Description
This is an idea for an improvement to the SpecificCompiler-generated record classes. There are two main issues to address:
- Default values specified in schemas are only used at read time, not when writing/serializing records. For example, a NullPointerException is thrown when attempting to write a record that has an uninitialized array or string type. I'm sure this was done for good reasons, like giving users maximum control and preventing unnecessary garbage collection, but I think it's also somewhat confusing and unintuitive for new users (myself included).
- Users have to create their own factory classes/methods for every record type, both to ensure that all non-primitive members are initialized and to facilitate the construction and initialization of record instances (i.e. constructing and setting values in a single statement).
These issues have been discussed previously here:
I'd like to propose a solution that is used by at least one other messaging framework. For each generated record class there will be a public static inner class called Builder. The Builder inner class has the same fields as the record class, as well as accessors and mutators for each of these fields. Whenever a mutator method is called, the Builder sets a boolean flag indicating that the field has been set. All mutators return a reference to 'this', so it's possible to chain a series of setter invocations, which makes it really easy to construct records in a single statement. The Builder also has a build() method which constructs a record instance using the values that were set in the Builder. When the build() method is invoked, if there are any fields that have not been set but have default values as defined in the schema, the Builder will set the values of these fields using their defaults.
One nice thing about implementing the builder pattern in a static inner Builder class rather than in the record itself is that this enhancement will be completely backwards-compatible with existing code. The record class itself would not change, and the public fields would still be there, so existing code would still work. Users would have the option to use the Builder or continue constructing records manually. Eventually the public fields could be phased out, and the record would be made immutable. All changes would have to be done through the Builder.
Here is an example of what this might look like:
// Person.newBuilder() returns a new Person.Builder instance // All Person.Builder setters return 'this' allowing us to chain set calls together for convenience // Person.Builder.build() returns a Person instance after setting any uninitialized values that have defaults Person me = Person.newBuilder().setName("James").setCountry("US").setState("MA").build(); // We still have direct access to Person's members, so the records are backwards-compatible me.state = "CA"; // Person has accessor methods now so that the public fields can be phased out later System.out.println(me.getState()); // No NPE here because the array<Person> field that stores this person's friends has been automatically // initialized by the Builder to a new java.util.ArrayList<Person> due to a @java_class annotation in the IDL System.out.println(me.getFriends().size());
What do people think about this approach? Any other ideas?
Attachments
Attachments
Issue Links
- is related to
-
AVRO-951 Records with field named "data" collide with new builder code from specific compiler
- Closed
-
AVRO-846 Error record classes do not allow you to set the exception message or cause on construction
- Closed
-
AVRO-770 Codegen enhancement per comments from Avro-726
- Open
-
AVRO-726 Make GenericDatumReader/GenericDatumWriter data member protected so that it can be used by the derived classes
- Closed
-
AVRO-784 SpecificCompiler should generate accessors
- Closed
- relates to
-
PIG-2463 AvroStorage doesn't work with Avro 1.6.0
- Open
-
AVRO-993 Add methods back to GenericDatumReader that were removed in AVRO-839
- Closed
-
AVRO-985 Remove byte-by-byte copying in RecordBuilderBase.defaultValue for non-complex types
- Open