Vladislav Pernin created FLINK-6061:
---------------------------------------

             Summary: NPE on TypeSerializer.serialize with a 
RocksDBStateBackend calling state.entries in the open() function
                 Key: FLINK-6061
                 URL: https://issues.apache.org/jira/browse/FLINK-6061
             Project: Flink
          Issue Type: Bug
          Components: DataStream API, State Backends, Checkpointing, Streaming
    Affects Versions: 1.3.0
            Reporter: Vladislav Pernin


With a default state (heap), the call to state.entries() "nicely fails" with a 
IllegalStateException :
{noformat}
Caused by: java.lang.IllegalStateException: No key set.
        at 
org.apache.flink.util.Preconditions.checkState(Preconditions.java:195)
        at 
org.apache.flink.runtime.state.heap.HeapMapState.entries(HeapMapState.java:188)
        at 
org.apache.flink.runtime.state.UserFacingMapState.entries(UserFacingMapState.java:77)
        at 
org.apache.flink.Reproducer$FailingMapWithState.open(Reproducer.java:78)
        at 
org.apache.flink.api.common.functions.util.FunctionUtils.openFunction(FunctionUtils.java:36)
        at 
org.apache.flink.streaming.api.operators.AbstractUdfStreamOperator.open(AbstractUdfStreamOperator.java:112)
        at 
org.apache.flink.streaming.runtime.tasks.StreamTask.openAllOperators(StreamTask.java:376)
        at 
org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:252)
        at org.apache.flink.runtime.taskmanager.Task.run(Task.java:670)
        at java.lang.Thread.run(Thread.java:745)
{noformat}

With a RocksDBStateBackend, it fails with a NPE :
{noformat}
Caused by: java.lang.NullPointerException
        at 
org.apache.flink.api.common.typeutils.base.ShortSerializer.serialize(ShortSerializer.java:64)
        at 
org.apache.flink.api.common.typeutils.base.ShortSerializer.serialize(ShortSerializer.java:27)
        at 
org.apache.flink.contrib.streaming.state.AbstractRocksDBState.writeKey(AbstractRocksDBState.java:181)
        at 
org.apache.flink.contrib.streaming.state.AbstractRocksDBState.writeKeyWithGroupAndNamespace(AbstractRocksDBState.java:163)
        at 
org.apache.flink.contrib.streaming.state.AbstractRocksDBState.writeCurrentKeyWithGroupAndNamespace(AbstractRocksDBState.java:148)
        at 
org.apache.flink.contrib.streaming.state.RocksDBMapState.serializeCurrentKeyAndNamespace(RocksDBMapState.java:263)
        at 
org.apache.flink.contrib.streaming.state.RocksDBMapState.iterator(RocksDBMapState.java:196)
        at 
org.apache.flink.contrib.streaming.state.RocksDBMapState.entries(RocksDBMapState.java:143)
        at 
org.apache.flink.runtime.state.UserFacingMapState.entries(UserFacingMapState.java:77)
        at 
org.apache.flink.Reproducer$FailingMapWithState.open(Reproducer.java:78)
        at 
org.apache.flink.api.common.functions.util.FunctionUtils.openFunction(FunctionUtils.java:36)
        at 
org.apache.flink.streaming.api.operators.AbstractUdfStreamOperator.open(AbstractUdfStreamOperator.java:112)
        at 
org.apache.flink.streaming.runtime.tasks.StreamTask.openAllOperators(StreamTask.java:376)
        at 
org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:252)
        at org.apache.flink.runtime.taskmanager.Task.run(Task.java:670)
        at java.lang.Thread.run(Thread.java:745)
{noformat}

The reason is that the record is null, because backend.getCurrentKey() is null 
(not yet set) in AbstractRocksDBState.
This may also be the case for other RockDBState implementations.

You can find the reproducer here based on 1.3-SNAPSHOT (needed for the 
MapState) :
https://github.com/vpernin/flink-rocksdbstate-npe

The reproducer is a non sense application. There is no MapState with TTL or 
expiration yet, so the goal is to try to shrink or expire the state at some 
interval.
This could be done by iterating over the entries of the state and removing some 
of them.

This could probably not be done in the open() method of a rich function.
I also tried to implement CheckpointListener and to access the state content in 
notifyCheckpointComplete() method, but it fails to, I guess due to the 
asynchronous nature of the checkpoint.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Reply via email to