`"LOCAL"` (case insensitive). If `"AUTO"`, tf.data service runtime
        decides which workers to read from. If `"ANY"`, TF workers read from any
        tf.data service workers. If `"LOCAL"`, TF workers will only read from
        local in-processs tf.data service workers. `"AUTO"` works well for most
        cases, while users can specify other targets. For example, `"LOCAL"`

Aqui estão algumas combinações e estratégias possíveis:

* **Uvicorn** com `--workers`
    * Um **gerenciador de processos** Uvicorn escutaria no **IP** e na **porta** e iniciaria **vários processos de trabalho Uvicorn**.
* **Kubernetes** e outros **sistemas de contêineres** distribuídos

// * int is used as an index into arrays.
// * Deletion functions are safe to call on nullptr.
//
// Questions left to address:
// * Might at some point need a way for callers to provide their own Env.
// * Maybe add TF_TensorShape that encapsulates dimension info.
//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation

func (rc *ReadCloser) Close() error {
	return rc.f.Close()
}

// DataOffset returns the offset of the file's possibly-compressed
// data, relative to the beginning of the zip file.
//
// Most callers should instead use [File.Open], which transparently
// decompresses data and verifies checksums.
func (f *File) DataOffset() (offset int64, err error) {
	bodyOffset, err := f.findBodyOffset()
	if err != nil {
		return

	if source.noWait {
		select {
		case globalBackgroundHealRoutine.tasks <- task:
			if serverDebugLog {
				fmt.Printf("Task in the queue: %#v\n", task)
			}
		default:
			// task queue is full, no more workers, we shall move on and heal later.
			return nil
		}
		// Don't wait for result
		return nil
	}

	// respCh must be set to wait for result.
	// We make it size 1, so a result can always be written

像 Kubernetes 这样的分布式容器管理系统通常有一些集成的方法来处理**容器的复制**，同时仍然支持传入请求的**负载均衡**。 全部都在**集群级别**。

在这些情况下，你可能希望从头开始构建一个 **Docker 镜像**，如[上面所解释](#dockerfile)的那样，安装依赖项并运行 **单个 Uvicorn 进程**，而不是运行 Gunicorn 和 Uvicorn workers这种。


### 负载均衡器

使用容器时，通常会有一些组件**监听主端口**。 它可能是处理 **HTTPS** 的 **TLS 终止代理** 或一些类似的工具的另一个容器。

由于该组件将接受请求的**负载**并（希望）以**平衡**的方式在worker之间分配该请求，因此它通常也称为**负载均衡器**。

/// tip

It is the recommended server for Starlette and **FastAPI**.

/// check | "**FastAPI** recommends it as"

The main web server to run **FastAPI** applications.

You can also use the `--workers` command line option to have an asynchronous multi-process server.

Check more details in the [Deployment](deployment/index.md){.internal-link target=_blank} section.

///

## Benchmarks and speed

	}
	if isHeaderOnlyType(flag) {
		size = 0
	}
	tw.curr = &regFileWriter{tw.w, size}
	tw.pad = blockPadding(size)
	return nil
}

// AddFS adds the files from fs.FS to the archive.
// It walks the directory tree starting at the root of the filesystem
// adding each file to the tar archive while maintaining the directory structure.
func (tw *Writer) AddFS(fsys fs.FS) error {

## Gunicornによる公式Dockerイメージ - Uvicorn

前の章で詳しく説明したように、Uvicornワーカーで動作するGunicornを含む公式のDockerイメージがあります： [Server Workers - Gunicorn と Uvicorn](server-workers.md){.internal-link target=_blank}で詳しく説明しています。

このイメージは、主に上記で説明した状況で役に立つでしょう： [複数のプロセスと特殊なケースを持つコンテナ（Containers with Multiple Processes and Special Cases）](#containers-with-multiple-processes-and-special-cases)

		ChecksumSHA1:   partInfo.Checksums["SHA1"],
		ChecksumSHA256: partInfo.Checksums["SHA256"],
	}, nil
}

// GetMultipartInfo returns multipart metadata uploaded during newMultipartUpload, used
// by callers to verify object states
// - encrypted
// - compressed
// Does not contain currently uploaded parts by design.