🦙 llama_deploy 🤖

llama_deploy (formerly llama-agents) is an async-first framework for deploying, scaling, and productionizing agentic multi-service systems based on workflows from llama_index. With llama_deploy, you can build any number of workflows in llama_index and then bring them into llama_deploy for deployment.

In llama_deploy, each workflow is seen as a service, endlessly processing incoming tasks. Each workflow pulls and publishes messages to and from a message queue.

At the top of a llama_deploy system is the control plane. The control plane handles ongoing tasks, manages state, keeps track of which services are in the network, and also decides which service should handle the next step of a task using an orchestrator. The default orchestrator is purely programmatic, handling failures, retries, and state-passing.

Why llama_deploy?

1. Seamless Deployment: It bridges the gap between development and production, allowing you to deploy llama_index workflows with minimal changes to your code.

2. Scalability: The microservices architecture enables easy scaling of individual components as your system grows.

3. Flexibility: By using a hub-and-spoke architecture, you can easily swap out components (like message queues) or add new services without disrupting the entire system.

4. Fault Tolerance: With built-in retry mechanisms and failure handling, llama_deploy ensures robustness in production environments.

5. State Management: The control plane manages state across services, simplifying complex multi-step processes.

6. Async-First: Designed for high-concurrency scenarios, making it suitable for real-time and high-throughput applications.

Wait, where is llama-agents?

The introduction of Workflows in llama_indexproduced the most intuitive way to develop agentic applications. The question then became: how can we close the gap between developing an agentic application as a workflow, and deploying it?

With llama_deploy, the goal is to make it as 1:1 as possible between something that you built in a notebook, and something running on the cloud in a cluster. llama_deploy enables this by simply being able to pass in and deploy any workflow.

Installation

llama_deploy can be installed with pip, and relies mainly on llama_index_core:

pip install llama_deploy

Getting Started

High-Level Deployment

llama_deploy provides a simple way to deploy your workflows using configuration objects and helper functions.

When deploying, generally you'll want to deploy the core services and workflows each from their own python scripts (or docker images, etc.).

Here's how you can deploy a core system and a workflow:

Deploying the Core System

To deploy the core system (message queue, control plane, and orchestrator), you can use the deploy_core function:

from llama_deploy import (
    deploy_core,
    ControlPlaneConfig,
    SimpleMessageQueueConfig,
)


async def main():
    await deploy_core(
        control_plane_config=ControlPlaneConfig(),
        message_queue_config=SimpleMessageQueueConfig(),
    )


if name == "main":
    import asyncio

    asyncio.run(main())

This will set up the basic infrastructure for your llama_deploy system. You can customize the configs to adjust ports and basic settings, as well as swap in different message queue configs (Redis, Kafka, RabbiMQ, etc.).

Deploying a Workflow

To deploy a workflow as a service, you can use the deploy_workflow function:

python
from llama_deploy import (
    deploy_workflow,
    WorkflowServiceConfig,
    ControlPlaneConfig,
    SimpleMessageQueueConfig,
)
from llama_index.core.workflow import Workflow, StartEvent, StopEvent, step


# create a dummy workflow
class MyWorkflow(Workflow):
    @step()
    async def run_step(self, ev: StartEvent) -> StopEvent:
        # Your workflow logic here
        arg1 = str(ev.get("arg1", ""))
        result = arg1 + "_result"
        return StopEvent(result=result)


async def main():
    await deploy_workflow(
        workflow=MyWorkflow(),
        workflow_config=WorkflowServiceConfig(
            host="127.0.0.1", port=8002, service_name="my_workflow"
        ),
        control_plane_config=ControlPlaneConfig(),
    )


if name == "main":
    import asyncio

    asyncio.run(main())

This will deploy your workflow as a service within the llama_deploy system, and register the service with the existing control plane and message queue.

Interacting with your Deployment

Once deployed, you can interact with your deployment using a client.

from llama_deploy import LlamaDeployClient

# points to deployed control plane
client = LlamaDeployClient(ControlPlaneConfig())

session = client.create_session()
result = session.run("my_workflow", arg1="hello_world")
print(result)
# prints 'hello_world_result'

Deploying Nested Workflows

Every Workflow is capable of injecting and running nested workflows. For example

from llama_index.core.workflow import Workflow, StartEvent, StopEvent, step


class InnerWorkflow(Workflow):
    @step()
    async def run_step(self, ev: StartEvent) -> StopEvent:
        arg1 = ev.get("arg1")
        if not arg1:
            raise ValueError("arg1 is required.")

        return StopEvent(result=str(arg1) + "_result")


class OuterWorkflow(Workflow):
    @step()
    async def run_step(
        self, ev: StartEvent, inner: InnerWorkflow
    ) -> StopEvent:
        arg1 = ev.get("arg1")
        if not arg1:
            raise ValueError("arg1 is required.")

        arg1 = await inner.run(arg1=arg1)

        return StopEvent(result=str(arg1) + "_result")


inner = InnerWorkflow()
outer = OuterWorkflow()
outer.add_workflows(inner=InnerWorkflow())

llama_deploy makes it dead simple to spin up each workflow above as a service, and run everything without any changes to your code!

