|| message.contains("not supported") || message.contains("capability")) {
                return true;
            }
        }

        // If we've had many errors, consider fallback
        double errorRate = statistics.getErrorRate();
        if (errorRate > 0.1) { // More than 10% error rate

### Threads

#### Application's calling thread

The application-layer must block on writing I/O. We can't return from a write until we've pushed its bytes onto the socket. Otherwise, if the write fails we are unable to deliver its IOException to the application. We would have told the application layer that the write succeeded, but it didn't!

     * https://stackoverflow.com/a/35754308/28465
     *
     * In that explanation's notation, our `overlap` check would be `x1 < y2 && y1 < x2`. We've
     * flipped one part of the check so that we're using "less than" in both cases (rather than a
     * mix of "less than" and "greater than"). We've also switched to "strictly less than" rather
     * than "less than or equal to" because of *handwave* the difference between "endpoints of

        same |= ~(hash1 ^ hash2);
        // test whether the hash values have different output bits
        diff |= hash1 ^ hash2;

        count++;
        // check whether we've exceeded the probabilistically
        // likely number of trials to have proven no funneling
        if (count > maxCount) {
          Assert.fail(
              "input bit("
                  + i

   * @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.
   */
  public static <T extends @Nullable Object> T[] newArray(T[] reference, int length) {

    for (double tolerance : BAD_TOLERANCE_CANDIDATES) {
      assertThrows(IllegalArgumentException.class, () -> DoubleMath.fuzzyEquals(1, 2, tolerance));
    }
  }

  /*
   * We've split testFuzzyCompare() into multiple tests so that our internal Android test runner has
   * a better chance of completing each within its per-test-method timeout.
   */

  public void testFuzzyCompare0() {

   * @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.
   */
  public static <T extends @Nullable Object> T[] newArray(T[] reference, int length) {

   * @return a long of a concatenated 8 bytes
   */
  static long load64Safely(byte[] input, int offset, int length) {
    long result = 0;
    // Due to the way we shift, we can stop iterating once we've run out of data, the rest
    // of the result already being filled with zeros.

    // This loop is critical to performance, so please check HashBenchmark if altering it.
    int limit = min(length, 8);

practice is to set conventions in a plugin (2). This keeps the underlying object "dumb", so it can be reused in multiple contexts and doesn't contain any special information about how conventions are calculated or what they could be.  In the wild, we've seen some objects set conventions in the object's constructor (3), but this can lead to unexpected assumptions or coupling between plugins.

For external plugins, it's difficult for an object to be used outside the application of a plugin,...

        FileOutputStream transfer = null;
        try {
          transfer = new FileOutputStream(temp);
          transfer.write(memory.getBuffer(), 0, memory.getCount());
          transfer.flush();
          // We've successfully transferred the data; switch to writing to file.
          out = transfer;
        } catch (IOException e) {
          if (transfer != null) {
            try {
              transfer.close();