Re: Geode for time series data

Sun, 28 Feb 2016 16:44:38 -0800 

Hi there Andrew.
As you have discovered, pushing 100s per second is just not feasible for multiple reasons. This is mostly due to the fact that the events are mostly redundant (as in not all of them actually have changed data entries) or the fact that downstream systems cannot deal with the volume of data events being pushed and fall behind processing the data entries really quickly. 


Coalescing data on the input is a great idea. As you can determine the 
amount of data you pushing into the servers to process. Which is 
beneficial to everything downstream from that point. But this is 
something you'd have to write yourself on the ingest client.



I previously implemented something similar to what you are trying to 
achieve, using an AsyncEventListener/AsyncEventQueue, which is a batch 
(size or batch) processor. AsyncEventListeners are a little under 
valued, as everyone always wants single event processing response (like 
CacheListeners).



Although the main use case for AsyncEventListeners is mostly for 
write-behind for DB's, it allows for the processing of events (in batch) 
on the server-side. But turning on batch-conflation for the 
AsyncEventListener you could coalesce data entries per batch. More on 
AsyncEventListeners can be found in the docco. AsyncEventListeners 
<http://gemfire.docs.pivotal.io/docs-gemfire/developing/events/implementing_write_behind_event_handler.html> 
(http://gemfire.docs.pivotal.io/docs-gemfire/developing/events/implementing_write_behind_event_handler.html)



The implementation I worked on, would ingest data on an "input-region", 
which had an AsyncEventListener attached to it. The AsyncEventListener 
would then fire processing when either the time (200ms) or batch size 
(1000) was hit. Each batch would only contain unique coalesced keys 
which would then be inserted onto a "client-facing" region, which could 
then have either CQ's or Subscriptions enabled. Susbcription Docco 
<cid:[email protected]> 
(http://gemfire.docs.pivotal.io/docs-gemfire/latest/developing/events/configure_client_server_event_messaging.html)



In this case we used subscriptions, as all the data events from the 
server were required to be forwarded to the subscribed client. The 
client would then make the decision if it required to process the data. 
CQ's can be used as well, but I prefer not to add the extra burden on 
the server to filter every data entry event to see if it should be sent 
or not. But you could test either one and see which one works better for 
you.



The subscription client would then provide a CacheListener 
implementation on the "client-facing" region, and it can decide how it 
wants to process any of the data events.


Hope this helps.

--Udo

On 28/02/2016 3:45 pm, Andrew Munn wrote:

What if you have one market data feed handler that receives the incoming
price changes rapidly from an exchange and you want to push those updated
values out to listening objects using (I think) Continuous Queries?  I
wrote a Coherence app years ago which does this and I would like to port
it over to Geode or GF.  Any tips on cache configuration, etc?  These
updates could come at rates of 100s per second.  In the past I would
coalesce the updates in the feed handler and push them into the coherence
cache only once per second.

Thanks,
Andrew



On Tue, 23 Feb 2016, Michael Stolz wrote:


Something like that.You might choose a smaller granularity than minute if 
you're really getting that many ticks per minute.
But you probably want a consistent granularity to make it relatively easy to 
find what you are looking for.
You'll probably also want the date in the key.


--Mike Stolz
Principal Engineer, GemFire Product Manager
Mobile: 631-835-4771

On Tue, Feb 23, 2016 at 11:07 AM, Andrew Munn <[email protected]> wrote:
       How does that work when you're appending incoming data in realtime?  Say
       you're getting 1,000,000 data points per day on each of 1,000 incoming
       stock symbols.  That is 1bln data points.  Are you using keys like this
       that bucket the data into one array per minute of the day

               MSFT-08:00
               MSFT-08:01
               ...
               MSFT-08:59
               etc?

       each array might have several thousand elements in that case.

       Thanks
       Andrew

       On Mon, 22 Feb 2016, Michael Stolz wrote:

       > You will definitely want to use arrays rather than storing each 
individual data point because the overhead of
       each entry in Geode is nearly 300 bytes.
       > You could choose to partition by day/week/month but it shouldn't be 
necessary because the default partitioning
       scheme should be random enough to get reasonable distribution
       > if you are using the metadata and starting timestamp of the array as 
the key.
       >
       >
       > --Mike Stolz
       > Principal Engineer, GemFire Product Manager
       > Mobile: 631-835-4771
       >
       > On Fri, Feb 19, 2016 at 1:43 PM, Alan Kash <[email protected]> wrote:
       >       Hi,
       > I am also building a dashboard prototype for time-series data,
       >
       > For time-series data, usually we target a single metric change (stock 
price, temperature, pressure, etc.) for
       an entity, but the associated metadata with event -
       > {StockName/Place, DeviceID, ApplicationID, EventType} remains constant.
       >
       > For a backend like Cassandra, we denormalize everything and put 
everything in a flat key-map with [Metric,
       Timestamp, DeviceID, Type] as the key. This results in data
       > duplication of the associated "Metadata".
       >
       > Do you recommend similar approach for Geode ?
       >
       > Alternatively,
       >
       > We can have an array for Metrics associated with a given Metadata key 
and store it in a Map ?
       >
       > Key = [Metadata, Timestamp]
       >
       > TSMAP<Key, Array<Metric>> series = [1,2,3,4,5,6,7,8,9]
       >
       > We can partition this at application level by day / week / month.
       >
       > Is this approach better ?
       >
       > There is a metrics spec for TS data modeling for those who are 
interested - http://metrics20.org
       >
       > Thanks
       >
       >
       >
       > On Fri, Feb 19, 2016 at 1:11 PM, Michael Stolz <[email protected]> 
wrote:
       >       You will likely get best results in terms of speed of access if 
you put some structure around the way you
       store the data in-memory.
       > First off, you would probably want to parse the data into the 
individual fields and create a Java object that
       represents that structure.
       >
       > Then you would probably want to bundle those Java structures into 
arrays in such a way that it is easy to get
       to the array for a particular date and time by the
       > combination of a ticker and a date and time as the key.
       >
       > Those arrays of Java objects is what you would store as entries in 
Geode.
       > I think this would give you the fastest access to the data.
       >
       > By the way, probably better to use an integer Julian date and a long 
integer for the time rather than a Java
       Date. Java Dates in Geode PDX are way bigger than you
       > want when you have millions of them.
       >
       > Looking at the sample dataset you provided it appears there is a lot 
of redundant data in there. Repeating
       1926.75 for instance.
       > In fact, every field but 2 are all the same. Are the repetitious 
fields necessary? If they are, then you might
       consider using a columnar approach instead of the
       > Java structures I mentioned. Make an array for each column and compact 
the repetitions with a count. It would
       be slower but more compact.
       > The timestamps are all the same too. Strange.
       >
       >
       >
       > --Mike Stolz
       > Principal Engineer, GemFire Product Manager
       > Mobile: 631-835-4771
       >
       > On Fri, Feb 19, 2016 at 12:15 AM, Gregory Chase <[email protected]> 
wrote:
       >       Hi Andrew,I'll let one of the committers answer to your specific 
data file question. However, you might
       find some inspiration in this open source demo
       >       that some of the Geode team presented at OSCON earlier this
       year: http://pivotal-open-source-hub.github.io/StockInference-Spark/
       >
       > This was based on a pre-release version of Geode, so you'll want to 
sub the M1 release in and see if any other
       tweaks are required at that point.
       >
       > I believe this video and presentation go with the Github
       project: http://www.infoq.com/presentations/r-gemfire-spring-xd
       >
       > On Thu, Feb 18, 2016 at 8:58 PM, Andrew Munn <[email protected]> wrote:
       >       What would be the best way to use Geode (or GF) to store and 
utilize
       >       financial time series data like a stream of stock trades?  I 
have ASCII
       >       files with timestamps that include microseconds:
       >
       >       2016-02-17 
18:00:00.000660,1926.75,5,5,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,80,85,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,1,86,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,6,92,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,27,119,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,3,122,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,5,127,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,4,131,1926.75,1926.75,14644971,C,43,01,
       >       2016-02-17 
18:00:00.000660,1926.75,2,133,1926.75,1926.75,14644971,C,43,01,
       >
       >       I have one file per day and each file can have over 1,000,000 
rows.  My
       >       thought is to fault in the files and parse the ASCII as needed.  
I know I
       >       could store the data as binary primitives in a file on disk 
instead of
       >       ASCII for a bit more speed.
       >
       >       I don't have a cluster of machines to create an HDFS cluster 
with.  My
       >       machine does have 128GB of RAM though.
       >
       >       Thanks!
       >
       >
       >
       >
       > --
       > Greg Chase
       > Global Head, Big Data Communities
       > http://www.pivotal.io/big-data
       >
       > Pivotal Software
       > http://www.pivotal.io/
       >
       > 650-215-0477
       > @GregChase
       > Blog: http://geekmarketing.biz/
       >
       >
       >
       >
       >
       >



























					Reply via email to