Just deploy each workflow:

[!NOTE] This code is launching both workflows from the same script, but these could easily be separate scripts, machines, or docker containers!

import asyncio
from llama_deploy import (
    WorkflowServiceConfig,
    ControlPlaneConfig,
    deploy_workflow,
)


async def main():
    inner_task = asyncio.create_task(
        deploy_workflow(
            inner,
            WorkflowServiceConfig(
                host="127.0.0.1", port=8003, service_name="inner"
            ),
            ControlPlaneConfig(),
        )
    )

    outer_task = asyncio.create_task(
        deploy_workflow(
            outer,
            WorkflowServiceConfig(
                host="127.0.0.1", port=8002, service_name="outer"
            ),
            ControlPlaneConfig(),
        )
    )

    await asyncio.gather(inner_task, outer_task)


if name == "main":
    import asyncio

    asyncio.run(main())

And then use it as before:

from llama_deploy import LlamaDeployClient

# points to deployed control plane
client = LlamaDeployClient(ControlPlaneConfig())

session = client.create_session()
result = session.run("outer", arg1="hello_world")
print(result)
# prints 'hello_world_result_result'

Components of a llama_deploy System

In llama_deploy, there are several key components that make up the overall system

• message queue -- the message queue acts as a queue for all services and the control plane. It has methods for publishing methods to named queues, and delegates messages to consumers.
• control plane -- the control plane is a the central gateway to the llama_deploy system. It keeps track of current tasks and the services that are registered to the system. The control plane also performs state and session management and utilizes the orchestrator.
• orchestrator -- The module handles incoming tasks and decides what service to send it to, as well as how to handle results from services. By default, the orchestrator is very simple, and assumes incoming tasks have a destination already specified. Beyond that, the default orchestrator handles retries, failures, and other nice-to-haves.
• services -- Services are where the actual work happens. A services accepts some incoming task and context, processes it, and publishes a result. When you deploy a workflow, it becomes a service.

Low-Level Deployment

For more control over the deployment process, you can use the lower-level API. Here's what's happening under the hood when you use deploy_core and deploy_workflow:

deploy_core

The deploy_core function sets up the message queue, control plane, and orchestrator. Here's what it does:

async def deploy_core(
    control_plane_config: ControlPlaneConfig,
    message_queue_config: BaseSettings,
    orchestrator_config: Optional[SimpleOrchestratorConfig] = None,
) -> None:
    orchestrator_config = orchestrator_config or SimpleOrchestratorConfig()

    message_queue_client = _get_message_queue_client(message_queue_config)

    control_plane = ControlPlaneServer(
        message_queue_client,
        SimpleOrchestrator(**orchestrator_config.model_dump()),
        **control_plane_config.model_dump(),
    )

    message_queue_task = None
    if isinstance(message_queue_config, SimpleMessageQueueConfig):
        message_queue_task = _deploy_local_message_queue(message_queue_config)

    control_plane_task = asyncio.create_task(control_plane.launch_server())

    # let services spin up
    await asyncio.sleep(1)

    # register the control plane as a consumer
    control_plane_consumer_fn = await control_plane.register_to_message_queue()

    consumer_task = asyncio.create_task(control_plane_consumer_fn())

    # let things sync up
    await asyncio.sleep(1)

    # let things run
    if message_queue_task:
        all_tasks = [control_plane_task, consumer_task, message_queue_task]
    else:
        all_tasks = [control_plane_task, consumer_task]

    shutdown_handler = _get_shutdown_handler(all_tasks)
    loop = asyncio.get_event_loop()
    while loop.is_running():
        await asyncio.sleep(0.1)
        signal.signal(signal.SIGINT, shutdown_handler)

        for task in all_tasks:
            if task.done() and task.exception():  # type: ignore
                raise task.exception()  # type: ignore

This function:

1. Sets up the message queue client
2. Creates the control plane server
3. Launches the message queue (if using SimpleMessageQueue)
4. Launches the control plane server
5. Registers the control plane as a consumer
6. Sets up a shutdown handler and keeps the event loop running

deploy_workflow

The deploy_workflow function deploys a workflow as a service. Here's what it does:

async def deploy_workflow(
    workflow: Workflow,
    workflow_config: WorkflowServiceConfig,
    control_plane_config: ControlPlaneConfig,
) -> None:
    control_plane_url = control_plane_config.url

    async with httpx.AsyncClient() as client:
        response = await client.get(f"{control_plane_url}/queue_config")
        queue_config_dict = response.json()

    message_queue_config = _get_message_queue_config(queue_config_dict)
    message_queue_client = _get_message_queue_client(message_queue_config)

    service = WorkflowService(
        workflow=workflow,
        message_queue=message_queue_client,
        **workflow_config.model_dump(),
    )

    service_task = asyncio.create_task(service.launch_server())

