Backend System¶

Introduction¶

The backend system in oslo.service provides a modular architecture that allows you to choose between different concurrency models without modifying your service logic.

Currently, two backends are implemented and supported:

Eventlet backend: The traditional concurrency backend based on green threads and cooperative multitasking (greenlet library).
Threading backend: A new backend using native Python threads, Cotyledon, and Futurist, designed to be compatible with modern Python versions, simplify debugging, and improve maintainability.

Why a Backend System?¶

This system was introduced to enable a gradual transition away from the Eventlet backend, which has known limitations for scaling, cooperative multitasking complexity, and debugging in some Python environments. By using the backend system, you can test, adopt, or migrate to the Threading backend while keeping the same public API and service structure.

In the future, the Eventlet backend may be deprecated once the Threading backend covers all use cases. Some related features tied to Eventlet, like the embedded WSGI server and the backdoor shell, are also planned for removal.

For more context, see:

How It Works¶

The backend system exposes the same public components (ServiceLauncher, ProcessLauncher, ThreadGroup, LoopingCall, etc.) through a dynamic selection mechanism.

At runtime, you can select the backend you want to use in two ways:

Explicit initialization by calling init_backend() with the desired BackendType:

from oslo_service.backend import init_backend, BackendType

# Use the Threading backend
init_backend(BackendType.THREADING)

Dynamic selection using a hook function that decides which backend to use:

from oslo_service.backend import register_backend_default_hook, \
    BackendType

def my_backend_decider():
    # Add custom logic here
    return BackendType.THREADING

register_backend_default_hook(my_backend_decider)

If no backend is explicitly initialized, the system will use the Eventlet backend by default, unless a hook has been registered to override this choice.

Once the backend is initialized, it cannot be changed. Trying to re-initialize it with a different backend will raise a BackendAlreadySelected exception.

If you see this exception, ensure that your backend is initialized only once early during your application startup (for example in your main() entry point). Avoid calling init_backend() in shared libraries or multiple times. If you need to reset the backend in unit tests, use the _reset_backend() helper to clear the backend cache:

from oslo_service.backend import _reset_backend

_reset_backend()

Backend Selection Mechanism¶

The backend system uses a sophisticated caching and selection mechanism to ensure efficient and consistent backend selection throughout your application lifecycle.

Core Functions:

``init_backend(type_: BackendType)``: Explicitly initializes and caches a specific backend. This function loads the backend module, instantiates the backend class, and caches both the backend instance and its components. Once called, the backend cannot be changed to a different type.
``get_backend()``: Returns the currently cached backend instance. If no backend has been explicitly initialized, it follows this selection process:
1. If a hook is registered, calls the hook to determine the backend type
2. If the hook raises an exception, falls back to the default backend (Eventlet)
3. If no hook is registered, uses the default backend (Eventlet)
4. Initializes the selected backend and returns it
``register_backend_default_hook(hook: Callable[[], BackendType])``: Registers a callback function that decides which backend to use when no explicit initialization has occurred. The hook function should return a BackendType and will be called by get_backend() when needed.
``get_component(name: str)``: Retrieves a specific component from the cached backend. This is the primary way to access backend-specific implementations of service components.

Backend Caching:

The system maintains four global caches:

_cached_backend_type: The selected backend type
_cached_backend: The instantiated backend object
_cached_components: Dictionary of available components from the backend
_backend_hook: The registered hook function for dynamic backend selection

Backend Loading Process:

When a backend is initialized, the following steps occur:

Module Import: The backend module is dynamically imported using importlib.import_module()
Class Instantiation: The backend class (e.g., EventletBackend, ThreadingBackend) is instantiated
Component Caching: All available components are retrieved and cached
Type Caching: The backend type is stored for future reference

This process ensures that backend components are loaded only once and efficiently cached for subsequent access.

For implementation details, see:

https://opendev.org/openstack/oslo.service/src/branch/master/oslo_service/backend/__init__.py

Hook Function Requirements:

Your hook function must:

Accept no parameters
Return a valid BackendType enum value
Handle any exceptions internally (the system will fall back to the default if your hook raises an exception)

Example hook function:

def environment_based_backend():
    import os
    if os.environ.get('USE_THREADING_BACKEND'):
        return BackendType.THREADING
    return BackendType.EVENTLET

register_backend_default_hook(environment_based_backend)

The hook mechanism is primarily useful for library authors or applications that wish to select the backend dynamically based on environment, configuration, or context — without enforcing it explicitly.

Warning

If get_backend() or init_backend() has already been called before your hook is registered, the hook will be ignored and the default backend (usually Eventlet) will be used. Make sure to register the hook as early as possible, ideally before any other imports that might trigger backend usage.

How to Select a Backend¶

There are two supported ways to select which backend to use:

Explicit initialization

Call init_backend() early in your application entry point:
```
from oslo_service.backend import init_backend, BackendType

init_backend(BackendType.THREADING)
```
This will load the Threading backend and cache its components.

