- 27 4月, 2017 9 次提交
-
-
由 Nico Kruber 提交于
This closes #3348.
-
由 Nico Kruber 提交于
Previous commit in e908b62ab3 caused a regression leading to unit test failures.
-
由 Nico Kruber 提交于
The metrics group given to a Task must always be non-null otherwise the code would have crashed anyway. Similarly, the group returned by getIOMetricGroup() is always present.
-
由 Stephan Ewen 提交于
These metrics are optimised go go through the channels only once in order to gather all metrics, i.e. min, max, avg and sum. Whenever a request to either of those is made, all metrics are refreshed and cached. Requests to the other metrics will be served from the cache. However, each value will be served only once from the cache and a second call to retrieve the minimum, for example, will refresh the cache for all values. This setup may at first be a bit strange but ensures that the statistics belong together logically and originate from a common point in time. This is not necessarily the point in time the metric was requested though.
-
由 manuzhang 提交于
-
由 manuzhang 提交于
And also the same bug in WindowOperatorMigrationTest
-
由 Till Rohrmann 提交于
The problem is that we don't wait until the JobManager becomes the leader. Due to this, the sent RegisterTaskManager messages might get dropped. This PR fixes the problem by waiting on the completion of the NotifyWhenLeader message.
-
由 Stephan Ewen 提交于
This includes - A interface for hooks that are called by the checkpoint coordinator to trigger/restore a checkpoint - A source extension that triggers the operator checkpoints and barrier injection on certain events Because this changes the checkpoint metadata format, the commit introduces a new metadata format version. This closes #3782
-
由 Stefan Richter 提交于
-
- 26 4月, 2017 19 次提交
-
-
由 Greg Hogan 提交于
Replace division and modulus with addition and subtraction. This closes #2628
-
由 twalthr 提交于
This closes #3663.
-
由 Till Rohrmann 提交于
The PythonStreamer used to check for the existence of the python binary by starting a python process. This process was not closed afterwards. This caused the PythonPlanBinderTest to fail locally. I think the check whether a python binary exists is not necessary since the subsequent python command would fail anyway if there is no binary available on the system. The system failure message is that there is no such file or directory. This error message should be descriptive enough in order to debug such a problem. This closes #3774.
-
由 kl0u 提交于
By default, the times() quantifier operates in eager mode and the user can disable eagerness by calling .allowCombinations().
-
由 kl0u 提交于
Allow all patterns to be optional in a pattern sequence. Singleton becomes simply optional, times(X) transforms from "exactly X" to "0 or X", and "one or more" becomes "zero or more".
-
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 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 Aljoscha Krettek 提交于
This is based on the Apache Beam checkstyle with some custom modifications for making it fit the exisiting Flink code base. The most notable change is that Flink uses Tabs for indentation while Beam uses spaces. This adds a lot of checks that are commented out because they require non-trivial changes to the code base. There are some trivial code changes for good practices checkstyle rules that where only broken in very few places.
-
由 Greg Hogan 提交于
-
由 Till Rohrmann 提交于
An endpoint id allows to assign a unique name to a RpcEndpoint. This name can be used to look up the rpc endpoint within the RpcService. Remove component endpoint name methods from the HighAvailabilityServices Remove JobMaster endpoint id This closes #3596.
-
由 Till Rohrmann 提交于
The problem was that the started JobManagers try to instantiate a MetricQueryService actor with a fixed name. Thus, if the previous JobManager was not shutdown properly, then there was still a MetricQueryServie actor around having the exact same name with which you try to start a new actor. The problem is solved twofold: 1. by properly stopping the instantiated actors 2. by starting the JobManager without a MetricsRegistry
-
- 25 4月, 2017 3 次提交
-
-
由 xccui 提交于
This closes #3768.
-
由 shaoxuan-wang 提交于
This closes #3762.
-
由 sunjincheng121 提交于
This closes #3743.
-
- 24 4月, 2017 1 次提交
-
-
由 Greg Hogan 提交于
Evaluate user-defined Java options immediately before starting services and rotate all log files. This closes #3701
-
- 22 4月, 2017 6 次提交
-
-
由 Stefan Richter 提交于
-
由 xiaogang.sxg 提交于
-
由 Ken Geis 提交于
This closes #3754.
-
由 shaoxuan-wang 提交于
This closes #3735.
-
由 godfreyhe 提交于
This closes #3739.
-
由 shijinkui 提交于
This closes #2460.
-
- 21 4月, 2017 2 次提交
-
-
由 Nico Kruber 提交于
This way, if all slaves refer to "localhost", we do not require ssh at all. This closes #3298
-
由 rami-alisawi 提交于
This closes #3657
-