Cool. 

Brad Davis
Red Hat Consulting
Email: bdavis@redhat.com | c:980.226.7865 |http://www.redhat.com 

On May 9, 2014, at 1:41 PM, Jess Sightler <jsightle@redhat.com> wrote:

I'm sorry for not responding sooner... somehow I missed this email earlier.

Basically I have added a new Type Manager (GraphTypeManager) that
augments the TypeField/TypeValue functionality to support the resolution
of multiple types. It also uses a addon scanner to avoid the need to
hardcode a list of classes for TypedGraphModuleBuilder.

So far this appears to work will for the windup usecase.

On 05/05/2014 01:15 PM, Brad Davis wrote:
Gotcha.  In the current design, it uses TypedGraphModuleBuilder, which is aware of the type via the TypeField annotation, and therefore sets it to one "type" of object.  In the current design, I took a similar approach to what you are saying, but accomplished that by the concept of "facets."

com.tinkerpop.frames.modules.typedgraph.TypeField;
com.tinkerpop.frames.modules.typedgraph.TypeValue;

I think if you were trying to accomplish what you are describing, you would need to implement type resolvers (I imagine), and rip out all of the @TypeValue annotations.  But haven't done this before in Frames, so it would take some investigation.  We should definitely make it easy for those integrating with the Windup API to resolve the appropriate type.  Just keep that in mind.  But, if there is some sort of DAO that lets you fetch all elements of a given type & casts appropriately, that might be all that is required.

Brad Davis
Red Hat Consulting
Email: bdavis@redhat.com | c: 980.226.7865 | http://www.redhat.com


----- Original Message -----
From: "Jess Sightler" <jsightle@redhat.com>
To: windup-dev@lists.jboss.org
Sent: Monday, May 5, 2014 11:30:06 AM
Subject: Re: [windup-dev] Let's replace the GraphVisitor interface concept with something generic

I think you are not reading this correctly. That's ok, though, as I
probably didn't describe it correctly. :) I am talking about supporting
types, but supporting them using a model more similar to OWL/RDF than to
classical OO-devleopment. In this model, you do not know all of the
typing information up front, and instead infer type information as you go.

So, for example, a Vertex may evolve over time to support additional
types as more information is learned. For example, the first rule may
provide a "File", but it later may become "File and ZipFile", and then
ultimately "File and ZipFile and EarFile".

During each step of this evolution it retains all of it's
characteristics of the previous step. You can think of it as subclassing
if you want, but eventually the Java model of subclassing will mislead
you about what this means, as it is not necessary for new types to be
subclasses of existing types.

Frames encompasses this concept, and this is the big reason that it uses
Interfaces instead of classes, AFAICT.

This document may be helpful as well:
http://www.w3.org/TR/sw-oosd-primer/#comparison


On 05/05/2014 11:00 AM, Brad Davis wrote:
I don't know that I agree.  I think it makes sense if we had a black box, but if you are giving Windup API to other people to mash up, I think it might be hard to read a map of many different types of content with many different properties without structure.  That is what having real entities provide you.  There are only so many types of entities that are present that matter today.

I think by making it untyped, if I am reading this correctly, and just making it meta objects, it will lead to more complexities for those trying to integrate Windup into projects downstream.

Brad Davis
Red Hat Consulting
Email: bdavis@redhat.com | c: 980.226.7865 | http://www.redhat.com


----- Original Message -----
From: "Lincoln Baxter, III" <lincolnbaxter@gmail.com>
To: "Windup-dev List" <windup-dev@lists.jboss.org>
Sent: Monday, May 5, 2014 10:34:28 AM
Subject: Re: [windup-dev] Let's replace the GraphVisitor interface concept with something generic

Yeah, I actually have to agree with Jess here. I think that the "add more and more information as you go" concept is probably the simplest, and also makes sense from the point-of-view of "improving" what we know the more rules are run.


