configure a `work_dir` when running your dispatcher server and set
    `dispatcher_fault_tolerance=True`. The dispatcher will store its state to
    `work_dir`, so that on restart it can continue from its previous state after
    restart.
*   Adds support for sharing dataset graphs via shared filesystem instead of
    over RPC. This reduces load on the dispatcher, improving performance of

 * terms</a>):
 *
 * <ul>
 *   <li>directed graphs
 *   <li>undirected graphs
 *   <li>graphs that do/don't allow self-loops
 *   <li>graphs whose nodes/edges are insertion-ordered, sorted, or unordered
 * </ul>
 *
 * <p>{@code Graph} explicitly does not support parallel edges, and forbids implementations or
 * extensions with parallel edges. If you need parallel edges, use {@link Network}.
 *

 * terms</a>):
 *
 * <ul>
 *   <li>directed graphs
 *   <li>undirected graphs
 *   <li>graphs that do/don't allow self-loops
 *   <li>graphs whose nodes/edges are insertion-ordered, sorted, or unordered
 * </ul>
 *
 * <p>{@code Graph} explicitly does not support parallel edges, and forbids implementations or
 * extensions with parallel edges. If you need parallel edges, use {@link Network}.
 *

    thereof, to be licensed as a whole at no charge to all third parties
    under the terms of this License.

    c) If the modified program normally reads commands interactively
    when run, you must cause it, when started running for such
    interactive use in the most ordinary way, to print or display an
    announcement including an appropriate copyright notice and a notice
    that there is no warranty (or else, saying that you provide a

     * We will still use the readXxxWireFormat return values to
     * indicate how many bytes(note: readParameterWordsWireFormat
     * returns bytes read and not the number of words(but the
     * wordCount member DOES store the number of words)) we
     * actually read. Incedentally this is important to the
     * AndXServerMessageBlock class that needs to potentially
     * read in another smb's parameter words and bytes based on

      //    internal char array (hard to avoid using StringBuilder)
      // 2. we avoid extra copies into temporary buffers altogether
      // The downside is that this will cause us to store the file bytes in memory twice for a short
      // amount of time.
      return new String(ByteSource.this.read(), charset);
    }

    @Override
    public String toString() {

  }

  public void testVerifyMegaStreamHalves() {
    assertThat(
            concat(megaPrimitiveDoubleStreamPart1(), megaPrimitiveDoubleStreamPart2())
                .sorted()
                .toArray())
        .isEqualTo(megaPrimitiveDoubleStream().toArray());
  }

  public void testAddAllPrimitiveDoubleStream() {
    StatsAccumulator accumulator = new StatsAccumulator();

### Contributing code

If you have improvements to TensorFlow, send us your pull requests! For those
just getting started, GitHub has a
[how-to](https://help.github.com/articles/using-pull-requests/).

TensorFlow team members will be assigned to review your pull requests. Once the

    end
    Client--Plain-->CZ
    CZ-."HBONE (target)".->Server
    CZ--"HBONE (actual)"-->SZ
    SZ--Plain-->Server
```

### Pooling

User connections can be multiplexed over shared HBONE connections.
This is done through standard HTTP/2 pooling.
The pooling is keyed off the `{source identity, destination identity, destination ip}`.

### Headers

t:before{content:"\f105"}.fa-angle-up:before{content:"\f106"}.fa-angry:before{content:"\f556"}.fa-angrycreative:before{content:"\f36e"}.fa-angular:before{content:"\f420"}.fa-ankh:before{content:"\f644"}.fa-app-store:before{content:"\f36f"}.fa-app-store-ios:before{content:"\f370"}.fa-apper:before{content:"\f371"}.fa-apple:before{content:"\f179"}.fa-apple-alt:before{content:"\f5d1"}.fa-apple-pay:before{content:"\f415"}.fa-archive:before{content:"\f187"}.fa-archway:before{content:"\f557"}.fa-arrow-...