[ https://issues.apache.org/jira/browse/FLINK-9487?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16511139#comment-16511139 ]
ASF GitHub Bot commented on FLINK-9487: --------------------------------------- GitHub user StefanRRichter opened a pull request: https://github.com/apache/flink/pull/6159 [FLINK-9487] Prepare InternalTimerHeap for asynchronous snapshots ## What is the purpose of the change This PR is the first step in the context of FLINK-9485. Purpose of this PR is to enhance ``InternalTimerHeap`` with the capability to produce asynchronous snapshots. This is very similar to the approach of asynchronous snapshots for the ``CopyOnWriteStateTable`` and we want to reuse as much of the existing code as possible to power both instances of snapshots. ## Brief change log The first commit generalizes the key-group-partitioning algorithm for async snapshots from the ``CopyOnWriteStateTable``. The newly introduced ``StateSnapshot`` interface outlines the asynchronous snapshot life-cycle, which typically looks as follows. In the synchronous part of a checkpoint, an instance of {StateSnapshot} is produced for a state and captures the state at this point in time. Then, in the asynchronous part of the checkpoint, the user calls `` #partitionByKeyGroup()`` to ensure that the snapshot is partitioned into key-groups. For state that is already partitioned, this can be a NOP. The returned ``KeyGroupPartitionedSnapshot`` can be used by the caller to write the state by key-group. As a last step, when the state is completely written, the user calls ``#release()``. The partitioning algorithm is also slightly modified to cache computed key-group-ids per element. This is improves runtime at the cost of some additional memory. The second commit introduced an implementation of ``StateSnapshot`` for the ``InternalTimerHeap`` data structure. ## Verifying this change This change added tests: ``TimerPartitionerTest`` , ``StateTableKeyGroupPartitionerTest``, ``AbstractKeyGroupPartitionedSnapshotTest`` ## Does this pull request potentially affect one of the following parts: - Dependencies (does it add or upgrade a dependency): (no) - The public API, i.e., is any changed class annotated with `@Public(Evolving)`: (no) - The serializers: (no) - The runtime per-record code paths (performance sensitive): (no) - Anything that affects deployment or recovery: JobManager (and its components), Checkpointing, Yarn/Mesos, ZooKeeper: (yes) - The S3 file system connector: (no) ## Documentation - Does this pull request introduce a new feature? (yes) - If yes, how is the feature documented? (not applicable) You can merge this pull request into a Git repository by running: $ git pull https://github.com/StefanRRichter/flink FLINK-9487 Alternatively you can review and apply these changes as the patch at: https://github.com/apache/flink/pull/6159.patch To close this pull request, make a commit to your master/trunk branch with (at least) the following in the commit message: This closes #6159 ---- commit 9e95e1a5f36b6c4f0b021f38b2a2a4c2f8dacf3a Author: Stefan Richter <s.richter@...> Date: 2018-06-11T12:48:06Z Refactor/generalize key-group partitioning. commit 800abd9743c5483f50835d0cf938ad50f550ae49 Author: Stefan Richter <s.richter@...> Date: 2018-06-11T15:44:10Z Introduce TimerHeap snapshots and key-group-partitioning ---- > Prepare InternalTimerHeap for asynchronous snapshots > ---------------------------------------------------- > > Key: FLINK-9487 > URL: https://issues.apache.org/jira/browse/FLINK-9487 > Project: Flink > Issue Type: Sub-task > Components: State Backends, Checkpointing, Streaming > Reporter: Stefan Richter > Assignee: Stefan Richter > Priority: Major > Fix For: 1.6.0 > > > When we want to snapshot timers with the keyed backend state, this must > happen as part of an asynchronous snapshot. > The data structure {{InternalTimerHeap}} needs to offer support for this > through a lightweight copy mechanism (e.g. arraycopy of the timer queue, > because timers are immutable w.r.t. serialization). > We can also stop keeping the dedup maps in {{InternalTimerHeap}} separated by > key-group, all timers can go into one map. > Instead, we can implement online-partitioning as part of the asynchronous > operation, similar to what we do in {{CopyOnWriteStateTable}} snapshots. > Notice that in this intermediate state, the code will still run in the > synchronous part until we are integrated with the backends for async > snapshotting (next subtask of this jira). -- This message was sent by Atlassian JIRA (v7.6.3#76005)