• Forecast as a Service
  • Introduction
  • System Architecture
  • Component Overview
    • Dash Frontend
    • Fastapi Backend
    • Forecast Engine
    • Data Flow
  • Environment and Deployment
  • Known Limitations
  • Future Improvements

Forecast as a Service¶

Introduction¶

• This guide:
  • Describes the architecture of the Forecast-as-a-Service system
  • Explains how the Dash frontend, FastAPI backend, and forecasting engine interact
  • Shows how to run, extend, and customize the system effectively

System Architecture¶

• There are two containers:
  • a Dash app for interaction
  • a FastAPI server for processing
    • Uses Prophet to perform time series forecasting
• The containers are Orchestrated via Docker Compose

Component Overview¶

Dash Frontend¶

• Built using Plotly Dash
• Handles CSV upload
• Sends data to the FastAPI /forecast endpoint
• Displays visual output

Fastapi Backend¶

• Handles /upload_data and /forecast
• Parses uploaded CSVs, validates input, and invokes forecasting logic
• Uses Pydantic models defined in api/schemas.py
• Main server entrypoint: api/main.py

Forecast Engine¶

• Uses Facebook Prophet for univariate forecasting
• Logic is encapsulated in api/services.py
• Can be extended to support other models by modifying the run_forecast() function

Data Flow¶

• User uploads a CSV via the UI
• Frontend sends POST to FastAPI /forecast
• Backend parses and validates the file
• Forecast is computed using Prophet
• Result is sent back and plotted interactively

Environment and Deployment¶

• Environment is configured via setenv.sh and invoke.yaml

• The Docker image is tagged with --version, allowing multiple builds

• Ports 8000 (backend) and 8050 (frontend) must be available

• To run forecast UI:

  > ./devops/docker_run/run_docker_forecast.sh <version>

Known Limitations¶

• The forecast service must be launched via run_docker_forecast.sh, not through invoke
• The frontend and backend services are hardcoded to ports 8050 and 8000, respectively. This can cause conflicts on shared environments
• Only Prophet model is currently supported
• Only single forecast per request

Future Improvements¶

• Create an invoke task (e.g., invoke docker_forecast) similar to docker_bash and docker_jupyter
• Externalize port configuration into .env or accept CLI args
• Refactor engine to use a plug-in registry for multiple model backends
• Integrate background job processing to allow batching or queuing system