As Thomas says, the fact that we ever produce values in "no window" is an implementation quirk that should probably be fixed. (IIRC, it's used for the output of a GBK before we've done the group-also-by-windows to figure out what window it really should be in, so "value in unknown windows" would be a better choice).
If a WindowFn doesn't assign a value to any windows, the system is free to drop it. There are pros and cons to supporting this degenerate case vs. making it an error. However, this should almost certainly not be in the public API... - Robert On Wed, Apr 13, 2016 at 9:06 AM, Thomas Groh <[email protected]> wrote: > Actually, my above claim isn't as strong as it can be. > > A value in no windows is considered to not exist. Values that are not > assigned to any window can be dropped by a runner at *any time*. A WindowFn > *must* assign all elements to at least one window. All elements that are > produced by any PTransform (including Sources) must be in a window, > potentially the GlobalWindow. > > On Wed, Apr 13, 2016 at 8:52 AM, Thomas Groh <[email protected]> wrote: > >> Values should almost always be part of at least one window. WindowFns >> should place all elements in at least one window, as values that are in no >> windows will be dropped when they reach a GroupByKey. >> >> Elements in no windows, for example those created by >> WindowedValue.valueInEmptyWindows(T) are generally an implementation >> detail of a transform; for example, in the InProcessPipelineRunner, the KV<K, >> Iterable<WindowedValue<V>>> elements output by a GroupByKeyOnly are in >> empty windows - but by the time the element reaches the boundary of the >> GroupByKey, the elements are reassigned to the appropriate window(s). >> >> On Tue, Apr 12, 2016 at 11:44 PM, Amit Sela <[email protected]> wrote: >> >>> My instinct tells me that if a value does not belong to a specific window >>> (in time) it's a part of a global window, but if so, what's the role of >>> the >>> "empty window". When should an element be a "value in an empty window" ? >>> >> >>
