See [] for details, but basically, @FlowScope causes the Flow instance produced by flowProvider.get() in FlowRunner to be the same each time it's called, which leads to the instance being re-used when a transactional retry (e.g. for a ConcurrentModificationException) causes a flow to be attempted more than once. Flow is not meant to be re-used and certain flows fail at runtime when this happens, so the effect is that a CME now aborts most EPP requests, which is bad.
This is a bit of a hacky fix; finding a better one is tracked in []
== TESTING ==
This is very hard to test because there isn't really a clean way to trigger a CME from within a flow's execution without hardcoding in assumptions about what a given flow is doing when it runs, and we can't easily supply a custom Flow for testing while also exercising the Flow daggerization process (since this bug only appears due to the specific way that dagger constructs the Provider<Flow>). Ideally a fix would improve the testability here as well.
For now, I've manually tested this change by pasting code into FlowRunner that explicitly throws a ConcurrentModificationException after running the flow (similar to DryRunException), but only on the first transaction attempt. With @FlowScope on provideFlow(), this change reproduces the UnsupportedOperationException issue in many tests; once it's removed (i.e. with this CL submitted) the problem goes away.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=126226066
This is more efficient (we were constructing all the
providers and the map anew for every flow) and prettier.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=125479328
Daggerizes all of the EPP flows. This does not change anything yet
about the flows themselves, just how they are invoked, but after
this CL it's safe to @Inject things into flow classes.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=125382478
