### Modify the OpenAPI schema { #modify-the-openapi-schema }

Now you can add the ReDoc extension, adding a custom `x-logo` to the `info` "object" in the OpenAPI schema:

{* ../../docs_src/extending_openapi/tutorial001_py310.py hl[22:24] *}

### Cache the OpenAPI schema { #cache-the-openapi-schema }

You can use the property `.openapi_schema` as a "cache", to store your generated schema.

// different from the number returned by TFE_TensorHandleNumDims.
// The return value was current at the time of TFE_TensorDebugInfo creation.
TF_CAPI_EXPORT extern int TFE_TensorDebugInfoOnDeviceNumDims(
    TFE_TensorDebugInfo* debug_info);

// Returns the number of elements in dimension `dim_index`.
// Tensor representation on device can be transposed from its representation

# Stream Data { #stream-data }

If you want to stream data that can be structured as JSON, you should [Stream JSON Lines](../tutorial/stream-json-lines.md).

But if you want to **stream pure binary data** or strings, here's how you can do it.

/// info

Added in FastAPI 0.134.0.

///

## Use Cases { #use-cases }

You could use this if you want to stream pure strings, for example directly from the output of an **AI LLM** service.

      if (compareAndSet(currentRunner, blocker)) {
        // Thread.interrupt can throw arbitrary exceptions due to the nio InterruptibleChannel API
        // This will make sure that tasks don't get stuck busy waiting.
        // Some of this is fixed in jdk11 (see https://bugs.openjdk.org/browse/JDK-8198692) but
        // not all.  See the test cases for examples on how this can happen.
        try {
          ((Thread) currentRunner).interrupt();

# Templates { #templates }

You can use any template engine you want with **FastAPI**.

A common choice is Jinja2, the same one used by Flask and other tools.

There are utilities to configure it easily that you can use directly in your **FastAPI** application (provided by Starlette).

## Install dependencies { #install-dependencies }

Make sure you create a [virtual environment](../virtual-environments.md), activate it, and install `jinja2`:

   * higher value than you need can waste space and time, and a significantly lower value can lead
   * to thread contention. But overestimates and underestimates within an order of magnitude do not
   * usually have much noticeable impact. A value of one permits only one thread to modify the cache
   * at a time, but since read operations and cache loading computations can proceed concurrently,

# Background Tasks - `BackgroundTasks`

You can declare a parameter in a *path operation function* or dependency function with the type `BackgroundTasks`, and then you can use it to schedule the execution of background tasks after the response is sent.

You can import it directly from `fastapi`:

```python
from fastapi import BackgroundTasks
```

We can do better...

## What makes a dependency { #what-makes-a-dependency }

Up to now you have seen dependencies declared as functions.

But that's not the only way to declare dependencies (although it would probably be the more common).

The key factor is that a dependency should be a "callable".

A "**callable**" in Python is anything that Python can "call" like a function.

    /**
     * Gets the decrypted proxies.
     *
     * @return The decrypted proxy, can be empty but never {@code null}.
     */
    List<Proxy> getProxies();

    /**
     * Gets the problems that were encountered during the settings decryption.
     *
     * @return The problems that were encountered during the decryption, can be empty but never {@code null}.
     */
    List<SettingsProblem> getProblems();

			out <- grid.NewBytesWithCopyOf(buf.Bytes())
		}
	}
}

// TraceHandler sends http trace messages back to peer rest client
func (s *peerRESTServer) TraceHandler(ctx context.Context, payload []byte, _ <-chan []byte, out chan<- []byte) *grid.RemoteErr {
	var traceOpts madmin.ServiceTraceOpts
	err := json.Unmarshal(payload, &traceOpts)
	if err != nil {
		return grid.NewRemoteErr(err)
	}
	var wg sync.WaitGroup