- 26 10月, 2017 6 次提交
-
-
由 yew1eb 提交于
This closes #4661.
-
由 kkloudas 提交于
This was revealed through the continuous file reader. Previously the StreamTask was calling the quiesceAndAwaitPending() of the TimerService before the close() of the operator. This meant that with a periodic watermark emitter and a small file (e.g. one split), the timer service would be closed before even starting to read (as soon as the reader received the first split), and no timers would be registered to emit watermarks.
-
由 Fabian Hueske 提交于
- Fixes [FLINK-7542] "AggregateITCase fails in different timezone." This closes #4796.
-
由 Fabian Hueske 提交于
-
由 twalthr 提交于
This closes #4199.
-
由 Till Rohrmann 提交于
-
- 25 10月, 2017 19 次提交
-
-
由 Aljoscha Krettek 提交于
-
由 Aljoscha Krettek 提交于
Before, when setting classloader.resolve-order Flink classes where also resolved via the child ClassLoader which could lead to problems if the user-jar contains copies of Flink classes. This change introduces a new setting classloader.parent-first-patterns. Classes that match this pattern will always be resolved by the parent first, even when using the child-first ClassLoader.
-
由 minwenjun 提交于
This closes #4665.
-
由 Tzu-Li (Gordon) Tai 提交于
-
由 Tzu-Li (Gordon) Tai 提交于
-
由 Tzu-Li (Gordon) Tai 提交于
Prior to this commit, there is no flushing of KPL outstanding records on checkpoints in the FlinkKinesisProducer. Likewise to the at-least-once issue on the Flink Kafka producer before, this may lead to data loss if there are asynchronous failing records after a checkpoint which the records was part of was completed. This closes #4871.
-
由 gyao 提交于
This closes #4856.
-
由 Piotr Nowojski 提交于
-
由 Fabian Hueske 提交于
This closes #4813.
-
由 Xingcan Cui 提交于
This closes #4874.
-
由 Fabian Hueske 提交于
- Fixes also [FLINK-5498] Add support for non-equi join and local predicates to outer joins This closes #4858. This closes #3379 (stale PR for FLINK-5498).
-
由 zentol 提交于
This closes #4836
-
由 Mike Kobit 提交于
This closes #4854
-
由 yew1eb 提交于
This closes #4885
-
由 Till Rohrmann 提交于
-
由 Till Rohrmann 提交于
[FLINK-6046] Either store serialized value or permanent blob key in ExecutionGraph and ExecutionJobVertex
-
由 Till Rohrmann 提交于
-
由 Nico Kruber 提交于
This only includes potentially big parts, i.e. serializedJobInformation and serializedTaskInformation, which are both offloaded only once for all parallel instances. - adds a configurable akka.rpc.offload.minsize threshold for large data - serialized task information is uploaded only once per task, irrespective of the parallelism [FLINK-6046][tests] added an integration test with a job with big payload This verifies that uploading a job with a payload of 100MB is successful. NOTE: This only works after also implementing JobGraph offloading during submission or akka.framesize will already limit it when sending it form the client to the jobmanager. This closes #4412. Set min offloading size to 1 MiB
-
由 Nico Kruber 提交于
-
- 24 10月, 2017 15 次提交
-
-
由 Till Rohrmann 提交于
-
由 Till Rohrmann 提交于
-
由 Bowen Li 提交于
-
由 Aljoscha Krettek 提交于
Timer callbacks on Trigger are not called if the window is empty, they are silently ignored. This has led to some confusion.
-
由 Piotr Nowojski 提交于
It depends whether to call hsync or hflush on the underlying file system and user preferences. Normally hflush is enough to protect against single machine HDFS failures and against TaskManagers failures. However if user is using S3 like file system, or wants to protect againt whole HDFS rack power loss hsync must be used instead.
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
-
由 Aljoscha Krettek 提交于
Now, checkpoints are only discarded when they are subsumed. The earlier behaviour was causing problems when the checkpoint store had transient outages.
-
由 Bowen Li 提交于
-
由 zentol 提交于
This closes #4849.
-
由 Aljoscha Krettek 提交于
-
由 Stephan Ewen 提交于
-
由 Stephan Ewen 提交于
-