- add suffixes
- adapt change-scala-version script

This closes #1529.

this commit support full outer join includes:
    1. left outer join with REPARTITION_HASH_FIRST.
    2. right outer join with REPARTITION_HASH_SECOND.
    3. fullouter join with REPARTITION_HASH_FIRST and REPARTITION_HASH_SECOND.

this close #1469

This closes #1539.

This closes #1550

This closes #1548

This closes #1522

- Unify usage of expression keys: FLINK-1666
- Unify supported key expressions (incl. support for partitioning on atomic types: FLINK-3233)
- Remove removal of duplicate keys: FLINK-1903
- Unify checks for sort keys
- Add more tests for SelectorFunctionKeys and ExpressionKeys

This closes #1520

This closes #1549.

This closes #1541

Only keep Time for specifying time durations/intervals.

This closes #1512

This closes #1547

This closes #1531

This closes #1530

This closes #1526

Before this commit, Kryo runs extra logic to track and resolve repeated references to
the same object (similar as JavaSerialization)

This disables reference tracking because
  - reference tracking is costly
  - it is virtually always unnecessary in the datatypes used in Flink
  - it is inconsistent with Flink's own serialization (which does not do reference tracking)
  - it may have problems if elements are read in a different order than they are written.

This closes #1528

[FLINK-3246] Remove unnecessary '-parent' suffix from projects 'flink-contrib' and 'flink-streaming-connectors'.

This closes #1515

This closes #1518

This closes #1545

Before, it could happen that a StreamingTask receives a Checkpoint
Trigger message while internally not being ready. The checkpoint
coordinator would then wait the specified timeout interval before
continuing. Now, tasks can signal that they are not ready and the
checkpoint coordinator will dicard a checkpoint for which is this the
case and trigger new checkpoints if necessary.

The newly triggered checkpoints will also release alignment locks in
streaming tasks that are still waiting for barriers from failed
checkpoints.

Only some of the LearningRateMethodTrait implementations use the decay value.
Thus, it should only be part of these methods. This commit makes the learning
rate methods which need additional parameters a case class which has to be
instantiated with the respective parameter values.

Corrects some of the learning rate formulas and the example in the documentation.

Added SGD gain calculation schemes

fixed optimal SGD calculation scheme

FLINK-1994: Added 4 new effective learning rates

[FLINK-1994] [ml] Add different gain calculation schemes to SGD

fixed long lines in GradientDescent.scala

[FLINK-1994][ml]Add different gain calculation schemes to SGD

[FLINK-1994][ml] Add different gain calculation schemes to SGD

[FLINK-1994][ml] Add different gain calculation schemes to SGD

Added SGD gain calculation schemes

fixed optimal SGD calculation scheme

[FLINK-1994] [ml] Add different gain calculation schemes to SGD

FLINK-1994: Added 3 new effective learning rates

Added SGD gain calculation schemes

fixed optimal SGD calculation scheme

[FLINK-1994] [ml] Add different gain calculation schemes to SGD

fixed long lines in GradientDescent.scala

[FLINK-1994][ml] Add different gain calculation schemes to SGD

[Flink-1994][ml] Add different gain calculation schemes to SGD

[FLINK-1994][ml] Updated docs, refactored optimizationMethod from Int to String

[FLINK-1994][ml] Added test and example to docs

[FLINK-1994][ml] Fixed Int Artifacts in LinearRegression.scala

Added LearningRateMethod to IterativeSolver

The learning rate method defines how the effective learning step is calculated for
each iteration step of the IterativeSolver.

Fixed docs, merged enumeration from Till, fixed typo in Wus method

[FLINK-1994][ml] Added 4 new effective learning rate methods

[FLINK-1994][ml] Add different gain calulation schemes to SGD

This closes #1397.

This closes #1535.

* updated algorithm and tests
* updated / corrected algorithm documentation

This closes #1521

For adding Kafka 0.9.0.0 support, this commit changes the following:
- Split up of the kafka connector into a flink-connector-kafka-(base|0.9|0.8) with different dependencies
- The base package contains common test cases, classes and implementations (the producer for 0.9 and 0.8 relies on exactly the same code)
- the 0.8 package contains a kafka connector implementation against the SimpleConsumer (low level) API of Kafka 0.8. There are some additional tests for the ZK offset committing
- The 0.9 package relies on the new Consumer API of Kafka 0.9.0.0
- Support for metrics for all producers and the 0.9 consumer through Flink's accumulators.

This closes #1489

The method WebMonitorUtils.startWebRuntimeMonitor instantiates the WebMonitor. However, in case of a failure null is returned. This failure case has to be handled in the method FlinkMiniCluster.startWebServer in order to avoid a possible NullPointerException.

Full changelog:

Major changes
===============
- Users no longer have to supply information about types
- Values are now stored as byte arrays on the Java side in
* a plain byte[] most of the time,
* a T2<b[],b[]> within a join/cross
* a T2<TX<b[]...b[]>, b[]> within keyed operations.
- Every value contains information about its type at the beginning of
each byte array.
- Implemented KeySelectors

Minor
===============
- improved error messages in several places
- defaultable operations now use a "usesUDF" flag
- reshuffled type ID's; tuple type encoded as 1-25
- broadcast variables are now sent via the tcp socket
- ProjectJoin/-Cross now executes projection on python side

Java
---------------
- Sort field now stored as String, continuation of FLINK-2431
- object->byte[] serializer code moved into separate utility class

Python
---------------
- Fixed NullSerializer not taking a read method argument
- Serializer/Deserializer interface added
- Refactored DataSet structure
* Set and ReduceSet merged into DataSet
- configure() now takes an OperationInfo argument
- Simplified GroupReduce tests
- removed unused Function._open()
- simplified chaining setup
- most functions now use super.configure()