Gaiaflow Documentation

Gaiaflow is a local-first MLOps infrastructure tool that simplifies the process of building, testing, and deploying ML workflows. It provides an opinionated CLI for managing Airflow, MLflow, and other dependencies, abstracting away complex configurations, and giving you a smooth developer experience.

Features

Local MLOps Infrastructure via CLI with pre-installed prerequisites.
Simplified Airflow + Xcom Integration with an easy interface for DAG creation.
Cookiecutter template for project scaffolding with standardized structure.
State-of-the-art services provided:
- Apache Airflow
- MLflow
- MinIO
- JupyterLab
Containerization support to package workflow steps as python packages in Docker images.
Deployment support:
- Local development
- Minikube production-like testing locally
Future roadmap:
- Model deploy locally and in production
- Streamlined production deployment documentation
- Centralized DAG repository for CI/CD pipelines
- Private onboarding steps for users

Gaiaflow-Core

Gaiaflow library provides a simplified interface to create airflow tasks:

from gaiaflow.core.create_task import create_task

The create_task function is a high-level abstraction on top of Apache Airflow operators.

This interface makes it easy and intuitive to develop a DAG in Airflow without worrying about the complexities of Airflow concepts and configurations about running them in different environments allowing you to define tasks for your DAGs in a uniform way.

Instead of writing different operator classes manually for development, containers, or production deployments, create_task automatically selects the correct operator and applies consistent configuration based on the mode argument.

It supports function execution via a dotted and colon string (e.g. my_package.module:function) with args/kwargs in any of the supported modes without changing the user code.

It is configurable with:

Docker images
Secrets
Environment variables
Retry policies
Airflow params

Arguments for: create_task

Name	Type	Default	Description
task_id	`str`	–	Unique identifier for the task in the DAG.
func_path	`str`	–	Path to the Python function to execute. Format: `"module:function_name"`.
func_args	`list`	`[]`	Positional arguments for the function. The user can also provide output from upstream tasks as args input using the `FromTask` class. Usage shown below.
func_kwargs	`dict`	`{}`	Keyword arguments to pass into the function. The user can also provide output from upstream tasks as kwargs input using the `FromTask` class. Usage shown below.
image	`str`	`None`	Docker image to run the task in `dev_docker`, `prod_local` or `prod` mode. Ignored in `dev` mode.
mode	`str`	`"dev"`	Execution mode. Must be one of: `["dev", "dev_docker", "prod_local", "prod"]`.
secrets	`list`	`None`	List of secrets (e.g., Kubernetes secrets) required by the task in `prod_local` or `prod` mode.
env_vars	`dict`	`{}`	Environment variables to inject into the task. Can be used in `dev_docker`, `prod_local` or `prod` mode.
retries	`int`	`3`	Number of times to retry the task on failure.
dag	`airflow.DAG`	`None`	Optional reference to the DAG object. Used to inherit DAG-level params. Should be passed if `params` are defined in your DAG

`FromTask`

When building workflows with Gaiaflow, tasks often depend on the output of a previous task.

The FromTask helper makes this easy: it allows you to pull values from upstream tasks and use them as arguments in downstream tasks.

What It Does - Provides a simple way to pass values between tasks without writing manual XCom logic. - Works seamlessly inside func_args and func_kwargs of create_task. - Lets you specify: - Which task to pull from (task) - Which key inside the upstream task’s return dictionary to extract (key)

Example usage:

from airflow import DAG
from airflow.utils.task_group import TaskGroup
from datetime import datetime
from gaiaflow.core.create_task import create_task
from gaiaflow.core.operators import FromTask

default_args = {
    "owner": "airflow",
}

MODE = "dev"  # or "dev_docker", "prod_local", "prod"

with DAG(
    "my_training_dag",
    default_args=default_args,
    description="Example DAG using create_task",
    schedule="0 12 * * *",
    start_date=datetime(2025, 1, 1),
    catchup=False,
    tags=["ml", "gaiaflow"],
    params={"experiment": "mnist"}, # This will be available to your task via 
                                    # `kwargs[params]`
) as dag:

    with TaskGroup(group_id="training_group") as trainer:

        preprocess = create_task(
            task_id="preprocess_data",
            func_path="my_package:preprocess",
            func_kwargs={"path": "/data/raw"},
            mode=MODE,
            dag=dag,
        )

        train = create_task(
            task_id="train_model",
            func_path="my_package:train",
            func_kwargs={
                "epochs": 10, 
                # The key `preprocessed_path` is the kwarg expected by the train method
                "preprocessed_path": FromTask(
                    task="training_group.preprocess_data", # This tells which task to pull the value from
                    key="preprocessed_path", # This tells that we want to extract the value of `preprocessed_path`. 
                    # The previous task must make sure that they return a dict with `preprocessed_path` in it
                    # like return {"preprocessed_path": processed_path}.
                )},
            mode=MODE,
            dag=dag,
        )

        evaluate = create_task(
            task_id="evaluate_model",
            func_path="my_package:evaluate",
            func_kwargs={"metrics": ["accuracy", "f1"]},
            mode=MODE,
            dag=dag,
        )

        preprocess >> train >> evaluate

Best Practices for passing values from one task to another

Always make your task functions return a dictionary of literals. This ensures downstream tasks can pull values by key.
Use clear, descriptive keys like "preprocessed_path", "model_path", "metrics".

