On Mon, May 20, 2019 at 5:24 PM Jan Lukavský <[email protected]> wrote:
> I don't see batch vs. streaming as part of the model. One can have
microbatch, or even a runner that alternates between different modes.
Although I understand motivation of this statement, this project name is
"Apache Beam: An advanced unified programming model". What does the
model unify, if "streaming vs. batch" is not part of the model?
What I mean is that streaming vs. batch is no longer part of the model
(or ideally API), but pushed down to be a concern of the runner
(executor) of the pipeline.
On Tue, May 21, 2019 at 10:32 AM Jan Lukavský <[email protected]> wrote:
Hi Kenn,
OK, so if we introduce annotation, we can have stateful ParDo with sorting,
that would perfectly resolve my issues. I still have some doubts, though. Let
me explain. The current behavior of stateful ParDo has the following properties:
a) might fail in batch, although runs fine in streaming (that is due to the
buffering, and unbounded lateness in batch, which was discussed back and forth
in this thread)
b) might be non deterministic (this is because the elements arrive at somewhat random
order, and even if you do the operation "assign unique ID to elements" this
might produce different results when run multiple times)
PCollections are *explicitly* unordered. Any operations that assume or
depend on a specific ordering for correctness (or determinism) must
provide that ordering themselves (i.e. tolerate "arbitrary shuffling
of inputs"). As you point out, that may be very expensive if you have
very hot keys with very large (unbounded) timestamp skew.
StatefulDoFns are low-level operations that should be used with care;
the simpler windowing model gives determinism in the face of unordered
data (though late data and non-end-of-window triggering introduces
some of the non-determanism back in).
What worries me most is the property b), because it seems to me to have serious
consequences - not only that if you run twice batch pipeline you would get
different results, but even on streaming, when pipeline fails and gets
restarted from checkpoint, produced output might differ from the previous run
and data from the first run might have already been persisted into sink. That
would create somewhat messy outputs.
Beam has an exactly-once model. If the data was consumed, state
mutated, and outputs written downstream (these three are committed
together atomically) it will not be replayed. That does not, of
course, solve the non-determanism due to ordering (including the fact
that two operations reading the same PCollection may view different
ordering).
These two properties makes me think that the current implementation is more of
a _special case_ than the general one. The general one would be that your state
doesn't have the properties to be able to tolerate buffering problems and/or
non-determinism. Which is the case where you need sorting in both streaming and
batch to be part of the model.
Let me point out one more analogy - that is merging vs. non-merging windows.
The general case (merging windows) implies sorting by timestamp in both batch
case (explicit) and streaming (buffering). The special case (non-merging
windows) doesn't rely on any timestamp ordering, so the sorting and buffering
can be dropped. The underlying root cause of this is the same for both stateful
ParDo and windowing (essentially, assigning window labels is a stateful
operation when windowing function is merging).
The reason for the current behavior of stateful ParDo seems to be performance,
but is it right to abandon correctness in favor of performance? Wouldn't it be
more consistent to have the default behavior prefer correctness and when you
have the specific conditions of state function having special properties, then
you can annotate your DoFn (with something like @TimeOrderingAgnostic), which
would yield a better performance in that case?
There are two separable questions here.
1) Is it correct for a (Stateful)DoFn to assume elements are received
in a specific order? In the current model, it is not. Being able to
read, handle, and produced out-of-order data, including late data, is
a pretty fundamental property of distributed systems.
2) Given that some operations are easier (or possibly only possible)
to write when operating on ordered data, and that different runners
may have (significantly) cheaper ways to provide this ordering than
can be done by the user themselves, should we elevate this to a
property of (Stateful?)DoFns that the runner can provide? I think a
compelling argument can be made here that we should.
- Robert
On 5/21/19 1:00 AM, Kenneth Knowles wrote:
Thanks for the nice small example of a calculation that depends on order. You are right
that many state machines have this property. I agree w/ you and Luke that it is
convenient for batch processing to sort by event timestamp before running a stateful
ParDo. In streaming you could also implement "sort by event timestamp" by
buffering until you know all earlier data will be dropped - a slack buffer up to allowed
lateness.
I do not think that it is OK to sort in batch and not in streaming. Many state machines
diverge very rapidly when things are out of order. So each runner if they see the
"@OrderByTimestamp" annotation (or whatever) needs to deliver sorted data (by
some mix of buffering and dropping), or to reject the pipeline as unsupported.
And also want to say that this is not the default case - many uses of state & timers in ParDo yield
different results at the element level, but the results are equivalent at in the big picture. Such as
the example of "assign a unique sequence number to each element" or "group into
batches" it doesn't matter exactly what the result is, only that it meets the spec. And other
cases like user funnels are monotonic enough that you also don't actually need sorting.
