[
https://issues.apache.org/jira/browse/FLINK-6363?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Tzu-Li (Gordon) Tai updated FLINK-6363:
---------------------------------------
Description:
Currently, there is no documentation / Javadoc mentioning the precedence rules
of Kryo registrations via the register methods in
{{StreamExecutionEnvironment}} / {{ExecutionEnvironment}}.
It is important for the user to be notified of the precedence because the
{{KryoSerializer}} applies the configurations in a specific order that is not
visible from the public API.
For example:
{code}
env.addDefaultKryoSerializer(SomeClass.class, SerializerA.class);
env.addDefaultKryoSerializer(SomeClass.class, new SerializerB());
{code}
from this API usage, it may seem as if {{SerializerA}} will be used as the
default serializer for {{SomeClass}} (or the other way around, depends really
on how the user perceives this).
However, whatever the called order in this example, {{SerializerB}} will always
be used because in the case of defining default serializers, due to the
ordering that the internal {{KryoSerializer}} applies these configurations,
defining default serializer by instance has a higher precedence than defining
by class. Since the existence of this precedence is not due to Kryo's
behaviour, but due to the applied ordering in {{KryoSerializer}}, even users
that are familiar with Kryo will be surprised with unexpected results.
These methods are also subject to the same issue:
{code}
env.registerType(SomeClass.class, SerializerA.class);
env.registerTypeWithKryoSerializer(SomeClass.class, SerializerA.class);
env.registerTypeWithKryoSerializer(SomeClass.class, new SerializerB());
{code}
was:
Currently, there is no documentation / Javadoc mentioning the precedence rules
of Kryo registrations via the register methods in
{{StreamExecutionEnvironment}} / {{ExecutionEnvironment}}.
It is important for the user to be notified of the precedence because the
{{KryoSerializer}} applies the configurations in a specific order that is not
visible from the public API.
For example:
{code}
env.addDefaultKryoSerializer(SomeClass.class, SerializerA.class);
env.addDefaultKryoSerializer(SomeClass.class, new SerializerB());
{code}
from this API usage, it may seem as if {{SerializerA}} will be used as the
default serializer for {{SomeClass}} (or the other way around, depends really
on how the user perceives this).
However, whatever the called order in this example, {{SerializerB}} will always
be used because in the case of defining default serializers, due to the
ordering that the internal {{KryoSerializer}} applies these configurations,
defining default serializer by instance has a higher precedence than defining
by class.
These methods are also subject to the same issue:
{code}
env.registerType(SomeClass.class, SerializerA.class);
env.registerTypeWithKryoSerializer(SomeClass.class, SerializerA.class);
env.registerTypeWithKryoSerializer(SomeClass.class, new SerializerB());
{code}
> Document precedence rules of Kryo serializer registrations
> ----------------------------------------------------------
>
> Key: FLINK-6363
> URL: https://issues.apache.org/jira/browse/FLINK-6363
> Project: Flink
> Issue Type: Improvement
> Components: Documentation, Type Serialization System
> Reporter: Tzu-Li (Gordon) Tai
> Assignee: Tzu-Li (Gordon) Tai
>
> Currently, there is no documentation / Javadoc mentioning the precedence
> rules of Kryo registrations via the register methods in
> {{StreamExecutionEnvironment}} / {{ExecutionEnvironment}}.
> It is important for the user to be notified of the precedence because the
> {{KryoSerializer}} applies the configurations in a specific order that is not
> visible from the public API.
> For example:
> {code}
> env.addDefaultKryoSerializer(SomeClass.class, SerializerA.class);
> env.addDefaultKryoSerializer(SomeClass.class, new SerializerB());
> {code}
> from this API usage, it may seem as if {{SerializerA}} will be used as the
> default serializer for {{SomeClass}} (or the other way around, depends really
> on how the user perceives this).
> However, whatever the called order in this example, {{SerializerB}} will
> always be used because in the case of defining default serializers, due to
> the ordering that the internal {{KryoSerializer}} applies these
> configurations, defining default serializer by instance has a higher
> precedence than defining by class. Since the existence of this precedence is
> not due to Kryo's behaviour, but due to the applied ordering in
> {{KryoSerializer}}, even users that are familiar with Kryo will be surprised
> with unexpected results.
> These methods are also subject to the same issue:
> {code}
> env.registerType(SomeClass.class, SerializerA.class);
> env.registerTypeWithKryoSerializer(SomeClass.class, SerializerA.class);
> env.registerTypeWithKryoSerializer(SomeClass.class, new SerializerB());
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
