flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <ted.dunn...@gmail.com>
Subject Re: Thoughts About Streaming
Date Tue, 23 Jun 2015 17:43:52 GMT
Out of order is ubiquitous in the real-world.  Typically, what happens is
that businesses will declare a maximum allowable delay for delayed
transactions and will commit to results when that delay is reached.
Transactions that arrive later than this cutoff are collected specially as
corrections which are reported/used when possible.

Clearly, ordering can also be violated during processing, but if the data
is originally out of order the situation can't be repaired by any protocol
fixes that prevent transactions from becoming disordered but has to handled
at the data level.




On Tue, Jun 23, 2015 at 9:29 AM, Aljoscha Krettek <aljoscha@apache.org>
wrote:

> I also don't like big changes but sometimes they are necessary. The reason
> why I'm so adamant about out-of-order processing is that out-of-order
> elements are not some exception that occurs once in a while; they occur
> constantly in a distributed system. For example, in this:
> https://gist.github.com/aljoscha/a367012646ab98208d27 the resulting
> windows
> are completely bogus because the current windowing system assumes elements
> to globally arrive in order, which is simply not true. (The example has a
> source that generates increasing integers. Then these pass through a map
> and are unioned with the original DataStream before a window operator.)
> This simulates elements arriving from different operators at a windowing
> operator. The example is also DOP=1, I imagine this to get worse with
> higher DOP.
>
> What do you mean by costly? As I said, I have a proof-of-concept windowing
> operator that can handle out-or-order elements. This is an example using
> the current Flink API:
> https://gist.github.com/aljoscha/f8dce0691732e344bbe8.
> (It is an infinite source of tuples and a 5 second window operator that
> counts the tuples.) The first problem is that this code deadlocks because
> of the thread that emits fake elements. If I disable the fake element code
> it works, but the throughput using my mockup is 4 times higher . The gap
> widens dramatically if the window size increases.
>
> So, it actually increases performance (unless I'm making a mistake in my
> explorations) and can handle elements that arrive out-of-order (which
> happens basically always in a real-world windowing use-cases).
>
>
> On Tue, 23 Jun 2015 at 12:51 Stephan Ewen <sewen@apache.org> wrote:
>
> > What I like a lot about Aljoscha's proposed design is that we need no
> > different code for "system time" vs. "event time". It only differs in
> where
> > the timestamps are assigned.
> >
> > The OOP approach also gives you the semantics of total ordering without
> > imposing merges on the streams.
> >
> > On Tue, Jun 23, 2015 at 10:43 AM, Matthias J. Sax <
> > mjsax@informatik.hu-berlin.de> wrote:
> >
> > > I agree that there should be multiple alternatives the user(!) can
> > > choose from. Partial out-of-order processing works for many/most
> > > aggregates. However, if you consider Event-Pattern-Matching, global
> > > ordering in necessary (even if the performance penalty might be high).
> > >
> > > I would also keep "system-time windows" as an alternative to "source
> > > assigned ts-windows".
> > >
> > > It might also be interesting to consider the following paper for
> > > overlapping windows: "Resource sharing in continuous sliding-window
> > > aggregates"
> > >
> > > > https://dl.acm.org/citation.cfm?id=1316720
> > >
> > > -Matthias
> > >
> > > On 06/23/2015 10:37 AM, Gyula Fóra wrote:
> > > > Hey
> > > >
> > > > I think we should not block PRs unnecessarily if your suggested
> changes
> > > > might touch them at some point.
> > > >
> > > > Also I still think we should not put everything in the Datastream
> > because
> > > > it will be a huge mess.
> > > >
> > > > Also we need to agree on the out of order processing, whether we want
> > it
> > > > the way you proposed it(which is quite costly). Another alternative
> > > > approach there which fits in the current windowing is to filter out
> if
> > > > order events and apply a special handling operator on them. This
> would
> > be
> > > > fairly lightweight.
> > > >
> > > > My point is that we need to consider some alternative solutions. And
> we
> > > > should not block contributions along the way.
> > > >
> > > > Cheers
> > > > Gyula
> > > >
> > > > On Tue, Jun 23, 2015 at 9:55 AM Aljoscha Krettek <
> aljoscha@apache.org>
> > > > wrote:
> > > >
> > > >> The reason I posted this now is that we need to think about the API
> > and
> > > >> windowing before proceeding with the PRs of Gabor (inverse reduce)
> and
> > > >> Gyula (removal of "aggregate" functions on DataStream).
> > > >>
> > > >> For the windowing, I think that the current model does not work for
> > > >> out-of-order processing. Therefore, the whole windowing
> infrastructure
> > > will
> > > >> basically have to be redone. Meaning also that any work on the
> > > >> pre-aggregators or optimizations that we do now becomes useless.
> > > >>
> > > >> For the API, I proposed to restructure the interactions between all
> > the
> > > >> different *DataStream classes and grouping/windowing. (See API
> section
> > > of
> > > >> the doc I posted.)
> > > >>
> > > >> On Mon, 22 Jun 2015 at 21:56 Gyula Fóra <gyula.fora@gmail.com>
> wrote:
> > > >>
> > > >>> Hi Aljoscha,
> > > >>>
> > > >>> Thanks for the nice summary, this is a very good initiative.
> > > >>>
> > > >>> I added some comments to the respective sections (where I didnt
> fully
> > > >> agree
> > > >>> :).).
> > > >>> At some point I think it would be good to have a public hangout
> > session
> > > >> on
> > > >>> this, which could make a more dynamic discussion.
> > > >>>
> > > >>> Cheers,
> > > >>> Gyula
> > > >>>
> > > >>> Aljoscha Krettek <aljoscha@apache.org> ezt írta (időpont:
2015.
> jún.
> > > >> 22.,
> > > >>> H, 21:34):
> > > >>>
> > > >>>> Hi,
> > > >>>> with people proposing changes to the streaming part I also
wanted
> to
> > > >>> throw
> > > >>>> my hat into the ring. :D
> > > >>>>
> > > >>>> During the last few months, while I was getting acquainted
with
> the
> > > >>>> streaming system, I wrote down some thoughts I had about how
> things
> > > >> could
> > > >>>> be improved. Hopefully, they are in somewhat coherent shape
now,
> so
> > > >>> please
> > > >>>> have a look if you are interested in this:
> > > >>>>
> > > >>>>
> > > >>>
> > > >>
> > >
> >
> https://docs.google.com/document/d/1rSoHyhUhm2IE30o5tkR8GEetjFvMRMNxvsCfoPsW6_4/edit?usp=sharing
> > > >>>>
> > > >>>> This mostly covers:
> > > >>>>  - Timestamps assigned at sources
> > > >>>>  - Out-of-order processing of elements in window operators
> > > >>>>  - API design
> > > >>>>
> > > >>>> Please let me know what you think. Comment in the document
or here
> > in
> > > >> the
> > > >>>> mailing list.
> > > >>>>
> > > >>>> I have a PR in the makings that would introduce source timestamps
> > and
> > > >>>> watermarks for keeping track of them. I also hacked a
> > proof-of-concept
> > > >>> of a
> > > >>>> windowing system that is able to process out-of-order elements
> > using a
> > > >>>> FlatMap operator. (It uses panes to perform efficient
> > > >> pre-aggregations.)
> > > >>>>
> > > >>>> Cheers,
> > > >>>> Aljoscha
> > > >>>>
> > > >>>
> > > >>
> > > >
> > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message