Kenn
On Mon, May 20, 2019 at 2:59 PM Jan Lukavský <[email protected]> wrote:
Yes, the problem will arise probably mostly when you have not well distributed
keys (or too few keys). I'm really not sure if a pure GBK with a trigger can
solve this - it might help to have data driven trigger. There would still be
some doubts, though. The main question is still here - people say, that sorting
by timestamp before stateful ParDo would be prohibitively slow, but I don't
really see why - the sorting is very probably already there. And if not (hash
grouping instead of sorted grouping), then the sorting would affect only user
defined StatefulParDos.
This would suggest that the best way out of this would be really to add
annotation, so that the author of the pipeline can decide.
If that would be acceptable I think I can try to prepare some basic
functionality, but I'm not sure, if I would be able to cover all runners / sdks.
On 5/20/19 11:36 PM, Lukasz Cwik wrote:
It is read all per key and window and not just read all (this still won't scale
with hot keys in the global window). The GBK preceding the StatefulParDo will
guarantee that you are processing all the values for a specific key and window
at any given time. Is there a specific window/trigger that is missing that you
feel would remove the need for you to use StatefulParDo?
On Mon, May 20, 2019 at 12:54 PM Jan Lukavský <[email protected]> wrote:
Hi Lukasz,
Today, if you must have a strict order, you must guarantee that your StatefulParDo implements
the necessary "buffering & sorting" into state.
Yes, no problem with that. But this whole discussion started, because *this
doesn't work on batch*. You simply cannot first read everything from
distributed storage and then buffer it all into memory, just to read it again,
but sorted. That will not work. And even if it would, it would be a terrible
waste of resources.
Jan
On 5/20/19 8:39 PM, Lukasz Cwik wrote:
On Mon, May 20, 2019 at 8:24 AM Jan Lukavský <[email protected]> wrote:
This discussion brings many really interesting questions for me. :-)
> I don't see batch vs. streaming as part of the model. One can have
microbatch, or even a runner that alternates between different modes.
Although I understand motivation of this statement, this project name is
"Apache Beam: An advanced unified programming model". What does the
model unify, if "streaming vs. batch" is not part of the model?
Using microbatching, chaining of batch jobs, or pure streaming are
exactly the "runtime conditions/characteristics" I refer to. All these
define several runtime parameters, which in turn define how well/badly
will the pipeline perform and how many resources might be needed. From
my point of view, pure streaming should be the most resource demanding
(if not, why bother with batch? why not run everything in streaming
only? what will there remain to "unify"?).
> Fortunately, for batch, only the state for a single key needs to be
preserved at a time, rather than the state for all keys across the range
of skew. Of course if you have few or hot keys, one can still have
issues (and this is not specific to StatefulDoFns).
Yes, but here is still the presumption that my stateful DoFn can
tolerate arbitrary shuffling of inputs. Let me explain the use case in
more detail.
Suppose you have input stream consisting of 1s and 0s (and some key for
each element, which is irrelevant for the demonstration). Your task is
to calculate in running global window the actual number of changes
between state 0 and state 1 and vice versa. When the state doesn't
change, you don't calculate anything. If input (for given key) would be
(tN denotes timestamp N):
t1: 1
t2: 0
t3: 0
t4: 1
t5: 1
t6: 0
then the output should yield (supposing that default state is zero):
t1: (one: 1, zero: 0)
t2: (one: 1, zero: 1)
t3: (one: 1, zero: 1)
t4: (one: 2, zero: 1)
t5: (one: 2, zero: 1)
t6: (one: 2, zero: 2)
How would you implement this in current Beam semantics?
I think your saying here that I know that my input is ordered in a specific way
and since I assume the order when writing my pipeline I can perform this
optimization. But there is nothing preventing a runner from noticing that your
processing in the global window with a specific type of trigger and re-ordering
your inputs/processing to get better performance (since you can't use an
AfterWatermark trigger for your pipeline in streaming for the GlobalWindow).
Today, if you must have a strict order, you must guarantee that your StatefulParDo implements the
necessary "buffering & sorting" into state. I can see why you would want an annotation
that says I must have timestamp ordered elements, since it makes writing certain StatefulParDos much
easier. StatefulParDo is a low-level function, it really is the "here you go and do whatever you
need to but here be dragons" function while windowing and triggering is meant to keep many people
from writing StatefulParDo in the first place.
> Pipelines that fail in the "worst case" batch scenario are likely to
degrade poorly (possibly catastrophically) when the watermark falls
behind in streaming mode as well.
But the worst case is defined by input of size (available resources +
single byte) -> pipeline fail. Although it could have finished, given
the right conditions.
> This might be reasonable, implemented by default by buffering
everything and releasing elements as the watermark (+lateness) advances,
but would likely lead to inefficient (though *maybe* easier to reason
about) code.
Sure, the pipeline will be less efficient, because it would have to
buffer and sort the inputs. But at least it will produce correct results
in cases where updates to state are order-sensitive.
> Would it be roughly equivalent to GBK + FlatMap(lambda (key, values):
[(key, value) for value in values])?
