[
https://issues.apache.org/jira/browse/FALCON-1107?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15141901#comment-15141901
]
Sowmya Ramesh commented on FALCON-1107:
---------------------------------------
Opening up for discussion about where to publish the recipe artifacts.
Recipe artifacts repo is structured as
{noformat}
RecipeRoot
|-- Recipe1
|-- README
|-- META
|-- libs
|-- build
|-- runtime
|-- resources
|-- build
|-- runtime
{noformat}
Falcon server should be aware of where recipe artifacts are hosted as it is
required for cooking the recipe and also to provide recipe repository
management feature. Multiple clients can use the same recipe, so this should be
hosted in one centralized location.
For trusted recipes which are provided by Falcon OOTB the artifacts are
published by Falcon and for custom recipes instructions are provided to the
user about where to publish the artifacts.
Also, this should be designed so that it works for both unsecure and secure
cluster setup.
Today, in Falcon during the process entity validation its assumed that
workflow[WF] and WF libs are present on the cluster where the process instance
runs. This has to change as in case of recipes WF and libs can reside on
different cluster. Also, in case of secure cluster, required NN principals
should be passed to access the file and configuration
"mapreduce.job.hdfs-servers" should be updated for job execution to succeed.
One approach is similar to config store uri introduce another config for recipe
store uri in startup properties and another config to set the NN principal. For
trusted recipes Ambari can be used to copy the required artifacts to this
location. In case of custom recipes user has to manually copy them. One issue
with this approach is that since its configurable if user changes this config
later say when implementing custom recipes, it will break the trusted recipes
if the artifacts are not copied to new recipe location.
Please let me know if there any suggestions or better approaches, thanks!
cc: [~sriksun], [~venkatnrangan]
> Move trusted recipe processing to server side
> ---------------------------------------------
>
> Key: FALCON-1107
> URL: https://issues.apache.org/jira/browse/FALCON-1107
> Project: Falcon
> Issue Type: Sub-task
> Reporter: Sowmya Ramesh
> Assignee: Sowmya Ramesh
> Labels: Recipe
> Fix For: trunk
>
> Attachments: ApacheFalcon-RecipeDesignDocument.V1.pdf,
> ApacheFalcon-RecipeDesignDocument.pdf
>
>
> Today Recipe cooking is a client side logic. Recipe also supports extensions
> i.e. user can cook his/her own custom recipes.
> Decision to make it client side logic was for the following reasons
> * Keep it isolated from falcon server
> * As custom recipe cooking is supported, user recipes can introduce
> security vulnerabilities and also can bring down the falcon server
> Today, falcon provides HDFS DR recipe out of the box. There is a plan to add
> UI support for DR in Falcon.
> Rest API support cannot be added for recipe as it is client side processing.
> If the UI is pure java script[JS] then all the recipe cooking logic has to be
> repeated in JS. This is not a feasible solution - if more recipes are added
> say DR for hive, hbase and others, UI won't be extensible.
> For the above mentioned reasons Recipe should me made a server side logic.
> Provided/Trusted recipes [recipes provided out of the box] can run as Falcon
> process. Recipe cooking will be done in a new process if its custom recipe
> [user code].
> For cooking of custom recipes, design proposed should consider handling
> security implications, handling the issues where the custom user code can
> bring down the Falcon server (trapping System.exit), handling class path
> isolation.
> Also it shouldn't in anyway destabilize the Falcon system.
> There are couple of approaches which was discussed
> *Approach 1:*
> Custom Recipe cooking can be carried out separately in another Oozie WF, this
> will ensure isolation. Oozie already has the ability to schedule jobs as a
> user and handles all the security aspects of it.
> Pros:
> - Provides isolation
> - Piggyback on Oozie as it already provides the required functionality
> Cons:
> - As recipe processing is done in different WF, from operations point of view
> user cannot figure out recipe processing status and thus adds to the
> operational pain. Operational issue with this approach is said to be the
> overall
> apparatus needed to monitor and manage the recipe-cooking workflows.
> Oozie scheduling can bring arbitrary delays Granted we can design around the
> limitations and make use of the strengths of the approach but it seems
> something we can avoid if we can.
> - There has been few discussions to move away from Oozie as scheduling engine
> for Falcon. If this is the plan going forward its good not to add new
> functionality using oozie.
> *Approach 2:*
> Custom recipe cooking is done on the server side in a separate independent
> process than Falcon process I.e. It runs in a different JVM. Throttling
> should be added for how many recipe cooking processes can be launched keeping
> in mind the machine configuration.
> Pros:
> - Provides isolation as recipe cooking is done in a independent process
> Cons:
> - Performance overhead as new process is launched for custom recipe cooking
> - Adds more complexity to the system
> This bug will be used to move recipe processing for trusted recipes to server
> side.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
