[CALCITE-4199] Add nullability annotations to the methods and fields, ensure consistency with checkerframework - ASF JIRA

XML

Word

Printable

JSON

Details

Type: New Feature
Status: Closed
Priority: Major
Resolution: Fixed
Affects Version/s: 1.25.0
Fix Version/s: 1.27.0
Component/s: None
Labels:
- pull-request-available

Description

The current codebase uses jsr305 javax.annotation.Nullable / NonNull which helps to catch bugs while developing Calcite and libraries.

Unfortunately, the current annotation is not enforced, and it lacks support for generics.
jsr305.jar is probably abandoned (see https://github.com/google/guava/issues/2960), so we should probably migrate to something else.

https://checkerframework.org/ is a solid framework for machine verification which is tailored to Java.

The releases are quite frequent: https://github.com/typetools/checker-framework/releases
Their annotations are recognized by major IDEs.

So I see the following options:
a) Leave the code as is
b) Annotate (gradually?) the code with checkerframework annotations
c) Migrate (gradually?) the code to Kotlin

I've created a PR to verify which changes would be needed, verify if CI is able to type check in a reasonable time, and so on: https://github.com/apache/calcite/pull/1929

My findings regarding checkerframework so far are:
0) It does detect NPEs (which were hidden in the current code), and it does detect initialized files in the constructors
1) Checkerframework takes ~2 minutes for :linq4j, and 9min for :core in GitHub Actions CI
2) "non-nullable by default" is quite close to the current Calcite conventions.
3) There are cases when javadoc comments says "or null", however, the code reads much easier if @Nullable nullable appears in the signature
4) If a third-party library supplies invalid type annotations, there's a way to fix that. For instance, Guava's Function is annotated as "always nullable", and we can overrule that (so the nullability is taken from generic signature rather than "always nullable"). The override files are placed to src/main/config/checkerframework
5) Generic-heavy code might be challenging (they are always like that), however, in the most obscure cases there's always a way to suppress the warning
6) I've filed a Gradle improvement so it schedules recently modified files first for the compilation: https://github.com/gradle/gradle/issues/14332

—

Option a: leave the code as is

Typically, Calcite code does have decent documentation if null is permissible, however, I often need to check the existing calls in order to figure out if null is expected.

Note: nullability does not mean just "null vs non-null" value, but it is related to other common pitfalls like "failing to initialize the value in constructor", "calling non-final method in the constructor"

Option b: annotate (gradually?) the code with checkerframework annotations

Pros:

Checkerframework verifies one method at a time, so the resulting code is easy to reason about. For instance, it requires that boolean equals(@Nullable Object other) overloads must explicitly declare @Nulable annotation. That might sound too verbose at first, however, it helps to read the code: you don't need to analyze class/interface declarations in order to figure out if nulls are expected or not.
Checkerframework supports nullability with generics. For instance, class Wrapper<T> means T can be used as either non-nullable or nullable. class Wrapper<T> means T is always nullable (all implementations can assume T is always nullable), and class Wrapper<T extends @NonNull Object> (nonnull is redundant in this case and can be omitted) means T can't be null.
Checkerframework supports nullability with arrays. @Nullable Object[] is non-nullable array with nullable elements, Object @Nullable [] is nullable array with non-nullable elements.
The verifier sees untested code
checkerframework has a coherent set of verifiers, and it is pluggable
checkerframework compiler understand lots of different nullability annotations, so we could use jsr305 nullability annotations instead of the ones from checkerframework
Nullability annotations might make it easier to contribute: one has to reason about nulls anyway, and the annotations make it simpler.

