We create an async function `lifespan()` with `yield` like this:

```Python hl_lines="16  19"
{!../../../docs_src/events/tutorial003.py!}
```

Here we are simulating the expensive *startup* operation of loading the model by putting the (fake) model function in the dictionary with machine learning models before the `yield`. This code will be executed **before** the application **starts taking requests**, during the *startup*.

  // The webhook cares about an operation if it matches _any_ Rule.
  // However, in order to prevent ValidatingAdmissionWebhooks and MutatingAdmissionWebhooks
  // from putting the cluster in a state which cannot be recovered from without completely
  // disabling the plugin, ValidatingAdmissionWebhooks and MutatingAdmissionWebhooks are never called

  // The webhook cares about an operation if it matches _any_ Rule.
  // However, in order to prevent ValidatingAdmissionWebhooks and MutatingAdmissionWebhooks
  // from putting the cluster in a state which cannot be recovered from without completely
  // disabling the plugin, ValidatingAdmissionWebhooks and MutatingAdmissionWebhooks are never called

    ```

### `Union` in Python 3.10

In this example we pass `Union[PlaneItem, CarItem]` as the value of the argument `response_model`.

Because we are passing it as a **value to an argument** instead of putting it in a **type annotation**, we have to use `Union` even in Python 3.10.

If it was in a type annotation we could have used the vertical bar, as:

```Python
some_variable: PlaneItem | CarItem
```

    }

    @Override
    @ParametricNullness
    public T get() {
      // Another variant of Double Checked Locking.
      //
      // We use two volatile reads. We could reduce this to one by
      // putting our fields into a holder class, but (at least on x86)
      // the extra memory consumption and indirection are more
      // expensive than the extra volatile reads.
      long nanos = expirationNanos;

    }

    @Override
    @ParametricNullness
    public T get() {
      // Another variant of Double Checked Locking.
      //
      // We use two volatile reads. We could reduce this to one by
      // putting our fields into a holder class, but (at least on x86)
      // the extra memory consumption and indirection are more
      // expensive than the extra volatile reads.
      long nanos = expirationNanos;

* It should not be client-specific.
  * In Istio sidecars, historically we had a lot of client-specific xDS. For example, putting the xDS-client's IP back into the xDS response. This makes efficient control plane implementation (most notably, caching), extremely challenging.

## Type Hints with Metadata Annotations

Python also has a feature that allows putting **additional metadata** in these type hints using `Annotated`.

=== "Python 3.9+"

    In Python 3.9, `Annotated` is part of the standard library, so you can import it from `typing`.

    ```Python hl_lines="1  4"

Then the cashier 💁 says "I'm finished with doing the burgers" by putting your number on the counter's display, but you don't jump like crazy immediately when the displayed number changes to your turn number. You know no one will steal your burgers because you have the number of your turn, and they have theirs.

				continue
			}

			// An online-disk means its a valid disk but it may be a re-connected disk
			// we verify that here based on LastConn(), however we make sure to avoid
			// putting it back into the s.erasureDisks by re-placing the disk again.
			_, setIndex, _ := cdisk.GetDiskLoc()
			if setIndex != -1 {
				continue
			}
		}
		if cdisk != nil {
			// Close previous offline disk.