app.add_middleware(MiddlewareB)
```

이 경우 실행 순서는 다음과 같습니다:

* **요청**: MiddlewareB → MiddlewareA → route

* **응답**: route → MiddlewareA → MiddlewareB

이러한 스태킹 동작은 미들웨어가 예측 가능하고 제어 가능한 순서로 실행되도록 보장합니다.

## 다른 미들웨어 { #other-middlewares }

다른 미들웨어에 대한 더 많은 정보는 나중에 [숙련된 사용자 안내서: 향상된 미들웨어](../advanced/middleware.md)에서 확인할 수 있습니다.

你可能記得，請求體也會用 Pydantic 模型宣告。

這裡因為你使用了 `Depends`，**FastAPI** 不會混淆。

///

/// check | 檢查

這個依賴系統的設計讓我們可以有不同的依賴（不同的 "dependables"），都回傳 `User` 模型。

我們不受限於只能有一個能回傳該類型資料的依賴。

///

## 其他模型 { #other-models }

現在你可以在*路徑操作函式*中直接取得目前使用者，並在**依賴注入**層處理安全機制，使用 `Depends`。

而且你可以為安全需求使用任意模型或資料（本例中是 Pydantic 模型 `User`）。

但你不受限於某個特定的資料模型、類別或型別。

想在模型中只有 `id` 與 `email` 而沒有任何 `username`？當然可以。你可以用同樣的工具達成。

///

/// warning | 警告

如果您的应用程序依赖于生命周期事件， `AsyncClient` 将不会触发这些事件。为了确保它们被触发，请使用 [florimondmanca/asgi-lifespan](https://github.com/florimondmanca/asgi-lifespan#usage) 中的 `LifespanManager` 。

///

## 其他异步函数调用 { #other-asynchronous-function-calls }

由于测试函数现在是异步的，因此除了在测试中向 FastAPI 应用程序发送请求之外，您现在还可以调用（和使用 `await` 等待）其他 `async` 异步函数，就和您在代码中的其他任何地方调用它们的方法一样。

/// tip | 提示

// Figure out a way to avoid this.
// TODO(srbs): Should ZerosLike check-fail instead of returning nullptr?
class TapeTensor {
 public:
  explicit TapeTensor(AbstractTensorHandle* handle);
  TapeTensor(const TapeTensor& other);
  ~TapeTensor();

  int64_t GetID() const;
  tensorflow::DataType GetDType() const;

  AbstractTensorHandle* ZerosLike() const;

  AbstractTensorHandle* GetHandle() const;

 private:

  - [Changelog since v1.34.3](#changelog-since-v1343)
  - [Changes by Kind](#changes-by-kind-2)
    - [Feature](#feature-1)
    - [Bug or Regression](#bug-or-regression-1)
    - [Other (Cleanup or Flake)](#other-cleanup-or-flake)
  - [Dependencies](#dependencies-2)
    - [Added](#added-2)
    - [Changed](#changed-2)
    - [Removed](#removed-2)
- [v1.34.3](#v1343)
  - [Downloads for v1.34.3](#downloads-for-v1343)

# wheels, we install them in a virtual environment and run some smoke tests on
# it. TFCI Mac VMs only have one Python version installed so we need to install
# the other versions manually.
if [[ "${TFCI_MACOS_PYENV_INSTALL_ENABLE}" == 1 ]]; then
  # Install the necessary Python, unless it's already present
  pyenv install -s "$TFCI_PYTHON_VERSION"
  pyenv local "$TFCI_PYTHON_VERSION"

```

The benchmark code defines default values for the parameters so if
you leave out any out JMH will run with each default value, one after
the other. This will run with `interval` set to `calendar year` then
`calendar hour` then `10d` then `5d` then `1h`:
```

                    throw new IllegalStateException(
                        "No " + DOCKER_COMPOSE_YML + " found for " + project.getPath() + " nor does it use other fixtures."
                    );
                }
            });
        }

        extension.fixtures.matching(fixtureProject -> fixtureProject.equals(project) == false)

 * includes both natural exceptions such as unreachable servers, as well as synthetic exceptions
 * when responses are of an unexpected type or cannot be decoded.
 *
 * Other exception types cancel the current call:
 *
 *  * For synchronous calls made with [Call.execute], the exception is propagated to the caller.
 *

   *       happens all the time in normal conditions (i.e., android) the OS must make a tradeoff
   *       between paging memory and killing other processes - so allocating a gigantic buffer and
   *       then sequentially accessing it could result in other processes dying. This is solvable
   *       via madvise(2), but that obviously doesn't exist in java.