My understanding is that you can only remove the exports by either a) removing all usages or b) migrating the project to modules and declare the JDK dependencies that way.

a) is not viable AFAIK and in at least one case would require dropping features (the JMX reporter), while b) appears to be a massive task; modularization for larger projects is not trivial, it does not play nicely with the shade-plugin (i.e., our entire packaging paradigm) and can bring additional headaches when also trying to be compatible with Java 8.

I don't believe this to be a big problem though to be honest; it seems unlikely that this "temporary workaround" is really that temporary.

Sooo....sure you can create a JIRA ticket, but I do not see anything we can change at the moment.

On 01/07/2020 17:31, Niels Basjes wrote:
Hi,

I've been fiddling around to try to see if this really works when compiling with Java 14. I have the simplest form of this fix working and I've put up a pull request for just that. https://github.com/apache/flink/pull/12799 I also found that there are two different values for java.version that make the build needlessly hard to debug (for which I created https://issues.apache.org/jira/browse/FLINK-18458 )

What I found is that building under Java 9+ uses a construct that is not 'clean'.
The build is not how it is intended by the Java developers.

What I found today:

When I simply change this (add the release):
                 <plugin>
                   <groupId>org.apache.maven.plugins</groupId>
                   <artifactId>maven-compiler-plugin</artifactId>
                   <configuration>
<!-- <source>11</source>--> <!-- <target>11</target>--> <release>11</release>
                      <compilerArgs combine.children="append">
                         
<arg>--add-exports=java.base/sun.net.util=ALL-UNNAMED</arg>
                         
<arg>--add-exports=java.management/sun.management=ALL-UNNAMED</arg>
                         
<arg>--add-exports=java.rmi/sun.rmi.registry=ALL-UNNAMED</arg>
                         
<arg>--add-exports=java.security.jgss/sun.security.krb5=ALL-UNNAMED</arg>
y</compilerArgs>
                   </configuration>
                </plugin>

I get

    *[ERROR] exporting a package from system module java.base is not
    allowed with --release*
    *[ERROR] exporting a package from system module java.management is
    not allowed with --release*
    *[ERROR] exporting a package from system module java.rmi is not
    allowed with --release*
    *[ERROR] exporting a package from system module java.security.jgss
    is not allowed with --release*


When I remove the --add-exports statements I get

    *[ERROR]
    
/home/nbasjes/workspace/Apache/flink/flink-core/src/main/java/org/apache/flink/util/NetUtils.java:[26,20]
    package sun.net.util does not exist*


According to the migration guides the --add-exports is seen as a temporary workaround

  * 
https://docs.oracle.com/javase/9/migrate/toc.htm#JSMIG-GUID-77874D97-46F3-4DB5-85E4-2ACB5F8D760B
  * Quote:
      o Critical internal JDK APIs such as sun.misc.Unsafe are still
        accessible in JDK 9, but most of the JDK’s internal APIs are
        not accessible at compile time. You may get compilation errors
        that indicate that your application or its libraries are
        dependent on internal APIs.
      o ...
      o You may use the --add-exports option as a temporary workaround
        to compile source code with references to JDK internal classes.

and thus is now no longer allowed in combination with the release flag.

  * 
https://stackoverflow.com/questions/45370178/exporting-a-package-from-system-module-is-not-allowed-with-release
  * https://bugs.openjdk.java.net/browse/JDK-8178152


Q: Is there an existing ticket for making the build 'clean' for java 9+ ? Or shall I create one?

Niels Basjes



On Wed, Jul 1, 2020 at 10:57 AM Chesnay Schepler <ches...@apache.org <mailto:ches...@apache.org>> wrote:

    oh cool, yes that should work nicely then.

    On 01/07/2020 10:50, Niels Basjes wrote:
    Hi,

    > This is difficult to do since you cannot have multiple jdk
    activations for a single profile in maven

    Well actually you can have multiple versions of JDK activate the
    java11 profile.
    http://maven.apache.org/guides/introduction/introduction-to-profiles.html

    quote:

     1. <profile>
     2. <activation>
     3. <jdk>[1.3,1.6)</jdk>
     4. </activation>



    So I actually had something really simple in mind:
    Have the Java11 profile activate on anything greater or equal to
    java 11 (i.e. also on java 12,13,14, ...)
    And then force it to build java11 code.

    Something like this (untested)

       <profile>
            <id>java11</id>
            <activation>
                <jdk>[11,)</jdk>
            </activation>

    ...
    <configuration>
        <source>11</source>
        <target>11</target>
        <release>11</release>

    Niels  Basjes




    On Tue, Jun 30, 2020 at 7:53 PM Chesnay Schepler
    <ches...@apache.org <mailto:ches...@apache.org>> wrote:

        This is difficult to do since you cannot have multiple jdk
        activations for a single profile in maven, and duplicating
        the entire profile for all jdk versions isn't an option.

