2.15.0-rc1 regression: transient fields are serialized

[emlun]

Describe the bug

In 2.14.2 and earlier, transient fields (with or without @JsonIgnore) with a getter (without @JsonIgnore) are ignored during serialization and do not emit a JSON property.

In 2.15.0-rc1 and 2.15.0-rc2, they do emit a JSON property.

Example

Given this data class:

public class Foo1 {
    private final Integer a;
    @JsonIgnore private final transient Integer minusA;

    @JsonCreator
    public Foo1(@JsonProperty("a") Integer a) {
        this.a = a;
        this.minusA = a == null ? null : -a;
    }

    public Integer getA() { return a; }
    public Integer getMinusA() { return minusA; }
}

and this ObjectMapper:

ObjectMapper json =
    new ObjectMapper()
        .configure(DeserializationFeature.FAIL_ON_UNKNOWN_PROPERTIES, true)
        .configure(SerializationFeature.FAIL_ON_EMPTY_BEANS, false)
        .setSerializationInclusion(Include.NON_ABSENT);

Jackson 2.14.2 serializes a Foo1 instance to {"a":0} while Jackson 2.15.0-rc1 serializes the same value to {"a":0,"minusA":0}.

Version information

Issue is not present in 2.14.2.

Issue is present in 2.15.0-rc1.

To Reproduce

1. Check out commit f40ff286 in Yubico/java-webauthn-server (branch jackson-regression-2-15-rc1 - might not persist after this issue is closed).
2. Run: ./gradlew :test-success-2-14:test. Observe that the test succeeds.
3. Run: ./gradlew :test-fail-2-15:test. Observe that the test fails.
4. Run: ./gradlew :test-success-2-14:dependencies. Observe that the resolved Jackson version is 2.14.2.
5. Run: ./gradlew :test-fail-2-15:dependencies. Observe that the resolved Jackson version is 2.15.0-rc1.
6. Run: git diff --no-index lib/src/main/java/org/example/Foo1.java lib/src/main/java/org/example/Foo2.java. Observe that the only (relevant) difference is that Foo1.minusA is transient while Foo2.minusA is not.

After running the above, you can see the failed test report at test-fail-2-15/build/reports/tests/test/index.html.

Additional context

One might run into this when using Lombok's @Value annotation on classes with transient fields - even if the field has @JsonIgnore, the generated getter does not inherit the @JsonIgnore annotation.

[pjfanning]

this is possibly because of public Integer getMinusA() { return minusA; } -- fields that are exposed when serializing can include getters

• you could try adding @JsonIgnore to the getter method
• you could also try setting the visibility settings on the mapper to disable the use of getter methods - https://www.javadoc.io/static/com.fasterxml.jackson.core/jackson-databind/2.15.0-rc2/com/fasterxml/jackson/databind/ObjectMapper.html#setDefaultVisibility-com.fasterxml.jackson.annotation.JsonAutoDetect.Value-

[cowtowncoder]

I think this is a side effect of some change related to pruning of transient fields during processing.

First of all: note that the reason 2.14 and earlier did not include minusA is NOT due to transient marker but rather @JsonIgnore that is used to remove "whole" property minusA (both field and getter).
Why this isn't happening 2.15.0-rc1/rc2 is unclear but I suspect Field is dropped first and linkage not established to drop getter as well.

But there are at least 2 changes that should make things work in a way that'd work in both 2.15 and 2.14 (and earlier):

1. Enable MapperFeature.PROPAGATE_TRANSIENT_MARKER and leave out @JsonIgnore from field
2. Move @JsonIgnore on getter; not sure why it is added to Field in the first place (getter is what would be used)

I suggest doing one of above; changes are behavior of 2.15 will remain the way rc2 has it (due to timing of things) so that is safer set up.

[emlun]

Yeah, it would make sense that this would have changed if @JsonIgnore is processed before removing transient fields in 2.14, but that order is reversed in 2.15.

[...] changes are behavior of 2.15 will remain the way rc2 has it [...]

Ok, I guess this was an unintentional "feature". Issue #296 appears to also deal with this. I'll admit I didn't look up whether Jackson promises anything about transient fields before I added the modifier. The docs for PROPAGATE_TRANSIENT_MARKER for 2.9 (the version used at the time) do indeed say (emphasis added):

Feature is disabled by default, meaning that existence of transient for a field does not necessarily lead to ignoral of getters or setters but just ignoring the use of field for access.

So yeah, looks like this is a "regression" of a feature that was not promised. Thanks for the replies!

[cowtowncoder]

@emlun Yes -- ideally behavior would not have changed, but it is bit of an edge case. I do consider it a regression but given how complicated the logic in that part is, it may be difficult to return handling of such cases.
transient is interesting in its own right; as well as what is intended to be relationship between fields, getters, setters etc etc.

But I hope your use case works with some modifications. Thank you for follow up!