You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@flink.apache.org by "Aljoscha Krettek (JIRA)" <ji...@apache.org> on 2016/01/05 12:27:39 UTC

[jira] [Created] (FLINK-3201) Enhance Partitioned State Interface with State Types

Aljoscha Krettek created FLINK-3201:
---------------------------------------

             Summary: Enhance Partitioned State Interface with State Types
                 Key: FLINK-3201
                 URL: https://issues.apache.org/jira/browse/FLINK-3201
             Project: Flink
          Issue Type: Sub-task
          Components: Streaming
    Affects Versions: 1.0.0
            Reporter: Aljoscha Krettek
            Assignee: Aljoscha Krettek


We should enhance the partitioned state with different state types, so that the system knows about the semantics of the State. I propose for now:
- ValueState, this behaves like the current OperatorState: state is one value that can be set and retrieved
- ListState, state is a list that can be appended to and iterated over
- ReducingState, state is one value that other values can be added to

ListState and ReducingState would share a common superclass to allow them to be used in the same places. For example, the WindowOperator would use ReducingState and ListState interchangeably, depending on whether we have a ReduceFunction or not.

These additions allow the system to be clever about how state is checkpointed in the future. Think ManageMemory/Out-of-core state and incremental checkpoints.

Also, state should be scoped to both a key and a namespace. This will allow the WindowOperator to use the interface. Right now, WindowOperator has a custom state implementation that uses a two-level Map (by key and by window). In this case the window would be the namespace.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)