Skip to content

SYSTEM Cited by 1 source

Mako (Netflix ML compute platform)

Mako is Netflix's internal ML compute platform. It provisions GPUs on AWS and is the substrate that sits underneath open-source components like PyTorch, Ray, and vLLM in Netflix's Post-Training Framework stack. First canonical wiki reference: sources/2026-02-13-netflix-scaling-llm-post-training-at-netflix.

Definition

"At the base is Mako, Netflix's internal ML compute platform, which provisions GPUs on AWS." (Source: sources/2026-02-13-netflix-scaling-llm-post-training-at-netflix)

Role

Mako is the GPU-provisioning compute substrate — Netflix's ML equivalent of a scheduler/allocator layer — above which the rest of the LLM post-training stack runs:

Post-Training Framework  ← library above
PyTorch / Ray / vLLM     ← OSS, unmodified
Mako                     ← this page: GPU provisioning on AWS
AWS GPU instances

This is a foundational-platform-plus-domain-libraries pattern: Mako handles compute allocation generically, domain-specific libraries (post-training framework, presumably others for inference/pretraining) live above it.

What the source reveals

  • Provisions GPUs on AWS (not on-prem).
  • Intended to be the stable lower layer: the post-training framework runs PyTorch/Ray/vLLM "largely out of the box" on top of Mako, and the framework's four-pillar value-add (Data/Model/Compute/Workflow) sits above that.

What the source does not reveal

  • Relationship (if any) to systems/metaflow, Netflix's other ML platform.
  • Whether Mako handles scheduling, quota, capacity management, or just provisioning.
  • Cluster topology, networking fabric, or job-lifecycle primitives.
  • Whether Mako is Netflix's "single ML platform" or one of several.
Last updated · 550 distilled / 1,221 read