Let's continue from the previous example. We wanted to **annotate the function with one type** but return something that includes **more data**.

We want FastAPI to keep **filtering** the data using the response model.

In the previous example, because the classes were different, we had to use the `response_model` parameter. But that also means that we don't get the support from the editor and tools checking the function return type.

    WeakReference<? extends Enum<?>> ref = Enums.getEnumConstants(enumClass).get(value);
    /*
     * We use `fromNullable` instead of `of` because `WeakReference.get()` has a nullable return
     * type.
     *
     * In practice, we are very unlikely to see `null`: The `WeakReference` to the enum constant

## Decision

We do not use Java serialization.
Instead, we use custom serialization where we explicitly describe how data objects should be serialized and deserialized.

For internal purposes, we use binary formats for their brevity.
We use the `Serializer` abstraction to separate the actual implementation of serialization from its uses.

When sharing data with external tools, we use JSON.

## Status

ACCEPTED

          if (rule.startsWith(EXCEPTION_RULE_MARKER)) {
            rule = rule.substring(1)
            // We use '\n' for end of value.
            totalExceptionRuleBytes += rule.size + 1
            sortedExceptionRules.add(rule)
          } else {
            totalRuleBytes += rule.size + 1 // We use '\n' for end of value.
            sortedRules.add(rule)
          }
        }
      }

Nevertheless, although we are not using the default integrated functionality, we are still using a Pydantic model to manually generate the JSON Schema for the data that we want to receive in YAML.

Then we use the request directly, and extract the body as `bytes`. This means that FastAPI won't even try to parse the request payload as JSON.

 * of project dependencies. This optimizes module loading during runtime, as we will only load external
 * modules that are not loaded transitively by other project modules.
 *
 * We perform this filtering, since if we simply include all external dependencies, regardless of whether
 * they are already loaded transitively, there is a measurable performance impact during module-loading.

		// There is at least one SSE-S3 single-part object.
		// For all SSE-S3 single-part objects we have to
		// fetch their decryption keys. We do this using
		// a Bulk-Decryption API call, if available.
		keys, err := crypto.S3.UnsealObjectKeys(ctx, k, metadata, buckets, names)
		if err != nil {
			return err
		}

		// Now, we have to decrypt the ETags of SSE-S3 single-part

	buf, err := xattr.LGet(s.formatFile, attr)
	if err != nil {
		// We start off with '0' if we can read the attributes
		return 0
	}
	return binary.LittleEndian.Uint64(buf[:8])
}

func (s *xlStorage) getWriteAttribute() uint64 {
	attr := "user.total_writes"
	buf, err := xattr.LGet(s.formatFile, attr)
	if err != nil {
		// We start off with '0' if we can read the attributes
		return 0
	}

	if askDisks == -1 {
		newDisks := getQuorumDisks(disks, infos, (len(disks)+1)/2)
		if newDisks != nil {
			// If we found disks signature in quorum, we proceed to list
			// from a single drive, shuffling of the drives is subsequently.
			disks = newDisks
			askDisks = 1
		} else {
			// If we did not find suitable disks, perform strict quorum listing
			// as no disk agrees on quorum anymore.

   * @param length the length of the new array
   */
  /*
   * The new array contains nulls, even if the old array did not. If we wanted to be accurate, we
   * would declare a return type of `@Nullable T[]`. However, we've decided not to think too hard
   * about arrays for now, as they're a mess. (We previously discussed this in the review of
   * ObjectArrays, which is the main caller of this method.)
   */