Actually yes, tagging PCollections and using stateful DoFn can solve
that in all cases - the memory requirements for batch should be solvable
using @RequiresTimeSortedInput, as that is what causes the issues on
batch on runners that support it.
Another yes, this is somewhat related to a full stream-to-stream join
with exact time characteristics, where each side should be matched to
the closest preceding update on the other side (again, some
oversimplification).
"Event time side inputs" could be a handy tool for solving these
problems in cases where one side is "reasonably small", so that it can
simply fir into memory, yet still be arbitrarily fast updatable (as
opposed to the "slowly changing caches" patterns that we sometimes see
using side inputs - the reasons for this "slowness" seem to be in the
inherent processing time matching in global windows). It should also
give a more efficient runtime characteristics.
On 4/27/21 10:27 PM, Robert Bradshaw wrote:
So, I was just thinking when Jan mentioned "timestamp indexed" that
side input and state happen to be implemented on top of roughly the
same thing (in streaming Dataflow, and over the FnAPI) but the
connection might be deeper than that. Side inputs could be viewed as a
special type of (typically bag or map) state that is fully populated
by another PCollection and has a set eviction property. Maybe there's
some deeper unification here that could generalize side inputs.
On the other hand, it seems you can get any desired side input
behavior by taking as your main input a tagged union of the various
PCollections and using state+timers. (This may, however, require
infeasibly large memory requirements for batch.) The value is in a
more general, modular pattern that both pipeline authors and systems
can reason about.
On Tue, Apr 27, 2021 at 1:01 PM Reuven Lax <[email protected]> wrote:
The watermark discussion may be a red herring here.
I think it's interesting to think about the case where the side input is
produced by a state/timers DoFn. in this scenario, there may be more
interesting ways to match the side input than simply using windowing. It might
be worth thinking about this topic more deeply.
On Tue, Apr 27, 2021 at 12:51 PM Jan Lukavský <[email protected]> wrote:
On 4/27/21 9:26 PM, Robert Bradshaw wrote:
On Tue, Apr 27, 2021 at 12:05 PM Jan Lukavský <[email protected]> wrote:
On 4/27/21 8:51 PM, Robert Bradshaw wrote:
On Tue, Apr 27, 2021 at 11:25 AM Jan Lukavský <[email protected]> wrote:
Are you asking for a way to ignore early triggers on side input mapping, and
only map to on-time triggered values for the window?
No, that could for sure be done before applying the View transform. I'd like a
know if it would be possible to create mode of the matching which would be
deterministic. One possibility to make it deterministic seems to be, that main
input elements would be pushed back until side input watermark 'catches up'
with main input. Whenever the side input watermark would be delayed after the
main input watermark, elements would start to be pushed back again. Not sure if
I'm explaining it using the right words. The side input watermark can be
controlled using timer in an upstream transform, so this defines which elements
in main input would be matched onto which pane of the side input.
Perhaps I'm not following the request correctly, but this is exactly
how side inputs work by default. It is only when one explicitly
requests a non-deterministic trigger upstream of the side input (e.g.
one that may fire multiple times or ahead of the watermark) that one
sees a side input with multiple variations or data in the side input
before the watermark of the side input is caught up to the main input.
Yes, exactly. But take the example of side input in global windows (on
both the main input and side input). Then there has to be multiple
firings per window, because otherwise the side input would be available
at the end of time, which is not practical. The trigger doesn't have to
be non-deterministic, the data might come from a stateful ParDo, using a
timer with output timestamp, which would make the downstream watermark
progress quite well defined. The matching would still be
nondeterministic in this case.
If everything is in the global window, things get non-deterministic
across PCollections. For example, say the main input has element m20
and the side input has elements s10 and s30 (with the obvious
timestamps). Suppose we have a total ordering of events as follows.
s10 arrives
side input watermark advances to 25
s30 arrives
side input watermark advances to 100
m20 arrives
In this case, while processing m20, one would see both s10 and s30.
Alternatively we could have had
s10 arrives
side input watermark advances to 25
m20 arrives
s30 arrives
side input watermark advances to 100
in which case m20 would only see s10.
