flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From M Singh <mans2si...@yahoo.com>
Subject Re: Flink : CEP processing
Date Wed, 10 Aug 2016 18:42:41 GMT
Thanks for the pointers Sameer.

The reason I wanted to find out about snapshotting with CEP is because I thought that CEP
state might also be snapshotted for recovery. If that is the case, then there are events in
the CEP might be in two snapshots.
Mans 

    On Tuesday, August 9, 2016 1:15 PM, Sameer W <sameer@axiomine.com> wrote:
 

 In one of the earlier thread Till explained this to me (http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/CEP-and-Within-Clause-td8159.html)
1. Within does not use time windows. It sort of uses session windows where the session begins
when the first event of the pattern is identified. The timer starts when the "first" event
in the pattern fires. If the pattern completes "within" the designated times (meaning the
"next" and "followed by" fire as will "within" the time specified) you have a match or else
the window is removed. I don't know how it is implemented but I doubt it stores all the events
in memory for the "within" window (there is not need to). It will only store the relevant
events (first, next, followed by, etc). So memory would not be an issue here. If two "first"
type events are identified I think two "within" sessions are created.
2. Snapshotting (I don't know much in this area so I cannot answer). Why should it be different
though? You are using operators and state. It should work the same way. But I am not too familiar
with that.
3. The "Within" window is not an issue. Even the window preceding that should not be unless
you are using WindowFunction (more memory friendly alternative is https://ci.apache.org/
projects/flink/flink-docs- master/apis/streaming/windows. html#window-functions ) by themselves
and using a really large window
4. The way I am using it, it is working fine. Some of the limitations I have seen are related
to this paper not being fully implemented (https://people.cs.umass.edu/ ~yanlei/publications/sase-
sigmod08.pdf). I don't know how to support negation in an event stream but I don't need it
for now.
Thanks,Sameer

On Tue, Aug 9, 2016 at 3:45 PM, M Singh <mans2singh@yahoo.com> wrote:

Hi Sameer:
If we use a within window for event series - 
1. Does it interfere with the default time windows ?2. How does it affect snapshotting ?  3.
If the window is too large are the events stored in a "processor" for the window to expire
?4. Are there any other know limitations and best practices of using CEP with Flink ?
Thanks again for your help.
 

    On Tuesday, August 9, 2016 11:29 AM, Sameer Wadkar <sameer@axiomine.com> wrote:
 

 In that case you need to get them into one stream somehow (keyBy a dummy value for example).
There is always some logical key to keyBy on when data is arriving from multiple sources (ex
some portion of the time stamp). 
You are looking for patterns within something (events happening around the same time but arriving
from multiple devices). That something should be the key. That's how I am using it. 

Sameer
Sent from my iPhone
On Aug 9, 2016, at 1:40 PM, M Singh <mans2singh@yahoo.com> wrote:


Thanks Sameer.
So does that mean that if the events keys are not same we cannot use the CEP pattern match
?  What if events are coming from different sources and need to be correlated ?
Mans 

    On Tuesday, August 9, 2016 9:40 AM, Sameer W <sameer@axiomine.com> wrote:
 

 Hi,
You will need to use keyBy operation first to get all the events you need monitored in a pattern
on the same node. Only then can you apply Pattern because it depends on the order of the events
(first, next, followed by). I even had to make sure that the events were correctly sorted
by timestamps to ensure that the first,next and followed by works correctly.
Sameer
On Tue, Aug 9, 2016 at 12:17 PM, M Singh <mans2singh@yahoo.com> wrote:

Hey Folks:
I have a question about CEP processing in Flink - How does flink processing work when we have
multiple partitions in which the events used in the pattern sequence might be scattered across
multiple partitions on multiple nodes ?
Thanks for your insight.
Mans



   


   



  
Mime
View raw message