Title here
Summary here
Stage 2: Image-Builder Deployment
The FLOps manager tries to look up matching container images for learners and aggregators. The manager checks for images in the FLOps image registry (part of the management suite) that match the ML repository that was part of the requested project SLA.
If a match is found, there is no need to build redundant images, and the project goes straight to stage 4.
If images for FL Actors (learners & aggregators) using the requested ML repository are missing, the manager will create and deploy a single image-builder service. This builder service is exclusive to its originating project.
╭─────────────────────┬──────────────────────────┬────────────────┬──────────────────┬──────────────────────────╮
│ Service Name │ Service ID │ Instances │ App Name │ App ID │
├─────────────────────┼──────────────────────────┼────────────────┼──────────────────┼──────────────────────────┤
│ │ │ │ │ │
│ builder8266202cd6db │ 6761bf5e59461659a24b1199 │ 0 RUNNING │ projc3fd78f56b75 │ 6761bf5d59461659a24b1197 │
│ │ │ │ │ │
╰─────────────────────┴──────────────────────────┴────────────────┴──────────────────┴──────────────────────────╯ To build or not to build?
Stage 2 is distinct because:
- Deciding if new images need to be built requires querying the remote ML Git repository and the FLOps image registry.
- The image-builder service can only run on worker nodes with the
image-builderaddon enabled. - Deploying the large image builder service (pulled size ~3GB) can take time.