Yes, this is absolutely true, I didn't want to make things too
complicated, so I ignored this fact, that absolutely correct solution
would require the PCollectioView to be timestamp indexed so that if it
would receive both s10 and s30 it would correctly return s10 (as s30
didn't exist at m20, right?). I simplified this at this moment for this
discussion, thanks for clarifying.
Windowing is exactly the mechanism that gives us a cross-PCollection
barrier with which to line things up.
Should not watermark do this?
On 4/27/21 8:03 PM, Reuven Lax wrote:
On Tue, Apr 27, 2021 at 10:53 AM Jan Lukavský <[email protected]> wrote:
If using early triggers, then the side input loads the latest trigger for that
window.
Does not the the word 'latest' imply processing time matching? The watermark of main
input might be arbitrarily delayed from the watermark of side input. If we consider
GlobalWindows on both sides, than "latest" trigger in side input looks exactly
as processing time matching. Yes, between different windows, the matching is in event
time. But within the same window (ignoring the window mapping for now), the matching
looks like processing time, right?
Not really. Think of each trigger of a window as a refinement - so the latest trigger for
that window is our best approximation of the "correct" value for that window.
For this reason, window panes are indexed by an integer (pane_index), not by a timestamp.
The idea here is that the main input element maps to the side input for the best,
most-recent knowledge of the window.
Are you asking for a way to ignore early triggers on side input mapping, and
only map to on-time triggered values for the window?
If we look at the SimplePushbackSideInputDoFnRunner used by runners exactly for
the side input matching, there is no testing of side input watermark to
determine if an element should be 'pushed back' or processed. Element is
processed if, and only if, all side inputs for particular window are ready.
On 4/27/21 7:24 PM, Reuven Lax wrote:
The windows shouldn't need to match - at least if the FlinkRunner
implementation is correct. By default, the side-input's WindowFn should be used
to map the main input's timestamp into a window, and that window is used to
determine which version of the side input to load. A custom WindowFn can be
used to to even more - e.g. if you want the main input element to map to the
_previous_ window in the side input (you would do this by overloading
getDefaultWindowMappingFn).
If using early triggers, then the side input loads the latest trigger for that
window. This is still an event-time mapping - for example two main-input events
in two different windows will still map the the side input for the matching
window. However if the side input PCollection is triggered, than the latest
trigger for each window's side input will be loaded.
It's possible that the FlinkRunner implementation is incomplete, in which case
it should be fixed.
On Tue, Apr 27, 2021 at 9:36 AM Jan Lukavský <[email protected]> wrote:
It seems to me, that this is true only with matching windows on both sides and
default trigger of the side input. Then it will (due to condition a)) work as
if the matching happenned in event time. But when using any early triggers then
it will work in processing time. At least, that is my understanding from
studying the code in FlinkRunner.
On 4/27/21 4:05 PM, Robert Burke wrote:
I thought the matching happened with elements in the matching window, in Event
time, not in Processing time.
Granted, I'm not that familiar with this area myself, but one key part of Beam
is nearly everything is Event time by default, not Processing time.
On Tue, Apr 27, 2021, 12:43 AM Jan Lukavský <[email protected]> wrote:
Hi,
I have a question about matching side inputs to main input. First a
recap, to make sure I understand the current state correctly:
a) elements from main input are pushed back (stored in state) until a
first side input pane arrives (that might be on time, or early)
b) after that, elements from the main input are matched to the current
side input view - the view is updated as new data arrives, but is
matched to the main input elements in processing time
If this is the current state, my question is, would it be possible to
add a new mode of matching of side inputs? Something like
ParDo.of(new MyDoFn()).withSideInput("name", myView,
TimeDomain.EVENT_TIME)
the semantics would be that the elements from the main PCollection would
be stored into state as pairs with the value of the current main input
watermark and on update of side-input watermark only main input elements
with associated input watermark less than that of the side input would
be matched with the side input and sent downstream. Although this
approach is necessarily more expensive and introducing additional
latency than processing time matching, there are situations when
processing time matching is inapropriate for correctness reasons.
WDYT?
Jan
