As far as I see in [1], Peter's/Gyula's suggestion is what Infosphere
Streams does: symmetric hash join.
From [1]:
When a tuple is received on an input port, it is inserted into the window
corresponding to the input port, which causes the window to trigger. As
part of the trigger processing, the tuple is compared against all tuples
inside the window of the opposing input port. If the tuples match, then an
output tuple will be produced for each match. If at least one output was
generated, a window punctuation will be generated after all the outputs.
Cheers,
Asterios
[1]
http://www-01.ibm.com/support/knowledgecenter/#!/SSCRJU_3.2.1/com.ibm.swg.im.infosphere.streams.spl-standard-toolkit-reference.doc/doc/join.html
On Thu, Apr 9, 2015 at 1:30 PM, Matthias J. Sax
mj...@informatik.hu-berlin.de wrote:
Hi Paris,
thanks for the pointer to the Naiad paper. That is quite interesting.
The paper I mentioned [1], does not describe the semantics in detail; it
is more about the implementation for the stream-joins. However, it uses
the same semantics (from my understanding) as proposed by Gyula.
-Matthias
[1] Kang, Naughton, Viglas. Evaluationg Window Joins over Unbounded
Streams. VLDB 2002.
On 04/07/2015 12:38 PM, Paris Carbone wrote:
Hello Matthias,
Sure, ordering guarantees are indeed a tricky thing, I recall having
that discussion back in TU Berlin. Bear in mind thought that DataStream,
our abstract data type, represents a *partitioned* unbounded sequence of
events. There are no *global* ordering guarantees made whatsoever in that
model across partitions. If you see it more generally there are many “race
conditions” in a distributed execution graph of vertices that process
multiple inputs asynchronously, especially when you add joins and
iterations into the mix (how do you deal with reprocessing “old” tuples
that iterate in the graph). Btw have you checked the Naiad paper [1]?
Stephan cited a while ago and it is quite relevant to that discussion.
Also, can you cite the paper with the joining semantics you are
referring to? That would be of good help I think.
Paris
[1] https://users.soe.ucsc.edu/~abadi/Papers/naiad_final.pdf
https://users.soe.ucsc.edu/~abadi/Papers/naiad_final.pdf
https://users.soe.ucsc.edu/~abadi/Papers/naiad_final.pdf
On 07 Apr 2015, at 11:50, Matthias J. Sax mj...@informatik.hu-berlin.de
mailto:mj...@informatik.hu-berlin.de wrote:
Hi @all,
please keep me in the loop for this work. I am highly interested and I
want to help on it.
My initial thoughts are as follows:
1) Currently, system timestamps are used and the suggested approach can
be seen as state-of-the-art (there is actually a research paper using
the exact same join semantic). Of course, the current approach is
inherently non-deterministic. The advantage is, that there is no
overhead in keeping track of the order of records and the latency should
be very low. (Additionally, state-recovery is simplified. Because, the
processing in inherently non-deterministic, recovery can be done with
relaxed guarantees).
2) The user should be able to switch on deterministic processing,
ie, records are timestamped (either externally when generated, or
timestamped at the sources). Because deterministic processing adds some
overhead, the user should decide for it actively.
In this case, the order must be preserved in each re-distribution step
(merging is sufficient, if order is preserved within each incoming
channel). Furthermore, deterministic processing can be achieved by sound
window semantics (and there is a bunch of them). Even for
single-stream-windows it's a tricky problem; for join-windows it's even
harder. From my point of view, it is less important which semantics are
chosen; however, the user must be aware how it works. The most tricky
part for deterministic processing, is to deal with duplicate timestamps
(which cannot be avoided). The timestamping for (intermediate) result
tuples, is also an important question to be answered.
-Matthias
On 04/07/2015 11:37 AM, Gyula Fóra wrote:
Hey,
I agree with Kostas, if we define the exact semantics how this works,
this
is not more ad-hoc than any other stateful operator with multiple inputs.
(And I don't think any other system support something similar)
We need to make some design choices that are similar to the issues we had
for windowing. We need to chose how we want to evaluate the windowing
policies (global or local) because that affects what kind of policies can
be parallel, but I can work on these things.
I think this is an amazing feature, so I wouldn't necessarily rush the
implementation for 0.9 though.
And thanks for helping writing these down.
Gyula
On Tue, Apr 7, 2015 at 11:11 AM, Kostas Tzoumas ktzou...@apache.org
mailto:ktzou...@apache.org wrote:
Yes, we should write these semantics down. I volunteer to help.
I don't