Operator Window and other changes

[akarnokd]

Operator Window with no-first-loss functionality.

Issue #1060

(I had to leave my computer so the detailed description will come within a few hours.)

[cloudbees-pull-request-builder]

RxJava-pull-requests #1051 SUCCESS
This pull request looks good

[akarnokd]

Changes in respect to OperatorWindow:

• Theoriginal buffered all events and emitted them at the end of windows as a single hot Observable. This version, similar to groupBy, uses a subject that buffers values until a client subscribes to a window, replays the buffered content, then relays the rest of the events as they appear from the source. This behavior allows code expecting the old buffer-all mode to continue functioning, but such unrevokable behavior might not be desired in the library.
• Fix for OperatorBufferWithTime: unsubscribe on OnCompleted to stop the associated periodic task.
• The rewritten exact operators (i.e., count == skip, timespan == timeshift) make extra efforts to ensure values end up in a window even if a window change happens concurrently. A sideeffect is that emission may hop threads, but are still ensured to apear serially.
• Expanded BufferUntilSubscriber to behave as a regular ReplaySubject for the second, etc. Subscribers instead of either failing with ClassCastException or stealing/overtaking the first client's events.

[benjchristensen]

BufferUntilSubscriber should not support multicasting, we're just recreating the "time gap" or memory leak problems it was intended to solve.

If it behaves as a PublishSubject then subsequent subscribers will be missing events.
If it behaves as a ReplaySubject then it's a memory leak.

If someone wants to multicast they need to ask for it and deal with ConnectableObservable.

[akarnokd]

I won't be able to change this PR until next monday. If the other behaviors of the new windows are okay, I'd suggest a partial merge leaving my BufferUntilSubscriber out and fixing its ClassCastException and stealing/overtaking behavior in a separate PR.

[benjchristensen]

suggest a partial merge leaving my BufferUntilSubscriber out and fixing its ClassCastException and stealing/overtaking behavior in a separate PR.

Agreed. I can fix the ClassCastException and just make it onError if someone tries to subscribe twice.

[DavidMGross]

Under what circumstances (if any) is buffer/window supposed to emit an
empty buffer/Observable -- in other words, if nothing is emitted by the
source Observable during a span of time when any emissions would be
allocated to a new buffer/Observable, will buffer/window emit nothing or
will it emit an empty buffer/Observable? I made some guesss when I came up
with my marble diagrams, but I'm not sure if those guesses still hold.

https://github.com/Netflix/RxJava/wiki/Transforming-Observables#buffer
https://github.com/Netflix/RxJava/wiki/Transforming-Observables#window

On Wed, Apr 30, 2014 at 9:37 AM, Ben Christensen
[email protected]:

suggest a partial merge leaving my BufferUntilSubscriber out and fixing
its ClassCastException and stealing/overtaking behavior in a separate PR.

Agreed. I can fix the ClassCastException and just make it onError if
someone tries to subscribe twice.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/1138#issuecomment-41818932
.

David M. Gross
PLP Consulting

[akarnokd]

Every buffer and window operator will emit empty lists and observables, except the buffer with size == skip variant. I kept this oddity as there was a test case for it.

[cloudbees-pull-request-builder]

RxJava-pull-requests #1064 SUCCESS
This pull request looks good