The change improves the display of SerializedListAccumulators in the web
interface by displaying the number of elements contained in the
accumulator instead of the hash code of the object.

This closes #1717.

This closes #1712.

The start computation states have a null event associated. When serializing these states
one has to check whether the event is null or not, because not all serializer can handle
null values.

A single state NFA failed to compute a matching pattern because it was regarded as a
terminal state in the pattern extraction algorithm. By letting the first states start
with a two level dewey number this problem is fixed.

After restart of the execution graph, the test driver waits for the
graph to assign the new executions in a loop. If switchToRunning()
is called multiple times in this loop, the job is canceled. This
could happen, if one execution had its resource assigned and called
switchToRunning(), but the other didn't. In this case, the complete
loop was retried and switchToRunning() was called again.

This closes #1708.

This closes #1706

  - IngestionTime now only auto-generates watermarks
  - EventTime does not auto-generate ingestion watermarks

This also removes the "ExecutionConfig.areTimestampsEnabled()" flag.

This closes #1699

- Move "Windowing" to separate page
  - Make "Working with Time" sub page of "Windowing"
  - Fix link in "Working with Time" page
- Move "Working with State" page to "Fault Tolerance"
  - Add Table of Contents to State Backends page
  - Hierarchical restart strategies sections
- CEP => Event Processing (CEP)

This closes #1705.

This closes #1707.

This closes #1701

This closes #1697

This closes #1581

This changes slot sharing settings to single method
slotSharingGroup(String) on DataStream.

Operations inherit the slot sharing group of the input if all input
operations are in the same slot sharing group.

The default slot sharing group is "default" this can also be explicitly
set using slotSharingGroup("default"). This overrides the inheriting
behaviour.

This closes #1641.

removed unused import

use new Path().toUri()

added missing braces

This closes #1651.

This was causing to many problems with security tokens and yarn. Now,
let the RocksDB backup run in a thread but
don't interrupt these Threads anymore on closing. The Threads will close
themselves because the copy operation will fail because of a
FileNotFoundException when the state directories are being cleaned up.

This also removes the ExternalProcessRunner because it is not needed
anymore and using it causes too many headaches.

This closes #1687.

This closes #1652.

This closes #1688.

Add test case

This closes #1643.

This closes #1666

[FLINK-3438] ExternalProcessRunner fails to detect ClassNotFound exception because of locale settings

[FLINK-3438] Improved solution, no workaround

[FLINK-3438] Change: a faulty process now causes a RuntimeException to be thrown

This closes #1665.

[FLINK-3460] [build] Set Flink dependencies in flink-streaming-connectors, flink-batch-connectors, cep, gelly and flink-ml modules to provided

The flink-streaming-connectors all depend on flink-streaming-java in compile scope.
This entails that this dependency is always pulled in, when you build a fat jar. By
setting this dependency to provided, this will be avoided. Furthermore, the connectors
are always used in conjunction with flink-streaming-java. This means that you will
always have an explicit dependency import of flink-streaming-java in your build script.
This allows you to run your program locally but it is also easy to exclude the dependency
from being included in the fat jar by setting it to provided, as well.

This closes #1683.

[FLINK-3459] [build] Fix conflicting dependencies commons-collections, commons-beanutils and commons-beanutils-core

The Hadoop dependencies have a dependency on commons-configuration which pulls in transitively the commons-collection, commons-beanutils and common-beanutils-core depedencies. Commons-beanutils and commons-collection contain classes which live in the same namespace. They are also binary compatible but not binary identical. This is a problem for the sbt assembly plugin which checks for binary identity. In order to solve the problem, we bump the commons-configuration version to 1.7 so that only commons-beanutils is pulled in. This is necessary, because the transitive promotion of dependencies of the shade plugin only excludes the commons-beanutils dependency only from the directly depending dependency. All parent dependencies won't have the exclusion. This is a problem for SBT which will pull the dependency as part of one of the parents, then. Moreover, we replace commons-beanutils by commons-beanutils-bean-collections which contains only the non-conflicting classes wrt commons-collections.

This closes #1682.

The shade-flink execution of the parent pom caused the problem that the guava
dependencies were relocated twice in the flink-shaded-hadoop jar. In order to
avoid this, this patch disables the shade-flink execution.

This closes #1681.

This closes #1674.

Adds further log statements in order to improve debuggability of
JobManagerHAJobGraphRecoveryITCase.

Hope to help for debugging:
- hpick://s3.amazonaws.com/archive.travis-ci.org/jobs/110095304/log.txt
- https://s3.amazonaws.com/archive.travis-ci.org/jobs/110085371/log.txt

This closes #1692

This closes #1588

This closes #1680

This closes #1684

The StateDescriptor can be serializer asynchronously in case of
asynchronous checkpoints. In that case two threads would try to
concurrently use the TypeSerializer: The normal state updating and the
checkpoint serialization. If the TypeSerializer is a KryoSerializer this
can lead to problems. Therefore the need to duplicate it before using in
"writeObject".