predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mars Hall <m...@heroku.com>
Subject Re: Ecom recommender adding like event
Date Tue, 10 Jan 2017 21:03:17 GMT
Hi Daniel,

If you think there's a problem with the PredictionIO buildpack for Heroku, please file as
issue on its GitHub repo https://github.com/heroku/predictionio-buildpack/issues

If it works locally but not on Heroku, my wild guess is that one of your algorithms is serializing
the model to the filesystem, and it's throwing an error when trying to retrieve that model.
Heroku does not support filesystem persistence between runs (e.g. files committed during release/train
will not persist to web/deploy)

Perhaps try enabling S3 HDFS for filesystem persistence https://github.com/heroku/predictionio-buildpack/blob/master/CUSTOM.md#optional-persistent-filesystem
Note that with HDFS filesystem path references must be from `/` root, not nested in a User
ID directory.

*Mars

( <> .. <> )

> On Jan 10, 2017, at 00:53, Daniel O' Shaughnessy <danieljamesdavid@gmail.com> wrote:
> 
> Hi Pat,
> 
> Ya I guess it's an issue with the pio heroku buildpack. It seems to build and train OK
based on the logs but then throws that error when trying to deploy.
> 
> I'm not actually using the similar products template I'm using this: https://github.com/apache/incubator-predictionio-template-ecom-recommender
> 
> I'm trying to extend the ecom template to include 'likes' to try and improve the results
and using the tutorial as a guideline. I'm using this for an app store recommender...Ideally
I'd be using the UR but I can't get this up on Heroku so far, when AWS approve it I may go
down that route for hosting eventually.
> 
> On Mon, 9 Jan 2017 at 20:44 Pat Ferrel <pat@occamsmachete.com> wrote:
> AFAIK, the Heroku support is fairly new so not surprising that a multiple engine template
has trouble. Multiple Algorithms in a Template are fairly rare. I’d ask Heroku what is going
on. 
> 
> BTW what are you using this for? There are other ways to get similar products—some
of which may work better. For instance is there a reason for multiple algorithms? 
> 
> 
> 
> On Jan 9, 2017, at 10:45 AM, Daniel O' Shaughnessy <danieljamesdavid@gmail.com>
wrote:
> 
> Hi,
> 
> I'm attempting to add a like/dislike event to the ecom recommender ( as per http://predictionio.incubator.apache.org/templates/similarproduct/multi-events-multi-algos/
). This method combines the results of 2 separate algorithms in the Serving class.
> 
> When I deploy this using the predictionIO Heroku buildpack it seems to deploy and run
the release command OK but when I check the Heroku logs it fails here:
> 
> 2017-01-09T15:32:44.905195+00:00 app[web.1]: [INFO] [Engine] Using persisted model
> 
> 2017-01-09T15:32:44.912783+00:00 app[web.1]: [INFO] [Engine] Loaded model com.dos.sfdc.ECommModel
for algorithm com.dos.sfdc.ECommAlgorithm
> 
> 2017-01-09T15:32:45.307111+00:00 app[web.1]: [INFO] [Engine] Custom-persisted model detected
for algorithm com.dos.sfdc.LikeAlgorithm
> 
> 2017-01-09T15:32:45.668678+00:00 app[web.1]: [ERROR] [OneForOneStrategy] Cannot call
methods on a stopped SparkContext.
> 
> 2017-01-09T15:32:45.668690+00:00 app[web.1]: This stopped SparkContext was created at:
> 
> 2017-01-09T15:32:45.668691+00:00 app[web.1]: 
> 
> 2017-01-09T15:32:45.668692+00:00 app[web.1]: org.apache.spark.SparkContext.<init>(SparkContext.scala:82)
> 
> 2017-01-09T15:32:45.668693+00:00 app[web.1]: org.apache.predictionio.workflow.WorkflowContext$.apply(WorkflowContext.scala:45)
> 
> 2017-01-09T15:32:45.668694+00:00 app[web.1]: org.apache.predictionio.workflow.CreateServer$.createServerActorWithEngine(CreateServer.scala:228)
> 
> Anyone come across this before?
> 
> 
> 
> Thanks,
> 
> 
> 
> Daniel.
> 
> 


Mime
View raw message