Cons:

Nullability annotations add an extra burden on developers. Checkerframework documentation is good, however, it might be challenging to figure out the proper annotations especially when generics and/or inheritance is involved
Verification takes significant time, which is especially sad for local development. For instance, core verification takes ~10 minutes. Note: extra time is spent only in case checkerframework verification is activated, and the verification itself is NOT required for regular Calcite operation (e.g. development, testing, etc). The nullability verification is similar to code style checks like Checkstyle: it can be executed as a completely separate task.
IDEs do not have full support of checkerframework annotations. For instance, they can't infer nullability from generic signatures, so the code might look OK in IDE, and it would fail verification. At the same time, IDE might show false positives (warning in IDE while the verification passes)
Other languages might miss nullability inference from checkerframework annotations. For instance, Kotlin compiler can infer nullability from jsr305 annotations, however, it does not understand checkerframework yet
Checkerframework verifies one method at a time, so does not know if certain methods are used in a specific order only. One of the cases is Enumerator interface with current, moveNext, and close methods. Checker assumes the clients might call the methods in any order, so it enforces that null-checks must be placed in all the methods. It might sound like "adding verbosity to make checker happy"
Checkerframework is not suitable for test code, as it would take significant effort to write nullability annotations, and it would make test code much harder to write. For instance, checkerframework knows nothing on JUnit's @Before... annotations, so it assumes the fields are not initialized in the test methods
Nullability annotations might sound like adding a new language. For instance, the difference between <T> public T get(), <@Nullable T> public T get(), <T> public @Nullable T get(), <T extends @Nullable Object> public T get(), <@PolyNull T> public @PolyNull T get() might be not that easy to understand.
Java's type system is unsound, so it is hard to write code without suppression. For instance: Map.get(...) sometimes permits nulls (HashMap), and sometimes it does not (ImmutableMap). That is one of the reasons Map.get(...) and Map.remove(..) are annotated to require non-nullable argument.
Nullability annotations might make it harder to contribute: checkerframework might be unfamiliar to the contributors
Nullability annotations do not replace runtime checks, so we still need to use requireNonNull(...) to verify inputs in the runtime
There's a risk of human error while migrating the code: a wrong assertion might be added, so the code might start failing

Option b: migrate (gradually?) the code to Kotlin

Pros:

Better IDE support than the one of checkerframework

Standard library contains lots of collection manipulation code, so typical-for-Calcite Itarables.transform. Lists.transform, and so on would be easier to read and write. Streams in Java are still too verbose, especially for one-time mapping.
For instance,

    return columns.stream()
        .map(columnOrdinal -> table.getRowType().getFieldNames()
            .get(columnOrdinal))
        .collect(Collectors.toList());

becomes

    return columns.map { columnOrdinal ->
        table.getRowType().getFieldNames().get(columnOrdinal)
    }

or even

    return columns.map { table.getRowType().getFieldNames().get(it) }

Faster compilation times. Even though Kotlin is slower to compile than regular Java, Kotlin is way faster than checkerframework compiler. So edit-compile cycle would likely be much faster
Some type annotations would work for Kotlin as well. For instance, @CheckReturnValue would work in IDEs even for Kotlin code
It might attract contributors. StackOverflow 2020 survey suggests Kotlin is the 4th loved (and 6th wanted) programming language
Kotlin integrates with Java nicely: Java classes can extend Kotlin classes and vice versa
Kotlin is safe and pleasant to use for test code as well. For instance, lateinit var enables to have non-nullable reference and initialize it in @Before...
Migration to Kotlin can be fully backward compatible with Java (see OkHttp migration case)
Kotlin advanced deprecation features simplify migrations for the clients. For instance, @Deprecated annotation in Kotlin can suggest the replacements, so IDEs would suggest replacing the deprecations. With Java we have to guess the alternatives
Better "public API" support: Kotlin code fails to compile if a public class inherits from a private (or package-private) one.
Non-nullable arguments would automatically be verified, so hand-crafted Objects.requireNonNull(...) could be removed

Cons:

Kotlin might make it harder to contribute: Kotlin might be unfamiliar to the contributors
There's a risk of human error while migrating the code: a wrong assertion might be added, so the code might start failing
Kotlin standard library becomes an extra 1MiB dependency