* expect a newly-started thread to call [Runnable.run]. We shouldn't request new threads until
   * the already-requested ones are in service, otherwise we might create more threads than we need.
   *
   * We use [executeCallCount] and [runCallCount] to defend against starting more threads than we
   * need. Both fields are guarded by [lock].
   */
  private var executeCallCount = 0
  private var runCallCount = 0

Import the `Response` class (sub-class) you want to use and declare it in the *path operation decorator*.

For large responses, returning a `Response` directly is much faster than returning a dictionary.

And as you can have parallelism and asynchronicity at the same time, you get higher performance than most of the tested NodeJS frameworks and on par with Go, which is a compiled language closer to C <a href="https://www.techempower.com/benchmarks/#section=data-r17&hw=ph&test=query&l=zijmkf-1" class="external-link" target="_blank">(all thanks to Starlette)</a>.

### Is concurrency better than parallelism?

Nope! That's not the moral of the story.

   * `promisedStreamId` to which response frames will be delivered. Push promise frames are sent as
   * a part of the response to `streamId`. The `promisedStreamId` has a priority of one greater than
   * `streamId`.
   *
   * @param streamId client-initiated stream ID.  Must be an odd number.
   * @param promisedStreamId server-initiated stream ID.  Must be an even number.

          taskRunner.lock.withLock {
            activeThreads--
            startNextTask()
          }
        }
      }
    }
  }

  /**
   * This blocking queue hooks into a fake clock rather than using regular JVM timing for functions
   * like [poll]. It is only usable within task faker tasks.
   */
  private inner class TaskFakerBlockingQueue<T>(
    val delegate: BlockingQueue<T>,

    fifty percent (50%) of the outstanding shares or beneficial
    ownership of such entity.

2. License Grants and Conditions
--------------------------------

2.1. Grants

Each Contributor hereby grants You a world-wide, royalty-free,
non-exclusive license:

(a) under intellectual property rights (other than patent or trademark)

                (calleeReference.calleeOrCandidateName != OperatorNameConventions.INVOKE && targetSymbol.isInvokeFunction())
            ) {
                // Implicit invoke (e.g., `x()`) will have a different callee symbol (e.g., `x`) than the candidate (e.g., `invoke`).
                createKtPartiallyAppliedSymbolForImplicitInvoke(
                    candidate.dispatchReceiver,
                    candidate.chosenExtensionReceiver,

  internal val lock: ReentrantLock = ReentrantLock()

  /**
   * The maximum number of requests to execute concurrently. Above this requests queue in memory,
   * waiting for the running calls to complete.
   *
   * If more than [maxRequests] requests are in flight when this is invoked, those requests will
   * remain in flight.
   */
  var maxRequests = 64
    get() = this.withLock { field }
    set(maxRequests) {

## Version 5.0.0-alpha.9

_2022-06-16_

 *  New: Enforce label length limits in URLs. `HttpUrl` now rejects URLs whose domains aren't valid.
    This includes overly-long domain names (longer than 253 characters), overly-long labels (more
    than 63 characters between dots), and empty labels.
 *  New: Don't include the `Content-Length` header in multipart bodies. Servers must delimit

  /** Settings we communicate to the peer. */
  val okHttpSettings =
    Settings().apply {
      // Flow control was designed more for servers, or proxies than edge clients. If we are a client,
      // set the flow control window to 16MiB.  This avoids thrashing window updates every 64KiB, yet
      // small enough to avoid blowing up the heap.
      if (builder.client) {