Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
102 lines (78 loc) · 3.97 KB

ARCHITECTURE.md

File metadata and controls

102 lines (78 loc) · 3.97 KB
Raw

AutoML Server Architecture

Overview

automl_server is the main FastAPI service. It owns the platform API surface for datasets, annotations, tasks, deployment, inference, AI pipeline bindings, and home-page statistics across image, text, multimodal conversation, and preference labeling workflows.

  • Default host: 0.0.0.0
  • Default port: 45678
  • Swagger UI: /swagger-ui
  • Health check: /health

Structure

automl_server/
├── app/
│   ├── common/      # shared response, exceptions, constants
│   ├── config/      # settings, database, S3, Nacos
│   ├── db/          # SQLAlchemy models and sessions
│   ├── modules/    # feature modules
│   ├── mq/          # RabbitMQ client, handlers, messages
│   ├── scheduler/   # periodic jobs
│   └── utils/       # HTTP, S3, SSE and file helpers
├── run.py
├── requirements.txt
└── Dockerfile

Feature Modules

  • dataset: dataset CRUD, upload, preview, export, sample management
  • annotation: annotation project CRUD and record handling for detection, classification, segmentation, pose, LLM/MLLM, and DPO workflows
  • task: training task submission, progress, logs, and SSE stream
  • deploy: model deployment management and status syncing
  • inference: inference API for deployed models
  • ai_pipeline: AI pipeline templates, providers, and bindings
  • home: platform statistics and recent activity

Runtime Dependencies

  • MySQL for primary persistence
  • MinIO for datasets, models, annotations, and artifacts
  • RabbitMQ for training jobs, AI Pipeline RPC, batch annotation execution, and deployment status callbacks
  • Nacos for runtime configuration
  • model_trainer as an optional training plug-in
  • model_deploy as an optional deployment plug-in
  • ai_pipeline_runtime as an optional AI-assisted annotation plug-in
  • ai_pipeline_sandbox as the batch annotation execution plug-in

Communication Paths

  • Training:
    • automl_server -> RabbitMQ -> model_trainer for training job delivery
    • model_trainer -> RabbitMQ -> automl_server for task status, task logs, and model registration
    • automl_server -> model_trainer over HTTP for /health and /tasks/{task_id}/cancel
  • Deployment:
    • automl_server -> model_deploy over HTTP for deploy, undeploy, deployment listing, deployment health, and inference
    • model_deploy -> RabbitMQ -> automl_server for model.deployed and model.undeployed
  • AI runtime:
    • automl_server <-> ai_pipeline_runtime via RabbitMQ RPC for assist annotation and pipeline execution
  • Batch annotation:
    • automl_server -> RabbitMQ -> ai_pipeline_sandbox for pipeline.batch.execute
    • ai_pipeline_sandbox -> RabbitMQ -> automl_server for progress and result callbacks

service.heartbeat still exists in the shared message definitions, but trainer and deploy do not currently use a standalone MQ heartbeat channel. Service status in the UI is derived from on-demand /health probes initiated by automl_server.

Request Flow

  1. Frontend calls automl_server
  2. automl_server reads config from Nacos or environment variables
  3. The service persists metadata to MySQL and artifacts to MinIO
  4. Training jobs are published to RabbitMQ when model_trainer is enabled
  5. Deployment control and inference requests are sent to model_deploy over HTTP
  6. Deployment status, training status, logs, model registration, and batch callbacks are consumed back from RabbitMQ
  7. AI-assisted annotation calls are routed to ai_pipeline_runtime via RabbitMQ RPC when enabled

Configuration

Main configuration lives in:

  • app/config/settings.py
  • app/config/s3_config.py
  • app/config/nacos_config_center.py

Environment examples:

  • automl_server/.env.example
  • root .env.example

Nacos config source:

  • nacos/automl-config.yaml

Local Run

cd automl_server
pip install -r requirements.txt
python run.py

The service reads APP_HOST and APP_PORT from the environment and falls back to 0.0.0.0:45678.