On Sun, May 4, 2014 at 7:30 PM, Jess Sightler < jsightle@redhat.com > wrote:






On May 2, 2014 5:31 PM, Ondrej Zizka < ozizka@redhat.com > wrote:
On 2.5.2014 21:17, Lincoln Baxter wrote:
Hey Ondra,

I'm moving this discussion to windup-dev@lists.jboss.org - please use windup-dev for all non-confidential development discussion.

In general I support your idea to replace the GraphVisitor interface - I don't think any of us is suggesting that we continue with that approach. But I am concerned that it is more complicated than necessary, and I do have a few concerns about what you've mocked up below:

#1) Your concept of replacing nodes is interesting, but what is stopping a rule from replacing a node, then subsequently overwriting or being overwritten by another replacement? I see potential for multiple rules to interfere with each other's types in this way.
That's the purpose. The rules would interfere. Or rather, infer.
They would have to be written in a way that they would not do that. You can write an EJB which kills JVM. You can write HQL query which will delete all your entities. You can call wait() in the DSL. Let's assume that user will not write silly rules. Let's give them some freedom and see the rules flourish.

I prefer to think of this in terms of RDF-style classes that operate on the open world assumption: http://en.m.wikipedia.org/wiki/Open_world_assumption

In this model, you don't replace types, you add additional type information, and inferred properties as you discover more things about a particular entity. If you'd like, I can try to put together a little mini presentation on rdf and owl, as I believe we may be re-solving some problems that those groups have already solved pretty well. :-)

From my research on Friday, I believe that this OWA type model may actually be supported pretty well by Tinkerpop frames.

We can discuss further next week, as I am sure the above is insufficient. :-)




Further - the replacement would simply happen for certain cases. E.g. Maven POM file will never be anything else, so a XmlFileNode can be changed to MavenPomFileNode. In other cases, rules would create new nodes and connect them. Also - these possible collisions can be easily detected - if some migrator asks for a XmlFileNode, and then asks to change it to some subtype while it already is another subtype, -> warning or error.

#2) I'm also not sure that the graph would allow you to dynamically replace nodes of one type with nodes of another. Can you verify that?
I didn't find anything related in https://github.com/thinkaurelius/titan/wiki/Titan-Limitations (those type limitations seem to apply on other things). I'll try practically.
#3) You asked: "Forge UI could be mapped to XML elements. I.e. <aaa foo="bar"/> could invoke command "aaa" with given params.
I believe Forge already has this way of input, right?"

I'm not completely following your example here, but in theory you could map XML elements and their attributes to forge commands and their options in this way; nonetheless, I don't think that this is a good use of the Forge command model. You're better off just mapping to Java objects of some type.
Ok, I thought forge UI could be the way to do the mapping, since it has the type conversion already done, but we can duplicate that outside Forge, too.
#4) This seems more complicated than the example Visitor you linked. Now instead of one Java file containing the rule, and two java interfaces to encapsulate the data storage in the graph, you have two very convoluted XML files. I actually think that the JAXB bit is fairly nice, but the code required to do that in your example would work in the current approach anyway because it would still need to be implemented somewhere.
Yes and no.

1) The Visitor is custom Java code and of course, with access to all Java libraries in the world, you can make it matter of few lines. But we are heading towards rules which allow non-trivial operations while being limited to just several concepts. In this proposal, free-form Java code would be replaced with those few mentioned: Service calls, Graph queries, JAXB, EL, iteration, rules dependency,
IMO, if we don't create something such, we will end up with rules which will be just a thin wrapper around services.

2) This XML example doesn't map to the visitor 1:1, it's better. The approach is different. In that java code, there are two tasks mixed into one: Discover Maven POM files, and load the dependencies.
What if the pom files will be added by a custom migrator? E.g. when they have different doctype. Then this visitor would miss them and not scan their dependencies.
In this approach, the second rule would pick up the MavenPomFileNode no matter where it comes from.
I could have written the example 1:1 but wanted to show this task separation advantage.