I'd say roughly yes, but difference would be in the trigger. The trigger
should ideally fire as soon as watermark (+lateness) crosses element
with lowest timestamp in the buffer. Although this could be somehow
emulated by fixed trigger each X millis.
> Or is the underlying desire just to be able to hint to the runner
that the code may perform better (e.g. require less resources) as skew
is reduced (and hence to order by timestamp iff it's cheap)?
No, the sorting would have to be done in streaming case as well. That is
an imperative of the unified model. I think it is possible to sort by
timestamp only in batch case (and do it for *all* batch stateful pardos
without annotation), or introduce annotation, but then make the same
guarantees for streaming case as well.
Jan
On 5/20/19 4:41 PM, Robert Bradshaw wrote:
On Mon, May 20, 2019 at 1:19 PM Jan Lukavský <[email protected]> wrote:
Hi Robert,
yes, I think you rephrased my point - although no *explicit* guarantees
of ordering are given in either mode, there is *implicit* ordering in
streaming case that is due to nature of the processing - the difference
between watermark and timestamp of elements flowing through the pipeline
are generally low (too high difference leads to the overbuffering
problem), but there is no such bound on batch.
Fortunately, for batch, only the state for a single key needs to be
preserved at a time, rather than the state for all keys across the
range of skew. Of course if you have few or hot keys, one can still
have issues (and this is not specific to StatefulDoFns).
As a result, I see a few possible solutions:
- the best and most natural seems to be extension of the model, so
that it defines batch as not only "streaming pipeline executed in batch
fashion", but "pipeline with at least as good runtime characteristics as
in streaming case, executed in batch fashion", I really don't think that
there are any conflicts with the current model, or that this could
affect performance, because the required sorting (as pointed by
Aljoscha) is very probably already done during translation of stateful
pardos. Also note that this definition only affects user defined
stateful pardos
I don't see batch vs. streaming as part of the model. One can have
microbatch, or even a runner that alternates between different modes.
The model describes what the valid outputs are given a (sometimes
partial) set of inputs. It becomes really hard to define things like
"as good runtime characteristics." Once you allow any
out-of-orderedness, it is not very feasible to try and define (and
more cheaply implement) a "upper bound" of acceptable
out-of-orderedness.
Pipelines that fail in the "worst case" batch scenario are likely to
degrade poorly (possibly catastrophically) when the watermark falls
behind in streaming mode as well.
- another option would be to introduce annotation for DoFns (e.g.
@RequiresStableTimeCharacteristics), which would result in the sorting
in batch case - but - this extension would have to ensure the sorting in
streaming mode also - it would require definition of allowed lateness,
and triggger (essentially similar to window)
This might be reasonable, implemented by default by buffering
everything and releasing elements as the watermark (+lateness)
advances, but would likely lead to inefficient (though *maybe* easier
to reason about) code. Not sure about the semantics of triggering
here, especially data-driven triggers. Would it be roughly equivalent
to GBK + FlatMap(lambda (key, values): [(key, value) for value in
values])?
Or is the underlying desire just to be able to hint to the runner that
the code may perform better (e.g. require less resources) as skew is
reduced (and hence to order by timestamp iff it's cheap)?
- last option would be to introduce these "higher order guarantees" in
some extension DSL (e.g. Euphoria), but that seems to be the worst
option to me
I see the first two options quite equally good, although the letter one
is probably more time consuming to implement. But it would bring
additional feature to streaming case as well.
Thanks for any thoughts.
Jan
On 5/20/19 12:41 PM, Robert Bradshaw wrote:
On Fri, May 17, 2019 at 4:48 PM Jan Lukavský <[email protected]> wrote:
Hi Reuven,
How so? AFAIK stateful DoFns work just fine in batch runners.
Stateful ParDo works in batch as far, as the logic inside the state works for absolutely unbounded
out-of-orderness of elements. That basically (practically) can work only for cases, where the order
of input elements doesn't matter. But, "state" can refer to "state machine",
and any time you have a state machine involved, then the ordering of elements would matter.
No guarantees on order are provided in *either* streaming or batch
mode by the model. However, it is the case that in order to make
forward progress most streaming runners attempt to limit the amount of
out-of-orderedness of elements (in terms of event time vs. processing
time) to make forward progress, which in turn could help cap the
amount of state that must be held concurrently, whereas a batch runner
may not allow any state to be safely discarded until the whole
timeline from infinite past to infinite future has been observed.
Also, as pointed out, state is not preserved "batch to batch" in batch mode.
On Thu, May 16, 2019 at 3:59 PM Maximilian Michels <[email protected]> wrote:
batch semantics and streaming semantics differs only in that I can have
GlobalWindow with default trigger on batch and cannot on stream
You can have a GlobalWindow in streaming with a default trigger. You
could define additional triggers that do early firings. And you could
even trigger the global window by advancing the watermark to +inf.
IIRC, as a pragmatic note, we prohibited global window with default
trigger on unbounded PCollections in the SDK because this is more
likely to be user error than an actual desire to have no output until
drain. But it's semantically valid in the model.
