ProgramLifecycle methods should have better control over transactions

Update to the Design: It would be better to use a more reusable annotation that will be able to express other aspects of transactions in the future. Instead of @NoTransaction, we will have an annotation @TransactionPolicy(TransactionControl.EXPLICIT).

In the future we can extend this, for example, to allow specifying the transaction timeout, too, or to allow program types (such as workers) that have explicit transaction control by default, to annotate @TransactionPolicy(TransactionControl.IMPLICIT).

Addition: Nesting transactions will not be allowed and throw an exception.

Design

Currently all life cycle methods (initialize, destroy) are executed in a (short) transaction. But:

• some methods do not require a transaction (or it could even get in the way by causing tx timeouts)

• some methods need a longer timeout, or they need to start multiple transactions

This can be done implementing the Transactional interface with the program context, which will allow these methods to control their own transactions and their timeout. At the same time, we need a way to declare that this method does not require an implicit transaction. That can be done using a new annotation @NoTransaction. Then, for example, an initialize() method could look like this:

This will apply to life cycle methods of flowlets, service handlers, workers, mapreduce, spark, workflow, but not to task-level life cycle methods in mapreduce or spark (they are executed in the context of the program's long tx).