Single vs. Multiple Filters in the Java Stream API

One of the key features of Java 8 is the stream. It is frequently used in conjunction with lambdas, and one of them is the filter.

Let's consider the following example:

long count = doubles
   .stream()
   .filter(d -> d < Math.PI)
   .filter(d -> d > Math.E)
   .filter(d -> d != 3.10040970053377777)
   .filter(d -> d != 2.96240970053377777)
   .count();

It doesn't do anything fancy—perhaps it has no practical use case. However, for now, let's consider how the filter works. Each filter() method returns a new stream, so there in effect four extra steam.

However, of the four filters that can be written, one which has slightly less overhead. Let's compare these two ideas and see much benefit we can derive. 

import org.openjdk.jmh.annotations.Benchmark;
import org.openjdk.jmh.annotations.BenchmarkMode;
import org.openjdk.jmh.annotations.Mode;
import org.openjdk.jmh.annotations.OutputTimeUnit;

import java.util.List;
import java.util.Random;
import java.util.concurrent.TimeUnit;
import java.util.stream.Collectors;

public class MyBenchmark {

	@Benchmark
	@BenchmarkMode(Mode.All)
	@OutputTimeUnit(TimeUnit.SECONDS)
	public long testStreamWithSingleFilter() {
		List<Double> doubles = new Random().doubles(1_000, 1, 4).boxed().collect(Collectors.toList());
		long count = doubles
			.stream()
			.filter(d -> d < Math.PI
				&& d > Math.E
				&& d != 3.10040970053377777
				&& d != 2.96240970053377777)
			.count();
		
		return count;
	}

	@Benchmark
	@BenchmarkMode(Mode.All)
	@OutputTimeUnit(TimeUnit.SECONDS)
	public long testStreamWithMultipleFilter() {
		List<Double> doubles = new Random().doubles(1_000, 1, 4).boxed().collect(Collectors.toList());
		long count = doubles
			.stream()
			.filter(d -> d > Math.E)
			.filter(d -> d < Math.PI)
			.filter(d -> d != 3.10040970053377777)
			.filter(d -> d != 2.96240970053377777)
			.count();
		
		return count;
	}
}

Output: 

# Run complete. Total time: 00:40:19

Benchmark                                                                        Mode      Cnt      Score     Error  Units
MyBenchmark.testStreamWithMultipleFilter                                        thrpt      200  24367.016 ± 169.686  ops/s
MyBenchmark.testStreamWithSingleFilter                                          thrpt      200  32779.157 ± 127.938  ops/s
MyBenchmark.testStreamWithMultipleFilter                                         avgt      200     ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithSingleFilter                                           avgt      200     ≈ 10⁻⁵             s/op
MyBenchmark.testStreamWithMultipleFilter                                       sample  2581418     ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p0.00    sample              ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p0.50    sample              ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p0.90    sample              ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p0.95    sample              ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p0.99    sample              ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p0.999   sample               0.001             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p0.9999  sample               0.001             s/op
MyBenchmark.testStreamWithMultipleFilter:testStreamWithMultipleFilter·p1.00    sample               0.006             s/op
MyBenchmark.testStreamWithSingleFilter                                         sample  3292270     ≈ 10⁻⁵             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p0.00        sample              ≈ 10⁻⁵             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p0.50        sample              ≈ 10⁻⁵             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p0.90        sample              ≈ 10⁻⁵             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p0.95        sample              ≈ 10⁻⁵             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p0.99        sample              ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p0.999       sample              ≈ 10⁻⁴             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p0.9999      sample               0.001             s/op
MyBenchmark.testStreamWithSingleFilter:testStreamWithSingleFilter·p1.00        sample               0.011             s/op
MyBenchmark.testStreamWithMultipleFilter                                           ss       10      0.010 ±   0.001   s/op
MyBenchmark.testStreamWithSingleFilter                                             ss       10      0.009 ±   0.001   s/op

As you can see, the single filter took less time than using multiple ones.

Source code: here

The takeaway: Multiple filters have some overhead; make sure to write good filters.