- 18 9月, 2017 2 次提交
- 16 9月, 2017 4 次提交
-
-
由 zjureel 提交于
This closes #4562
-
由 yew1eb 提交于
This closes #4676
-
由 yew1eb 提交于
This closes #4575
-
由 Michael Fong 提交于
Program output is written to TaskManager logs. This closes #4641
-
- 15 9月, 2017 2 次提交
-
-
由 Greg Hogan 提交于
The Simplify parameter should accept and set the parallelism when calling the Simplify algorithms. The LocalClusteringCoefficient "count triangles" reduce now uses the assigned ("little") parallelism as this computation is proportional to the number of vertices (the combine computation is proportional to the potentially much larger number of triangles). The ignored CombineHint on the HITS all-reduces have been removed. This closes #4346
-
由 Greg Hogan 提交于
There exist some tests with empty graphs but the `EmptyGraph` in `AsmTestBase` contained vertices but no edges. Add a new `EmptyGraph` without vertices and test both empty graphs for each algorithm. EmptyGraph now generates the proper TypeInformation (for Edge<> not Tuple3) which had prevented adding edges due to a union incompatibility. GraphGeneratorUtils#vertexSet now uses a hash-combine for distinct. `PageRank` optionally includes zero-degree vertices in the results (at a performance cost). This closes #4405
-
- 14 9月, 2017 2 次提交
-
-
由 JingsongLi 提交于
[FLINK-7617] [runtime] Remove string format in BitSet to improve the performance of BuildSideOuterjoin This closes #4668.
-
由 bbayani 提交于
This closes #4622.
-
- 07 9月, 2017 6 次提交
-
-
由 Tzu-Li (Gordon) Tai 提交于
- Improve deprecation message in Javadocs - Remove usage of ProducerConfigConstants in code wherever possible - Remove usage of ProducerConfigConstants in documentation code snippets This closes #4473.
-
由 Bowen Li 提交于
-
由 Tzu-Li (Gordon) Tai 提交于
[FLINK-7440] [kinesis] Eagerly check that provided schema and partitioner is serializable in FlinkKinesisProducer This commit also adds a test to verify that the FlinkKinesisProducer is serializable. This closes #4537.
-
由 Tzu-Li (Gordon) Tai 提交于
[FLINK-7440] [kinesis] Eagerly check serializability of deserialization schema in FlinkKinesisConsumer This commit also adds tests for verifying that the FlinkKinesisConsumer itself is serializable.
-
由 Tzu-Li (Gordon) Tai 提交于
This closes #4526.
-
由 Tzu-Li (Gordon) Tai 提交于
Previously, the AbstractPartitionDiscoverer tracked discovered partitions by keeping only the largest discovered partition id. All fetched partition metadata with ids smaller than this id would be considered as discovered. This assumption of contiguous partition ids is too naive for corner cases where there may be undiscovered partitions that were temporariliy unavilable before and were shadowed by discoverered partitions with largerer partition ids. This commit changes to use a set to track seen partitions. This also removes the need of pre-sorting fetched partitions.
-
- 06 9月, 2017 3 次提交
-
-
由 Piotr Nowojski 提交于
-
由 Till Rohrmann 提交于
Closing StreamOperators is still part of the StreamTask's running lifecycle, because operators which perform asynchronous operations usually finish their work when the StreamOperator is closed. Since this also entails that errors can occur and that a checkpointing operation is triggered, we should only set the StreamTask's isRunning to false after all StreamOperators have been closed. Furthermore, this commit introduces a while guard for the waiting condition in ContinuousFileReaderOperator#close. This closes #4650.
-
由 Till Rohrmann 提交于
FutureUtils.retryWithDelay executes the given operation of type Callable<CompletableFuture<T>> n times and waits in between retries the given delay. This allows to retry an operation with a specified delay. Make retry and retry with delay future properly cancellable This closes #4637.
