@GwtCompatible
abstract class CollectionFuture<V extends @Nullable Object, C extends @Nullable Object>
    extends AggregateFuture<V, C> {
  /*
   * We access this field racily but safely. For discussion of a similar situation, see the comments
   * on the fields of TimeoutFuture. This field is slightly different from the fields discussed
   * there: cancel() never reads this field, only writes to it. That makes the race here completely

@GwtCompatible
abstract class CollectionFuture<V extends @Nullable Object, C extends @Nullable Object>
    extends AggregateFuture<V, C> {
  /*
   * We access this field racily but safely. For discussion of a similar situation, see the comments
   * on the fields of TimeoutFuture. This field is slightly different from the fields discussed
   * there: cancel() never reads this field, only writes to it. That makes the race here completely

 *   <li>or the usual, <i>robust</i> behavior, which all known {@link Collection} implementations
 *       have, of returning a new, independent iterator
 * </ul>
 *
 * <p>Because of this situation, any public method accepting an iterable should invoke the {@code
 * iterator} method only once, and should be tested using this class. Exceptions to this rule should
 * be clearly documented.
 *

 *   <li>or the usual, <i>robust</i> behavior, which all known {@link Collection} implementations
 *       have, of returning a new, independent iterator
 * </ul>
 *
 * <p>Because of this situation, any public method accepting an iterable should invoke the {@code
 * iterator} method only once, and should be tested using this class. Exceptions to this rule should
 * be clearly documented.
 *

   * {@link IllegalArgumentException} if passed a {@code toKey} greater than an earlier {@code
   * toKey}. However, this method doesn't throw an exception in that situation, but instead keeps
   * the original {@code toKey}.
   */
  @Override
  public ImmutableSortedMap<K, V> headMap(K toKey) {
    return headMap(toKey, false);
  }

  /**

    // below the values of characters in surrogate pairs, this cannot occur.
    // This approach does mean that we break out of the fast path code in cases
    // where we don't strictly need to, but this situation will almost never
    // occur in practice.
    if (safeMin >= Character.MIN_HIGH_SURROGATE) {
      // The safe range is empty or the all safe code points lie in or above the

   * if Fire.run is called by delegate.addListener since addListener is called after the assignment
   * to timer, and importantly this is the main situation in which we need to be able to see the
   * write.
   *
   * 2. visibility of the writes to an afterDone() call triggered by cancel():
   *

 *       "safe" edge is created.
 *   <li>If a cycle is detected, an "unsafe" (cyclic) edge is created to represent a potential
 *       deadlock situation, and the appropriate Policy is executed.
 * </ul>
 *
 * <p>Note that detection of potential deadlock does not necessarily indicate that deadlock will

Symmetrically, interceptors can rewrite response headers and transform the response body. This is generally more dangerous than rewriting request headers because it may violate the webserver's expectations!

If you're in a tricky situation and prepared to deal with the consequences, rewriting response headers is a powerful way to work around problems. For example, you can fix a server's misconfigured `Cache-Control` response header to enable better response caching:

 *       "safe" edge is created.
 *   <li>If a cycle is detected, an "unsafe" (cyclic) edge is created to represent a potential
 *       deadlock situation, and the appropriate Policy is executed.
 * </ul>
 *
 * <p>Note that detection of potential deadlock does not necessarily indicate that deadlock will