This closes #3653.
This closes #3574.

This closes #3649.

Allows looping states (oneOrMore, zeroOrMore, times) to specify
if they want their elements to be consecutive or allow non-matching
elements in-between.

This closes #3629.

Replace the method via identical method 'getTaskNameWithSubtaskIndex'.

ExecutionGraph-based metrics should be in their own package 'org.apache.flink.runtime.executiongraph.metrics'.
They are instantiated by whoever builds the execution graph, but not by the execution graph itself.
This separates concerns more elegantly.

This closes #3636.
This closes #3637.

[FLINK-6034] [checkpoints] Introduce KeyedStateHandle abstraction for the snapshots in keyed streams

So far, the where clause only supported simple FilterFunction
conditions. With this, we add support for conditions where an
event is accepted not only based on its own properties, e.g.
name, as it was before, but also based on some statistic
computed over previously accepted events in the pattern, e.g.
if the price is higher than the average of the prices of the
previously accepted events.

Create interface for parsing command-line parameters using ParameterTool
and generic implementations for boolean, long, double, string, choice.

This closes #3433

This closes #3618.

This closes #3269.

This closes #3386.

This closes #3585.

[FLINK-6089] [table] Add decoration phase for stream queries to rewrite plans after the cost-based optimization.

This closes #3564.

This closes #3409.

Decrease the time to run TravisCI builds and reduce OOM by
- increasing the heap size
- switching to Java's serial garbage collector
- removing the UseGCOverheadLimit option

This closes #3571

The initial fix for this ticket is not working on larger data sets.

Reduce supports returning the left input, right input, a new object, or
a locally reused object. The trouble with the initial fix was that the
returned local object was reusing fields from the input tuples.

The problem is with ReduceDriver#run managing two values (reuse1 and
reuse2) and with a third, local value returned by
GatherSumApplyIteration.SumUDF. After the first grouping value.f1 ==
reuse1.f1. Following UDF calls may swap value.f1 and reuse2.f1, which
causes reuse1.f1 == reuse2.f1. With an odd number of swaps the next
grouping will reduce with reuse1 and reuse2 sharing a field and
deserialization will overwrite stored values.

The simple fix is to only use and return the provided inputs.

This closes #3515

This closes #3606.

Currently, the TestingSerialRpcService does not play well together with the
MainThreadValidatorUtil which assumes that rpc calls are dispatched to a mailbox
from where they are picked up. In order to support the TestingSerialRpcService
we will have to extend the MainThreadValidatorUtil to allow entering the main thread
if the RpcEndpoint already runs in the context of the main thread.

This commit also renames isCompatibleWith() to canRestoreFrom() to make
the method asymetric because in the case of KryoSerializer we
can restore from state that was stored using no registed
types/serializers while the other way around is not possible.

This closes #3534.
This closes #3603.

This closes #3587.

The existing WindowOperatorContractTest is turned into a test base while
RegularWindowOperatorContract test tests WindowOperator and
EvictingWindowOperatorTest tests EvictingWindowOperator. For this to
work, the base tests now always use List windows and we have specific
tests for reducing/folding windows in RegularWindowOperatorContractTest.

This also patches in the missing side output support for
EvictingWindowOperator.

This PR properly shuts down the query actor of the MetricRegistry upon shut down.

Add locking to the MetricRegistry

This closes #3573.

This closes #3591.