KAFKA-12357: Do not inline methods from the scala package by default (#10174)

As mentioned in #9548, users currently use the kafka jar (`core` module)
for integration testing and the current inlining behavior causes
problems when the user's classpath contains a different Scala version
than the one that was used for compilation (e.g. 2.13.4 versus 2.13.3).

An example error:

`java.lang.NoClassDefFoundError: scala/math/Ordering$$anon$7`

We now disable inlining of the `scala` package by default, but make it
easy to enable it for those who so desire (a good option if you can
ensure the scala library version matches the one used for compilation).
While at it, we make it possible to disable scala compiler optimizations
(`none`) or to use only method local optimizations (`method`). This can
be useful if optimizing for compilation time during development.

Verified behavior by running gradlew with `--debug` and checking the
output after `[zinc] The Scala compiler is invoked with:`

Reviewers: Chia-Ping Tsai <chia7712@gmail.com>
This commit is contained in:
Ismael Juma 2021-02-22 21:11:17 -08:00 committed by GitHub
parent da1d9da7ea
commit b5265e98eb
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 25 additions and 14 deletions

View File

@ -221,6 +221,13 @@ The following options should be set with a `-P` switch, for example `./gradlew -
* `enableTestCoverage`: enables test coverage plugins and tasks, including bytecode enhancement of classes required to track said * `enableTestCoverage`: enables test coverage plugins and tasks, including bytecode enhancement of classes required to track said
coverage. Note that this introduces some overhead when running tests and hence why it's disabled by default (the overhead coverage. Note that this introduces some overhead when running tests and hence why it's disabled by default (the overhead
varies, but 15-20% is a reasonable estimate). varies, but 15-20% is a reasonable estimate).
* `scalaOptimizerMode`: configures the optimizing behavior of the scala compiler, the value should be one of `none`, `method`, `inline-kafka` or
`inline-scala` (the default is `inline-kafka`). `none` is the scala compiler default, which only eliminates unreachable code. `method` also
includes method-local optimizations. `inline-kafka` adds inlining of methods within the kafka packages. Finally, `inline-scala` also
includes inlining of methods within the scala library (which avoids lambda allocations for methods like `Option.exists`). `inline-scala` is
only safe if the Scala library version is the same at compile time and runtime. Since we cannot guarantee this for all cases (for example, users
may depend on the kafka jar for integration tests where they may include a scala library with a different version), we don't enable it by
default. See https://www.lightbend.com/blog/scala-inliner-optimizer for more details.
### Dependency Analysis ### ### Dependency Analysis ###

View File

@ -134,6 +134,12 @@ ext {
userEnableTestCoverage = project.hasProperty("enableTestCoverage") ? enableTestCoverage : false userEnableTestCoverage = project.hasProperty("enableTestCoverage") ? enableTestCoverage : false
// See README.md for details on this option and the reasoning for the default
userScalaOptimizerMode = project.hasProperty("scalaOptimizerMode") ? scalaOptimizerMode : "inline-kafka"
def scalaOptimizerValues = ["none", "method", "inline-kafka", "inline-scala"]
if (!scalaOptimizerValues.contains(userScalaOptimizerMode))
throw new GradleException("Unexpected value for scalaOptimizerMode property. Expected one of $scalaOptimizerValues), but received: $userScalaOptimizerMode")
generatedDocsDir = new File("${project.rootDir}/docs/generated") generatedDocsDir = new File("${project.rootDir}/docs/generated")
commitId = project.hasProperty('commitId') ? commitId : null commitId = project.hasProperty('commitId') ? commitId : null
@ -515,21 +521,19 @@ subprojects {
"-Xlint:unused" "-Xlint:unused"
] ]
// Inline more aggressively when compiling the `core` jar since it's not meant to be used as a library. // See README.md for details on this option and the meaning of each value
// More specifically, inline classes from the Scala library so that we can inline methods like `Option.exists` if (userScalaOptimizerMode.equals("method"))
// and avoid lambda allocations. This is only safe if the Scala library version is the same at compile time scalaCompileOptions.additionalParameters += ["-opt:l:method"]
// and runtime. We cannot guarantee this for libraries like kafka streams, so only inline classes from the else if (userScalaOptimizerMode.startsWith("inline-")) {
// Kafka project in that case. List<String> inlineFrom = ["-opt-inline-from:org.apache.kafka.**"]
List<String> inlineFrom if (project.name.equals('core'))
if (project.name.equals('core')) inlineFrom.add("-opt-inline-from:kafka.**")
inlineFrom = ["-opt-inline-from:scala.**", "-opt-inline-from:kafka.**", "-opt-inline-from:org.apache.kafka.**"] if (userScalaOptimizerMode.equals("inline-scala"))
else inlineFrom.add("-opt-inline-from:scala.**")
inlineFrom = ["-opt-inline-from:org.apache.kafka.**"]
// Somewhat confusingly, `-opt:l:inline` enables all optimizations. `inlineFrom` configures what can be inlined. scalaCompileOptions.additionalParameters += ["-opt:l:inline"]
// See https://www.lightbend.com/blog/scala-inliner-optimizer for more information about the optimizer. scalaCompileOptions.additionalParameters += inlineFrom
scalaCompileOptions.additionalParameters += ["-opt:l:inline"] }
scalaCompileOptions.additionalParameters += inlineFrom
if (versions.baseScala != '2.12') { if (versions.baseScala != '2.12') {
scalaCompileOptions.additionalParameters += ["-opt-warnings", "-Xlint:strict-unsealed-patmat"] scalaCompileOptions.additionalParameters += ["-opt-warnings", "-Xlint:strict-unsealed-patmat"]