For practical usage, see how Nova does it: https://review.opendev.org/c/openstack/nova/+/948311

Dynamic hook

You may register a Python hook to dynamically decide the backend if init_backend() is not called:

from oslo_service.backend \
    import register_backend_default_hook, BackendType

def my_backend_decider():
    # Add custom logic here if needed
    return BackendType.THREADING

register_backend_default_hook(my_backend_decider)

If no explicit backend is set, the system will call your hook to decide which backend to load.

When to Use init_backend() vs register_backend_default_hook()¶

Use Case	Recommended Method
You control the application entry point	`init_backend()`
You write a shared library or plugin	`register_backend_default_hook()`
You want full predictability	`init_backend()`
You want context-based dynamic selection	`register_backend_default_hook()`

Note

There is no automatic oslo.config option for service_backend. If you want to make the backend configurable through your application’s config file, you need to define and parse that option yourself, then call init_backend() with the selected value.

Accessing Backend Components¶

Once the backend is initialized, you can access its components via the get_component() helper function.

Example: getting a ProcessLauncher instance:

from oslo_service.backend import get_component

ProcessLauncher = get_component("ProcessLauncher")

launcher = ProcessLauncher(conf)
launcher.launch_service(my_service)
launcher.wait()

Available components include:

ServiceLauncher
ProcessLauncher
ThreadGroup
LoopingCall variants
SignalHandler
And other related service utilities

If you try to access a component that does not exist in the current backend, a KeyError will be raised.

Example Usage¶

Here is a minimal example that shows how to initialize the Threading backend and launch a service:

from oslo_service.backend import (
    init_backend,
    get_component,
    BackendType
)

def main():
    # Select the Threading backend
    init_backend(BackendType.THREADING)

    # Get the ProcessLauncher component
    ProcessLauncher = get_component("ProcessLauncher")

    # Initialize your service and launch it
    launcher = ProcessLauncher(conf)
    launcher.launch_service(my_service)
    launcher.wait()

if __name__ == "__main__":
    main()

Here’s an example using a hook for dynamic backend selection:

from oslo_service.backend import (
    register_backend_default_hook,
    get_component,
    BackendType
)

def config_based_backend():
    # Read from configuration or environment
    if config.get('backend_type') == 'threading':
        return BackendType.THREADING
    return BackendType.EVENTLET

def main():
    # Register the hook before any backend access
    register_backend_default_hook(config_based_backend)

    # The backend will be selected automatically when needed
    ProcessLauncher = get_component("ProcessLauncher")
    launcher = ProcessLauncher(conf)
    launcher.launch_service(my_service)
    launcher.wait()

Common Backend Errors and Exceptions¶

When working with the backend system, you may encounter several specific exceptions and errors:

BackendAlreadySelected

Raised when trying to initialize a backend after one has already been selected. This typically happens when init_backend() is called multiple times in your application.

Solution: Ensure that init_backend() is called only once early in your application startup, typically in your main entry point.

KeyError

Raised when trying to access a component that doesn’t exist in the current backend using get_component().

Solution: Check that the component name is correct and that the component is available in your selected backend.

ValueError

Raised when trying to initialize an unknown backend type.

Solution: Ensure you’re using a valid BackendType enum value (BackendType.EVENTLET or BackendType.THREADING).

ImportError

Raised when the backend module cannot be imported or the backend class is not found in the module.

Solution: This usually indicates an installation issue or missing dependencies for the selected backend.

Hook Exceptions

If your hook function raises an exception, the system will log the error and fall back to the default backend (Eventlet).

Solution: Ensure your hook function handles all potential errors and always returns a valid BackendType.

Debugging Tips:

Use _reset_backend() in unit tests to clear the backend cache
Check that your backend initialization happens before any component access
Verify that all required dependencies are installed for your chosen backend
Use logging to track backend loading and component access
Test your hook function independently to ensure it works correctly

Understanding the Implications of Migrating from Eventlet to Threading¶

The following table is not about choosing a backend, but about understanding the implications of migrating away from Eventlet to the Threading backend:

Compared Aspect	Eventlet Backend	Threading Backend
Concurrency model	Cooperative multitasking using green threads (greenlet)	Native Python threads (preemptive)
Process management	Uses built-in eventlet processes	Uses Cotyledon for multi-process support
Looping calls	Uses eventlet green thread pools	Uses Futurist thread pools
Embedded WSGI server	Provided (with eventlet)	Planned for removal; use an external WSGI server instead
Backdoor shell	Provided (eventlet only)	Planned for removal
API compatibility	Same public API	Same public API
Debugging	More complex due to cooperative multitasking and green threads. See eventlet_backdoor.py	May produce clearer native thread stack traces and benefits from recent CPython improvements (e.g. PEP 768).
Python compatibility	Known issues with CPython native threads and RLocks. May break with new Python versions.	Fully compatible with modern CPython threading.

Test Setup in Devstack and CI Gates¶

If you want to test the Threading backend in Devstack or your CI gates, refer to:

References¶

For any questions, please refer to the OpenStack mailing list or the oslo.service maintainers.

Backend System