This is similar to what I am prototyping in the config addon, but in XML not in Java. If you want to continue with this idea of reducing the operations to operate more closely with the graph, I support that, but let's please try to find a way to mock it up using the config DSL instead.
Okay, let's see what we have.
As far as implementing this goes - the syntax you've described below would probably work by mapping to our Java DSL using Reflection, but that's to be done once the java config API is established.

Java first. Then XML (or whatever).
Hmm. About 4 months ago, it was exactly opposite: XML rules, no Java.
It was also one of the main reasons to abandon WindRide the XML rules were implemented.
The argument was that rules authors will not create Java projects and study some framework (e.g. Forge) to be able to write a trivial rule like "if com.foo.Bar is found, report a warning with a comment and a link to docs XY".
When did this change, and what's the guarantee that it won't change again?

Ondra

~Lincoln

________________________________
From: "Ondrej Zizka" < ozizka@redhat.com >
To: jboss-migration@redhat.com
Sent: Thursday, May 1, 2014 12:05:23 AM
Subject: Re: Let's replace the GraphVisitor interface concept with something generic

I've put it to this doc so we can edit/comment there.

https://docs.google.com/document/d/1UOihPv_zryFilCb7T0k8992wPUt3YNP4goh9rxTC7ng/edit#

Ondra



On 1.5.2014 04:35, Ondrej Zizka wrote:
Hi,

as we discussed before, I'd like to replace the GraphVisitor interface with something generic.
Seriously, having hard-coded interface with methods specific for e.g. MessageDrivenBean, EjbEntity, SpringConfiguration, etc. is IMO not the way to go. It is hard to extend. A rule system created around this would be cumbersome.

public void visitEjbEntity(EjbEntityFacet entry);
public void visitEjbService(EjbSessionBeanFacet entry);
public void visitMessageDrivenBean(MessageDrivenBeanFacet entry);
public void visitEjbEntity(SpringBeanFacet entry);
...



Instead, we should focus on the graph, and have just very few node types in the core - FileNode, and the rest would be subtypes defined in addons. Addons would, through rules:
* replace a node with a subclass, e.g.
FileNode => XmlFileNode => MavenPomNode,
FileNode => JavaFileNode => AnnotationNode

* add properties, e.g.
XmlFileNode's "doctype",
ClassNode's "blacklisted"
* connect nodes to them, e.g.
MavenPom ---contains---> MavenDependencyNode
JavaFile --- imports --> [ ClassNode, ClassNode, ... ]

This approach would:
* Leverage of Forge modularity (e.g. Maven addon depending on XmlFile addon)
* Improve extendability (no need to squeeze everything into the GraphVisitor interface's methods or extend it)
* Lead to much more straightforward rules implementation - all rules would reduce to:
* matching graph information (Gremlin?)
* using DAO's / Services (for mining data from the files/..., and for writing them during active migration)
1) Bundled - XPath, AST query, properties, Maven remote fetch, ...
2) User's: .class packed within the addon or a Groovy class
* writing back to the graph
* rendering pieces of HTML for the report.

Who's in? I need some scenarios where this wouldn't work. But from what I can tell, this would be more generic, but still simpler, than current "God-object" suffering GraphVisitor.

As an example, take e.g. MavenRemoteFetchVisitor. https://github.com/windup/windup/blob/master/engine/rules/impl/src/main/java/org/jboss/windup/engine/visitor/inspector/MavenRemoteFetchVisitor.java

All that is doable using few simple building blocks, directed by few lines of a rule like this:

------------------------------------------------------------------

<var name="pomNS" val=" http://maven.apache.org/POM/4.0.0 ">

Rule 1) which would process all POM files and load the info into the graph.