-
- 05 9月, 2017 12 次提交
-
-
由 Aljoscha Krettek 提交于
-
由 Aljoscha Krettek 提交于
-
由 Aljoscha Krettek 提交于
-
由 Aljoscha Krettek 提交于
-
由 Aljoscha Krettek 提交于
-
由 Usman Younas 提交于
This closes #4608.
-
由 Fabian Hueske 提交于
This closes #4507.
-
由 sunjincheng121 提交于
This closes #4640.
-
由 Xingcan Cui 提交于
This closes #4633.
-
由 shaoxuan-wang 提交于
This closes #4546.
-
由 Till Rohrmann 提交于
This commit lets the JobMaster extend from FencedRpcEndpoint. This enables automatic fencing of all messages. Moreover, this PR introduces the JobMasterId which is the new leader id/fencing token replacing the UUID. This improves type safety when passing multiple fencing tokens around. This closes #4583.
-
由 Till Rohrmann 提交于
This commit waits for the completion of the shutdown of the ResourceManager before shutting down the ResourceManagerRuntimeServices. The JobManagerServices are now exclusively passed in to the JobManagerRunner which means that it is no longer responsible for shutting the JobManagerServices down. Additionally, it waits until the JobMaster has been shut down before closing the LeaderElectionService as well as the JobManagerMetricGroup. The JobManagerServices are now managed by the caller of the JobManagerRunner. This allows to reuse them across multiple JobManagerRunners. The RpcEndpoint#postStop method is now called by the UntypedActor#postStop method, which ensures that the RpcEndpoint's method is also called if only the underlying RpcService is shut down (without explicitly shutting down the RpcEndpoint). This closes #4596.
-
- 04 9月, 2017 9 次提交
-
-
由 Till Rohrmann 提交于
Properly fences the ResourceManager by letting it extend the FencedRpcEndpoint. Moreover, this PR introduces a ResourceManagerId which replaces the UUID as leader id/fencing token. This will give us more type safety when defining rpc interfaces. This closes #4582.
-
由 Till Rohrmann 提交于
Let the Dispatcher extend the FencedRpcEndpoint and introduce DispatcherId which replaces the UUID as leader id/fencing token. This closes #4584.
-
由 Till Rohrmann 提交于
This commit filters out duplicate JobManager gained leadership messges coming from the JobLeaderService. This avoid opening multiple connections to the JobManager which consumes resources. Moreover, this commit properly closes all JobManagerConnections in case of a shut down. This closes #4595.
-
由 Till Rohrmann 提交于
This commit changes the JobMaster such that it's leader session id is only set in the main thread. This is done by passing the leader session id to the startJobExecution method. Moreover, this commit returns a future from the start and suspend methods which can be used to wait on the completion of these operations. This closes #4579.
-
由 Till Rohrmann 提交于
Make all external calls from the RpcEndpoint's main thread non blocking by executing them as a runnable in an Executor. Make FatalErrorHandler calls non asynchronous This closes #4540.
-
由 Till Rohrmann 提交于
This commit changes the behaviour of the WebRuntimeMonitor to not longer block serving threads by waiting on the result of futures. Instead the RequestHandler now returns a CompletableFuture<FullHttpResponse> which is written out to the Netty channel upon completion. This will improve the performance of our WebRuntimeMonitor. This closes #4527.
-
由 Till Rohrmann 提交于
Introduce FencedRpcEndpoint which requires all RPC messages to have a fencing token attached. Based on the received fencing token and the actual fencing token, the message will either be discarded if they are not equal or it will be processed. That way we are able to filter out old messages or messages which originate from a split brain situation Add support for callAsyncWithoutFencing Introduce local and remote fenced messages This closes #4578.
-
由 Till Rohrmann 提交于
Enrich the RestClientException with the HttpResponseStatus in case of failures. This makes the exception handling on the client side easier. This closes #4588.
-
由 Till Rohrmann 提交于
The termination future is completed when the ClusterEntrypoint shuts down. This allows for easier testing. This closes #4589.
-