Disposing savepoints via the JobManager fails for state handles or descriptors,
which contain user classes (for example custom folding state or RocksDB handles).

With this change, the user can provide the job JAR when disposing a savepoint in
order to use the user code class loader of that job. The JAR is optional, hence
not breaking the CLI API.

This closes #2083.

This solves issues when a host has multiple IPs

Before, ValueState.update() was not called after changing the NFA or the
priority queue in CEP operators. This means that the operators don't
work with state backends such as RocksDB that strictly require that
update() be called when state changes.

This changes the operators to always call update() and also introduces a
test that verifies the changes.

Before, these were using StreamRecordSerializer, which does not serialize
timestamps. Now it uses MultiplexingStreamRecordSerializer.

This also extends the tests in CEPOperatorTest to test that timestamps
are correctly checkpointed/restored.

NFA is Serializable and has readObject()/writeObject() methods. In
AbstractKeyedCEPPatternOperator a KryoSerializer was used as the
TypeSerializer for the ValueState that holds NFA instances. Kryo does
not call readObject()/writeObject() therefore the state of the NFA was
invalid after deserialization.

This change adds a new TypeSerializer for NFA that uses
Java Serialization. In the long run it will be better to get rid of the
readObject()/writeObject() methods and instead efficiently serialize
using a specialized TypeSerializer.

State handles are cancelable, to make sure long running checkpoint restore operations do
finish early on cancallation, even if the code does not properly react to interrupts.

This is especially important since HDFS client code is so buggy that it deadlocks when
interrupted without closing.

This commit also adds monitoring for the current and committed offset

This closes #2236

The GraphiteReporter and GangliaReporter report metric names which can contain invalid
characters. These characters include quotes and dots. In order to properly report metrics
to these systems, the afore-mentioned characters have to be replaced in metric names.

The PR also removes quotes from the garbage collector metric name.

The PR sets the default value for TTL in the GangliaReporter to 1, because -1 causes the
reporter to fail.

Introduce CharacterFilter to filter invalid characters from the metric name out

The character filter is applied to all components of the fully qualified metric name.
The ScheduledDropwizardReporter and AbstractReporter implement this interface to generate
compatible metric names.

Correct AbstractMetricGroup.getMetricIdentifier; Add test cases to check that reporters filter out invalid characters

This closes #2220.

Assembles and applies a GroupReduceFunction using pre-defined
AggregationOperations in Python. References to aggregations in
PythonOperationInfo and other Java classes in the Python API
removed since aggregations are now handled by Python.

This closes #2115

This closes #2128

This closes #2186

Flink does not include the anchor.js file but loads it dynamically when displaying
the documentation. Therefore, we don't have to include anchor.js in the LICENSE file.

This closes #2255

This closes #2135

This closes #2254

This closes #2246
This closes #2247

This closes #2242

The 'recoveryTimestamp' was an unsafe wall clock timestamp attached by the master
upon recovery. This this timestamp cannot be relied upon in distributed setups,
it is removed.

This closes #2228

This closes #2237

This closes #1517 (2/2)

This closes #1517 (1/2)

This closes #2234

This closes #2227

This adds log messages about created result partition consumers and
spilling.

Introduce ProxyMetricGroup which forwards metric registration calls to its parent.
The IOMetricGroup then extends ProxyMetricGroup so that all metric registrations
are executed on the parent of IOMetricGroup.

This closes #2210.

This closes #2235

Include 'json' as a required dependency for building the documentation.