<rule id="maven.pomFiles" desc=" Analyze Maven POM files "
phase="StaticConfigAnalysis">
<graph match="XmlFileNode[doctype=' http://maven.apache.org/POM/4.0.0' ]" toVar="xmls">
<for var="pom" in="xmls">
<graph:replaceSingle ref="pom" with="MavenPomFileNode">
<properties>
<!-- <jaxb> would invoce a call to a Service.
<properties> hander would take returned object's bean props.
toClass would load the class from CL or compile from .groovy coming with the migrator (addon).
-->
<jaxb toClass="PomJaxb.groovy" fromFile="${pom.path}">
<ns name="pom" uri="${pomNS}"/>
</jaxb>
</properties>
</graph>
</for>
</rule>

@XmlRoot
class PomJaxb {
@XmlXPath("/pom:project/pom:modelVersion") String modelVersion;
@XmlXPath("/pom:project/pom:name") String name;
@XmlXPath("/pom:project/pom:description" String description;
@XmlXPath("/pom:project/pom:url" String url;
}


Rule 2) which would load the dependencies and describe them into Nodes and Edges.

<rule id="maven.pomDependencies" desc=" Analyze Maven POM file dependencies "
phase="StaticConfigAnalysis">
<after rule="maven.pomFiles">

<graph match="MavenPomFileNode" toVar="pomVertexes">
<for var="pomV" in="pomVertexes">

<xpath toVar="deps" fromFile="${pomV.path}" match="/pom:project/pom:dependencies/pom:dependency"/>
<for var="depElement" in="deps">
<jaxb toVar="dep" toClass="MavenDepencencyJaxb.groovy" fromElement="depElement" ns="pom ${pomNS}"/>
<graph:query toVar="isBlacklisted"
q=" /* I don't know Gremlin so far, imagine an equiv of this XPath: */
MavenDependencyNode[
g=${dep.groupId} and a=${dep.artifactId} and v=${dep.version}
]@blacklisted
" />
<continue if="isBlacklisted /* var, or Groovy (or EL) expression */" />
<!-- This would be useful for blacklists and filters in general, which appear often in real life rules. -->

<graph:insertIfNotExists type="MavenDependencyNode" toVar="depVertex>
<properties from="dep"/>
</graph>
<graph:edge type="dependsOn" from="pomV" to="depVertex"/>
<!-- Maybe Gremlin could replace this? -->
</for>
</for>
</rule>

@XmlRoot
class MavenDepencencyJaxb {
// GraphKeyProperty identifies those which are compared for insertIfNotExists.
@GraphKeyProperty @XmlXPath("./pom:groupId") String groupId;
@GraphKeyProperty @XmlXPath("./pom:artifactId") String artifactId;
@GraphKeyProperty @XmlXPath("./pom:version") String version;
}
------------------------------------------------------------------

As you can see, It creates independent rules which only communicate indirectly through the graph.
You can also see how nicely Java classes fit into this, and how Groovy could make this easier.

SERVICES INVOCATION
Forge UI could be mapped to XML elements. I.e. <aaa foo="bar"/> could invoke command "aaa" with given params.
I believe Forge already has this way of input, right?

GRAPH OPERATION
There would be several <graph:...> operations - CRUD plus some special.

EXECUTION FLOW
The flow would be simple, from top to bottom, creating variables along the way, containing objects or iterable collections of objects. Those iterable could be used in <for>.

Does Lincoln's executor fit this? I haven't still looked how it works. This tree would likely be executed classically with a stack and using tree reduction for operation arguments.

For more complex logic, users would break the task into multiple rules, storing data into the graph intermediately.

I'll check few more visitors to see if this is powerful enough to satisfy all the baneeds.


.............................

Also, I'd like to eradicate any mention of an archive from most of the code - archives should be totally transparent for the migrators. There should be just FileNodes, connected with ArchiveNodes, and whoever needs an information that a file came from an archive, may look that up. See https://docs.google.com/drawings/d/1IMnds3Qu8Wwcf7_mr7NJ9a3YgtcGJ7dejl09EhWl7Vc for illustration.

Ondra






_______________________________________________
windup-dev mailing list
windup-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/windup-dev



_______________________________________________
windup-dev mailing list
windup-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/windup-dev
_______________________________________________
windup-dev mailing list
windup-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/windup-dev

_______________________________________________
windup-dev mailing list
windup-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/windup-dev