stream(tags).of(stream -> stream.sorted().reduce((l, r) -> l + r).ifPresent(v -> buf.append(v)));
        buf.append(CACHE_KEY_SPLITTER);
        stream(roles).of(stream -> stream.sorted().reduce((l, r) -> l + r).ifPresent(v -> buf.append(v)));
        buf.append(CACHE_KEY_SPLITTER);
        stream(fields).of(stream -> stream.sorted().reduce((l, r) -> l + r).ifPresent(v -> buf.append(v)));

    public static String createMavenVersionString(Properties buildProperties) {
        String timestamp = reduce(buildProperties.getProperty("timestamp"));
        String version = reduce(buildProperties.getProperty(BUILD_VERSION_PROPERTY));
        String rev = reduce(buildProperties.getProperty("buildNumber"));
        String distributionName = reduce(buildProperties.getProperty("distributionName"));

        String msg = distributionName + " ";

      }
    },
    REDUCE_LAST {
      @Override
      Object operate(Stream<?> stream) {
        return stream.reduce((a, b) -> b);
      }
    },
    REDUCE_LAST_PARALLEL {
      @Override
      Object operate(Stream<?> stream) {
        return stream.parallel().reduce((a, b) -> b);
      }
    };

    abstract Object operate(Stream<?> stream);
  }

  @Param private Operation operation;

Doing this provides the following specific benefits:
- **Reduce the API surface** - We no longer need to maintain two methods.
- **Consistency** - All languages have consistent access to the same APIs and ergonomics in the DSL.
- **Reduce the size of the Gradle distribution** - We no longer need to carry multiple standard libraries for different languages.

## Decision

This container branches off from
/tensorflow/tools/tf_sig_build_dockerfiles/. However, since
hermetic CUDA and hermetic Python is now available for Tensorflow, a lot of the
requirements installed on the original container can be removed to reduce the
footprint of the container and make it more reusable across different ML

mapreduce.fileoutputcommitter.algorithm.version=2 # Use the latest committer version
mapreduce.job.reduce.slowstart.completedmaps=0.99 # 99% map, then reduce
mapreduce.reduce.shuffle.input.buffer.percent=0.9 # Min % buffer in RAM
mapreduce.reduce.shuffle.merge.percent=0.9 # Minimum % merges in RAM
mapreduce.reduce.speculative=false # Disable speculation for reducing
mapreduce.task.io.sort.factor=999 # Threshold before writing to disk

				// Number of probes.
				// ~ cat /proc/sys/net/ipv4/tcp_keepalive_probes (defaults to 9, we reduce it to 5)
				_ = syscall.SetsockoptInt(fd, syscall.IPPROTO_TCP, syscall.TCP_KEEPCNT, 5)

				// Wait time after successful probe in seconds.
				// ~ cat /proc/sys/net/ipv4/tcp_keepalive_intvl (defaults to 75 secs, we reduce it to 15 secs)
				_ = syscall.SetsockoptInt(fd, syscall.IPPROTO_TCP, syscall.TCP_KEEPINTVL, 15)
			}

  TF_DeleteTensor(model_out_tensor);

  // If the output is a scalar, then return the scalar output
  if (num_dims_out == 0) {
    outputs[0] = model_out.release();
    return absl::OkStatus();
  }

  // Else, reduce sum the output to get a scalar

  // Will sum all dimensions, so get a Tensor containing [0,...,num_dims_out-1].
  AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);

net.ipv4.tcp_max_syn_backlog=16384
net.core.wmem_max=4194304
net.core.rmem_max=4194304
net.core.rmem_default=4194304
net.core.wmem_default=4194304
net.ipv4.tcp_rmem="4096 87380 4194304"
net.ipv4.tcp_wmem="4096 65536 4194304"

# Reduce CPU utilization
net.ipv4.tcp_timestamps=0

# Increase throughput
net.ipv4.tcp_sack=1

# Low latency mode for TCP
net.ipv4.tcp_low_latency=1

# The following variable is used to tell the kernel how 

			need = 1 // to ensure we read at least one byte for every Read
			tokens = b
		}
	} else { // all tokens go towards payload
		need = int(math.Min(float64(b), float64(need)))
		tokens = need
	}
	// reduce tokens requested according to availability
	av := int(r.throttle.Tokens())
	if av < tokens && av > 0 {
		tokens = av
		need = int(math.Min(float64(tokens), float64(need)))
	}
	err = r.throttle.WaitN(r.ctx, tokens)