custom partisan membership layer plus a sprinter bootstrap and a “workflow CRDT” for distributed orchestration. Early runs reveal concrete tradeoffs: 5 GB of state shipped in 90s for five nodes, later optimizations drop per-instance RAM from ~2 GB to ~75 MB, reproducible 300-node experiments succeed while 1k+ runs see partition races, 10k remains future work.
AI-written summary · about
chapters (11)
- 0:00Start
- 3:37Weak Synchronization
- 11:08Selective Hearing
- 13:05Layered Approach
- 20:28Experiment Design
- 22:09Ideal Experiment
- 22:53Initial Evaluation
- 31:27Debugging Sprinter
- 32:23Evaluation Strategy
- 33:31Completion Detection
- 36:51Results Partisan
related talks
14mScalable Solutions for Running Large Language Models
17mRLM: Recursive Language Models for Large Codebases
31mDeploying Large Language Models with Ease
39mErlang on Microcontrollers: The Research Continues
05mProject Lightning Talk: Extend Large Language Model Training Beyond Single Kubernetes Cluster
42mStop Treating LLMs Like REST APIs
19mSystem for Scaling Large Language Models
28mAnyscale's Unified Platform for LLM Development and Deployment