Remove unnecessary error generation when JsUndefined.asOpt is used

[Dubinka94]

Pull Request Checklist

• Have you read through the contributor guidelines?
• Have you squashed your commits?
• Have you added copyright headers to new files?
• Have you updated the documentation?
• Have you added tests for any changed functionality?

Purpose

We discovered that in many cases, when JSON is parsed, asOpt is used to read an optional key. This results in a performance penalty because an error is generated each time, but it is not used in any way, as described in the documentation for asOpt.

When asOpt is used on a missing key, it is called on the JsUndefined object. This triggers the execution of the asOpt function in JsReadable, which performs unnecessary validation and ultimately returns None after constructing a large number of unnecessary strings.

An example of this issue can be seen in the following flame graph:

A simple fix would be to always return None when asOpt is called on the JsUndefined object.

We tested the performance of the function on version 3.0.4 witch Scala 2.13.14 and Java 21 with and without the fix, and the results are as follows:

Original version

[info] # Warmup Iteration   1: 16656.253 ops/ms
[info] # Warmup Iteration   2: 16971.747 ops/ms
[info] # Warmup Iteration   3: 17315.722 ops/ms
[info] # Warmup Iteration   4: 17300.985 ops/ms
[info] # Warmup Iteration   5: 17355.655 ops/ms
[info] Iteration   1: 17318.713 ops/ms
[info] Iteration   2: 17324.471 ops/ms
[info] Iteration   3: 17366.688 ops/ms
[info] Iteration   4: 17339.576 ops/ms
[info] Iteration   5: 17317.734 ops/ms
[info] Result "play.api.libs.json.AsOptJsonBenchmark.readNonExistentField":
[info]   17333.436 ±(99.9%) 79.080 ops/ms [Average]
[info]   (min, avg, max) = (17317.734, 17333.436, 17366.688), stdev = 20.537
[info]   CI (99.9%): [17254.356, 17412.517] (assumes normal distribution)
[info] # Run complete. Total time: 00:01:40
[info] Benchmark                                 Mode  Cnt      Score    Error   Units
[info] AsOptJsonBenchmark.readNonExistentField  thrpt    5  17333.436 ± 79.080  ops/ms

Fixed version

[info] # Warmup Iteration   1: 387805.387 ops/ms
[info] # Warmup Iteration   2: 395776.593 ops/ms
[info] # Warmup Iteration   3: 396521.421 ops/ms
[info] # Warmup Iteration   4: 396692.480 ops/ms
[info] # Warmup Iteration   5: 396071.469 ops/ms
[info] Iteration   1: 395516.846 ops/ms
[info] Iteration   2: 396267.266 ops/ms
[info] Iteration   3: 396287.878 ops/ms
[info] Iteration   4: 396425.990 ops/ms
[info] Iteration   5: 396137.772 ops/ms
[info] Result "play.api.libs.json.AsOptJsonBenchmark.readNonExistentField":
[info]   396127.150 ±(99.9%) 1371.385 ops/ms [Average]
[info]   (min, avg, max) = (395516.846, 396127.150, 396425.990), stdev = 356.144
[info]   CI (99.9%): [394755.766, 397498.535] (assumes normal distribution)
[info] # Run complete. Total time: 00:01:40
[info] Benchmark                                 Mode  Cnt       Score      Error   Units
[info] AsOptJsonBenchmark.readNonExistentField  thrpt    5  396127.150 ± 1371.385  ops/ms

As can easily be seen, the throughput increase is significant, and in our experience, in real-world applications, usually 70% of the time spent in asOpt will be saved in applications that frequently try to access missing keys.

The following code was used to test the throughput:

package play.api.libs.json

import org.openjdk.jmh.annotations._
import org.openjdk.jmh.infra.Blackhole
import java.util.concurrent.TimeUnit

@BenchmarkMode(Array(Mode.Throughput))
@OutputTimeUnit(TimeUnit.MILLISECONDS)
@State(Scope.Thread)
@Fork(value = 1)
class AsOptJsonBenchmark {
  var json: JsObject = _

  @Setup(Level.Trial)
  def prepare(): Unit = {
    json = Json.obj("existingField" -> "value")
  }

  @Benchmark
  def readNonExistentField(blackhole: Blackhole): Unit = {
    blackhole.consume((json \ "nonExistentField").asOpt[String])
  }
}

[cchantep]

👍 Make sure it's test covered

[Dubinka94]

What kind of test would you like me to add?
Testing that asOpt works as expected for both valid and invalid keys is already covered by other tests. Benchmarking to ensure performance stays within a certain threshold might be unreliable due to different test environments.
And I don’t have access to the intermediate object created inside asOpt to verify whether it was created or not in case of the older implementation.

[cchantep]

There is no direct test of JsUndefined whereas it can be easily done.

[cchantep]

revert