Gaiaflow behaviour by `mode`

The mode argument in create_task controls where and how your function runs. Internally, each mode maps to a different Airflow operator, with environment-specific defaults for networking, MLflow, MinIO, and execution backend.

dev → Local Python execution
- Airflow Operator: ExternalPythonOperator
- Runs your function directly on the Python environment created for you in the docker containers using your dependencies.
- Ignores the image, secrets and env_vars params.
- Uses your local .env file for environment variables.
- You will be using this mode most of the time for iteration, development and debugging.
dev_docker → Local Docker container execution
- Airflow Operator: DockerOperator
- Runs your function inside a Docker container
- Requires image parameter
- You can pass in env_vars which are then available as environment variables.
- This mode is useful when you want to test the docker image locally before moving to Kubernetes which helps you in catching bugs early on.
prod_local → Simulates production on your machine using Kubernetes via Local Minikube
- Operator: KubernetesPodOperator
- Runs tasks as Kubernetes pods in Minikube.
- Requires image
- You can pass in env_vars and/or secrets if any.
- For creating secrets in your minikube cluster, you can run
```
gaiaflow prod-local create-secret -p . --name <your-secret-name> --data SOME_KEY=value --data SOME_SECRET=secret
```
prod → Kubernetes in-cluster (production)
- Operator: KubernetesPodOperator
- Runs tasks as Kubernetes pods in Minikube.
- Requires image
- You can pass in env_vars and/or secrets if any.
- For creating secrets in your minikube cluster (coming soon)

Quick rule of thumb:

Use dev for rapid iteration and development.
Use dev_docker to validate your Docker image locally.
Use prod_local for testing end-to-end workflows on production-like settngs.
Use prod for production pipelines in the real cluster.

User workflow

A typical user workflow could look like this:

flowchart TD
    A1[Start]--> A
    A{Prerequisites installed except template?}-->|Yes| D{Which OS?}
    A -->|No| STOP[Stop]

    D -->|Windows WSL2| E1[Install template via conda/mamba/miniforge prompt<br/>then create env in WSL2 CLI + install gaiaflow]
    D -->|Linux| E2[Install template + create env<br/>then install gaiaflow]
    E1 --> M
    E2 --> M

    M[Start services in Dev Mode<br/>`gaiaflow start dev -p .`] --> N[Experiment in JupyterLab if needed]

    N --> P[Refactor to production code<br/>inside Python package and logging experiements to MLFlow and artifacts to S3 MinIO if needed]
    P --> Q[Write tests in tests/ folder]
    Q --> R[Create workflows with create_task mode=dev in Airflow]
    R --> S[Run & monitor DAG in Airflow]
    S --> T[Track experiments in MLflow & outputs in MinIO]
    T --> V{Workflow ran successfully in dev mode?}
    V --> |Yes| V1{Test your package in docker mode?}

    V1 --> |Yes| V2[Change create_task mode to dev_docker and run gaiaflow dev dockerize -p .]

    V2 --> V3{Docker mode ran successfully?}
    V3 --> |Yes| V4{Move to production?}

    V4 --> |Yes| X1[Start Prod-Local infra<br/>`gaiaflow prod-local start -p .`]

    V --> |No| V5[Fix bugs]
    V1 --> |No| STOP[Stop]
    V3 --> |No| V6[Fix bugs]
    V4 --> |No| STOP[Stop]

    V5 --> R
    V6 --> V3

    X1 --> X2[Build Docker image<br/>`gaiaflow prod-local dockerize -p .`]
    X2 --> X3[Configure secrets if needed]
    X3 --> Y{Prod-Local Success?}
    Y2 --> Y
    Y -->|Yes| Y1[Set create_task mode to prod - Local services are not needed anymore]
    Y1 --> Z[Deploy to Production Cluster ]

    Y -->|No| Y2[Fix the bugs]

    class A,D,U,Y,V,V1,V3,V4 decision;
    class E1,F1,G1,H1 windows;
    class E2,G2,H2 linux;
    class Z prod;
    class STOP,STOP2 stop;

    classDef decision fill:#FFDD99,stroke:#E67E22,stroke-width:2px;
    classDef windows fill:#AED6F1,stroke:#2874A6,stroke-width:2px;
    classDef linux fill:#ABEBC6,stroke:#1E8449,stroke-width:2px;
    classDef prod fill:#D7BDE2,stroke:#7D3C98,stroke-width:2px;
    classDef stop fill:#F5B7B1,stroke:#922B21,stroke-width:2px;