predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pat Ferrel <...@occamsmachete.com>
Subject Re: Can I train and deploy on different machine
Date Wed, 29 Mar 2017 19:25:54 GMT
The Machine running the PredictionSever should not be configured to connect to the Spark Cluster.

This is why I explained that we use a machine for training that is a Spark cluster “driver”
machine. The driver machine connects to the Spark cluster but the PredictionServer should
not. 

The PredictionServer should have default config that does not know how to connect to the Spark
cluster. In this case it will default to running spark-submit to launch with MASTER=local,
which puts Spark in the same process with the PredictionServer and you will not get the cluster
error. Note that the PredictionServer should be configured to know how to connect to Elasticsearch
and HBase and optionally HDFS, only Spark needs to be local. Note also that no config in pio-env.sh
needs to change, Spark local setup is done in the Spark conf, it has nothing to do with PIO
setup.  

After running `pio build` and `pio train` copy the UR directory to *the same location* on
the PredictionServer. Then, with Spark setup to be local, on the PredictionServer machine
run `pio deploy` From then on if you do not change `engine.json` you will have newly trained
models hot-swapped into all PredictionServers running the UR.


On Mar 29, 2017, at 11:57 AM, Marius Rabenarivo <mariusrabenarivo@gmail.com> wrote:

Let me be more explicit.

What I want to do is not using the host where PredictionServer will run as a slave in the
Spark cluser.

When I do this I got "Initial job has not accepted any resources" error message.

2017-03-29 22:18 GMT+04:00 Pat Ferrel <pat@occamsmachete.com <mailto:pat@occamsmachete.com>>:
yes

My answer below was needlessly verbose.


On Mar 28, 2017, at 8:41 AM, Marius Rabenarivo <mariusrabenarivo@gmail.com <mailto:mariusrabenarivo@gmail.com>>
wrote:

But I want to run the driver outside the server where I'll run the PredictionServer.

As Spark will be used only for launching there.

Is it possible to run the driver outside the host where I'll deploy the engine? I mean for
deploying

I'm reading documentation about Spark right now for having insight on how I can do it but
I want to know if someone has tried to do something similar.

2017-03-28 19:34 GMT+04:00 Pat Ferrel <pat@occamsmachete.com <mailto:pat@occamsmachete.com>>:
Spark must be installed locally (so spark-submit will work) but Spark is only used to launch
the PredictionServer. No job is run on Spark for the UR during query serving.

We typically train on a Spark driver machine that is like part of the Spark cluster and deploy
on a server separate from the Spark cluster. This is so that the cluster can be stopped when
not training and no AWS charges are incurred. 

So yes you can and often there are good reasons to do so.

See the Spark overview here: http://actionml.com/docs/intro_to_spark <http://actionml.com/docs/intro_to_spark>


On Mar 27, 2017, at 11:48 PM, Marius Rabenarivo <mariusrabenarivo@gmail.com <mailto:mariusrabenarivo@gmail.com>>
wrote:

Hello,

For the pio train command, I understand that I can use another machine with PIO, Spark Driver,
Master and Worker.

But, is it possible to deploy in a machine without Spark locally installed as it is use spark-submit
during deployment
and 
org.apache.predictionio.workflow.CreateServer
references sparkContext.

I'm using UR v0.4.2 and PredictionIO 0.10.0

Regards,

Marius

P.S. I also posted in the ActionML Google group forum : https://groups.google.com/forum/#!topic/actionml-user/9yNQgVIODvI
<https://groups.google.com/forum/#!topic/actionml-user/9yNQgVIODvI>





Mime
View raw message