    # let service spin up
    await asyncio.sleep(1)

    # register to message queue
    consumer_fn = await service.register_to_message_queue()

    # register to control plane
    control_plane_url = (
        f"http://{control_plane_config.host}:{control_plane_config.port}"
    )
    await service.register_to_control_plane(control_plane_url)

    # create consumer task
    consumer_task = asyncio.create_task(consumer_fn())

    # let things sync up
    await asyncio.sleep(1)

    all_tasks = [consumer_task, service_task]

    shutdown_handler = _get_shutdown_handler(all_tasks)
    loop = asyncio.get_event_loop()
    while loop.is_running():
        await asyncio.sleep(0.1)
        signal.signal(signal.SIGINT, shutdown_handler)

        for task in all_tasks:
            if task.done() and task.exception():  # type: ignore
                raise task.exception()  # type: ignore

This function:

1. Sets up the message queue client
2. Creates a WorkflowService with the provided workflow
3. Launches the service server
4. Registers the service to the message queue
5. Registers the service to the control plane
6. Sets up a consumer task for the service
7. Sets up a shutdown handler and keeps the event loop running

Using the llama_deploy client

llama_deploy provides both a synchronous and an asynchronous client for interacting with a deployed system.

Both clients have the same interface, but the asynchronous client is recommended for production use to enable concurrent operations.

Generally, there is a top-level client for interacting with the control plane, and a session client for interacting with a specific session. The session client is created automatically for you by the top-level client and returned from specific methods.

To create a client, you need to point it to a control plane.

from llama_deploy import (
    LlamaDeployClient,
    AsyncLlamaDeployClient,
    ControlPlaneConfig,
)

client = LlamaDeployClient(ControlPlaneConfig())
async_client = AsyncLlamaDeployClient(ControlPlaneConfig())

Client Methods

• client.create_session(poll_interval=DEFAULT_POLL_INTERVAL): Creates a new session for running workflows and returns a SessionClient for it. A session encapsulates the context and state for a single workflow run. Example:

session = client.create_session()

• client.list_sessions(): Lists all sessions registered with the control plane. Example:

sessions = client.list_sessions()
for session in sessions:
    print(session.session_id)

• client.get_session(session_id, poll_interval=DEFAULT_POLL_INTERVAL): Gets an existing session by ID and returns a SessionClient for it. Example:

session = client.get_session("session_123")

• client.get_or_create_session(session_id, poll_interval=DEFAULT_POLL_INTERVAL): Gets an existing session by ID, or creates a new one if it doesn't exist. Example:

session = client.get_or_create_session("session_123")

• client.get_service(service_name): Gets the definition of a service by name. Example:

service = client.get_service("my_workflow")
print(service.service_name, service.host, service.port)

• client.delete_session(session_id): Deletes a session by ID. Example:

client.delete_session("session_123")

• client.list_services(): Lists all services registered with the control plane. Example:

services = client.list_services()
for service in services:
    print(service.service_name)

• client.register_service(service_def): Registers a service with the control plane. Example:

service_def = ServiceDefinition(
    service_name="my_workflow", host="localhost", port=8000
)
client.register_service(service_def)

• client.deregister_service(service_name): Deregisters a service from the control plane. Example:

  client.deregister_service("my_workflow")
  

SessionClient Methods

• session.run(service_name, **run_kwargs): Implements the workflow-based run API for a session. Example:

result = session.run("my_workflow", arg1="hello", arg2="world")
print(result)

• session.create_task(task_def): Creates a new task in the session. Example:

task_def = TaskDefinition(input='{"arg1": "hello"}', agent_id="my_workflow")
task_id = session.create_task(task_def)

• session.get_tasks(): Gets all tasks in the session. Example:

tasks = session.get_tasks()
for task in tasks:
    print(task.task_id, task.status)

• session.get_current_task(): Gets the current (most recent) task in the session. Example:

current_task = session.get_current_task()
if current_task:
    print(current_task.task_id, current_task.status)

• session.get_task_result(task_id): Gets the result of a task in the session if it has one. Example:

  result = session.get_task_result("task_123")
  if result:
      print(result.result)
  

Examples

We have many examples showing how to use various message queues, and different ways to scaffold your project for deployment with docker and kubernetes.

You can find all examples in the examples folder for the llama-deploy repository.