        We _maybe_ could invert the behavior such that the Java 11
        behavior is the default, with a JDK 8 profile, but there may
        be some areas where this wouldn't work (e.g., adding
        dependencies in JDK 11), and it may catch developers off-guard.

        On 30/06/2020 17:57, Niels Basjes wrote:
        Hi Chesnay,

        Ok, so if someone uses a non-LTS version of Java (like 14)
        then how about simply "pinning" it to the Java 11 compatibility?
        I'm assuming no one uses Java 9 and/or 10 anymore so I'm
        ignoring those.
        Then building with Java 8 will result in Java 8 code.
        Building with Java 11, 12, 13, 14, ... will result in Java
        11 code.

        That way the code generated using Java 14 will fail
        immediately on Java 8 because of a completely
        incompatible binary format.
        To me this error would make a lot more sense (I would
        immediately know what I was doing should not work) than the
        strange error about the non-existent method.

        Do you agree?

        Niels




        On Tue, Jun 30, 2020 at 1:15 PM Chesnay Schepler
        <ches...@apache.org <mailto:ches...@apache.org>> wrote:

            > What is the Java version Apache Flink is supposed to
            work with?

            8 and 11. Non-LTS Java11+ releases _should work_, but we
            don't put in
            effort to make it as convenient as for LTS releases. As
            such you have to
            manually enable the java11 profile when compiling Flink.

            I set the target version to 11 since IIRC we ran into
            more errors that
            way, ensuring a smoother transition once Java 11 is the
            default.

            On 30/06/2020 13:01, Niels Basjes wrote:
            > Hi,
            >
            > I have both JDK 8 and 14 on my system and yesterday I
            ran into this
            > exception (I put the info I have in this ticket
            > https://issues.apache.org/jira/browse/FLINK-18455 ) :
            >
            >         java.lang.NoSuchMethodError:
            > java.nio.ByteBuffer.position(I)Ljava/nio/ByteBuffer;
            >
            > >From digging around (
            >
            
https://stackoverflow.com/questions/61267495/exception-in-thread-main-java-lang-nosuchmethoderror-java-nio-bytebuffer-flip
            > )
            > it seems this is caused when using JDK 9+ without
            setting '-release 8' and
            > then running with JDK 8.
            >
            > Essentially there are two solutions I see a lot:
            > 1) Add the -release 8 flag to the compiler
            > 2) Use the flaky workaround to cast all problem cases
            to the superclass
            > implementing the method (i.e. cast to java.nio.Buffer)
            >
            > Looking at the actual Flink code I found that
            >
            > In a JDK 8 build both source and target are set to Java 8
            >
            
https://github.com/apache/flink/blob/d735d8cd8e5d9fae5322001099097581822453ae/pom.xml#L109
            >  <java.version>1.8</java.version>
            >
            
https://github.com/apache/flink/blob/d735d8cd8e5d9fae5322001099097581822453ae/pom.xml#L115
            >
            <maven.compiler.source>${java.version}</maven.compiler.source>
            >
            <maven.compiler.target>${java.version}</maven.compiler.target>
            >
            > In a JDK 11 build a profile is activated that
            overrides it to Java 11
            >
            
https://github.com/apache/flink/blob/d735d8cd8e5d9fae5322001099097581822453ae/pom.xml#L938
            >    <profile>
            >       <id>java11</id>
            >       <activation>
            >           <jdk>11</jdk>
            >
            
https://github.com/apache/flink/blob/d735d8cd8e5d9fae5322001099097581822453ae/pom.xml#L1004
            > <artifactId>maven-compiler-plugin</artifactId>
            >        <configuration>
            > <source>11</source>
            > <target>11</target>
            >
            > So when building with Java 11 the output classes are
            Java 11 compatible
            > binaries.
            >
            > However I have Java 14 (and the 'java11' profile is
            only activated at the
            > EXACT version of java 11) so it stays at source and
            target 1.8 but does not
            > specify the "release 8" setting. ... which causes the
            problems I see.
            >
            > Looking at the current build settings I was puzzled a
            bit and I have this
            > main question:
            >
            >      What is the Java version Apache Flink is supposed
            to work with?
            >
            > Currently I would expect Java 8.
            >
            > So what I propose for this problem as a fix is to set
            source, target and
            > release to Java 8 for all compiler versions (i.e. also
            Java 9, 11, 14, ...).
            > That way you can use any compiler and get the correct
            results.
            >
            > I also am curious if that would fix the tests that
            seem to fail under Java
            > 11.
            >
            > What do you think is the correct approach for this?
            >



-- Best regards / Met vriendelijke groeten,

        Niels Basjes




-- Best regards / Met vriendelijke groeten,

    Niels Basjes




--
Best regards / Met vriendelijke groeten,

Niels Basjes


Reply via email to