Open Stack

On 01/20/2014 12:50 PM, Andrey Lazarev wrote:
> Inlined.
>
>
> On Mon, Jan 20, 2014 at 8:15 AM, Matthew Farrellee <matt at redhat.com
> <mailto:matt at redhat.com>> wrote:
>
>     (inline, trying to make this readable by a text-only mail client
>     that doesn't use tabs to indicate quoting)
>
>     On 01/20/2014 02:50 AM, Andrey Lazarev wrote:
>
>                  ------
>                  FIX - @rest.get('/jobs/config-hints/____<job_type>') -
>         should move to
>                  GET /plugins/<plugin_name>/<____plugin_version>, similar to
>                  get_node_processes
>                  and get_required_image_tags
>                  ------
>                  Not sure if it should be plugin specific right now. EDP
>         uses it
>                  to show some
>                  configs to users in the dashboard. it's just a cosmetic
>         thing.
>                  Also when user
>                  starts define some configs for some job he might not define
>                  cluster yet and
>                  thus plugin to run this job. I think we should leave it
>         as is
>                  and leave only
>                  abstract configs like Mapper/Reducer class and allow
>         users to
>                  apply any
>                  key/value configs if needed.
>
>
>              FYI, the code contains comments suggesting it should be
>         plugin specific.
>
>         https://github.com/openstack/____savanna/blob/master/savanna/____service/edp/workflow_creator/____workflow_factory.py#L179
>         <https://github.com/openstack/__savanna/blob/master/savanna/__service/edp/workflow_creator/__workflow_factory.py#L179>
>
>         <https://github.com/openstack/__savanna/blob/master/savanna/__service/edp/workflow_creator/__workflow_factory.py#L179
>         <https://github.com/openstack/savanna/blob/master/savanna/service/edp/workflow_creator/workflow_factory.py#L179>>
>
>              IMHO, the EDP should have no plugin specific dependencies.
>
>              If it currently does, we should look into why and see if we
>         can't
>              eliminate this entirely.
>
>         [AL] EDP uses plugins in two ways:
>         1. for HDFS user
>         2. for config hints
>         I think both items should not be plugin specific on EDP API
>         level. But
>         implementation should go to plugin and call plugin API for result.
>
>
>     In fact they are both plugin specific. The user is forced to click
>     through a plugin selection (when launching a job on transient
>     cluster) or the plugin selection has already occurred (when
>     launching a job on an existing cluster).
>
>     Since the config is something that is plugin specific, you might not
>     have hbase hints from vanilla but you would from hdp, and you
>     already have plugin information whenever you ask for a hint, my view
>     that this be under the /plugins namespace is growing stronger.
>
>
> [AL] Disagree. They are plugin specific, but EDP itself could have
> additional plugin-independent logic inside. Now config hints return EDP
> properties (like mapred.input.dir) as well as plugin-specific
> properties. Placing it under /plugins namespace will give a vision that
> it is fully plugin specific.
>
> I like to see EDP API fully plugin independent and in one workspace. If
> core side needs some information internally it can easily go into the
> plugin.

I'm not sure if we're disagreeing. We may, in fact, be in violent agreement.

The EDP API is fully plugin independent, and should stay that way as a 
project goal. config-hints is extra data that the horizon app can use to 
help give users suggestions about what config they may want to 
optionally add to their job. Those config options are independent of the 
job and specific to the cluster where the job will run, which is the 
purview of the plugin.

Moving config-hints out of the EDP API will make this even more clear.

Best,


matt