On Tue, Jul 10, 2012 at 8:15 PM, Patrik Sundberg wrote:I've done a bit more studying and I think a combination of SystemAnnouncer events to catch method changes, pragmas to mark "graph nodes" and use of method wrappers could create a very nice way to shield the user of my graph API from thinking of identity caches and graph node details.
The other cool feature of use is the new class builder and custom slots. For nodes that are not dynamic/calculations (what I call properties or relationship nodes) I think custom slots could be interesting but need to test out some ideas to organize thoughts. Obviously won't work for calculated/dynamic nodes since I need ability to take arguments for those and can't map to a slot. In principle all I'm after is to have an identity map and a graph (DAG) in the background whose management is hidden away from user so he can focus on domain logic and not graph stuff.
Going on holiday for a week and a bit, shall start experimenting after that.
Ok, so I'm back on this. My first idea is to do this:- Every node in my graph is either a piece of data (I call it a leaf or a property node) or a node depending on other nodes that perform a calculation (I call it a calc node for now)- All the nodes are part of an identity map which means a node can only exist in one copy and any relations and dependencies will sit between these unique nodes- I'd like to shield the user from all the identify map stuff etc as much as possible.- A smalltalk object is an instance of a class, and it can participate in the graph by providing leaf nodes and calc nodes. Those nodes are associated with the messages of the object.- I have some ideas for how to bootstrap the dependency graph between nodes etc and I'm not worried about that bit (on first calculation create structure via reflection and a dependency stack)- The explicit and uggly way I can do things are like this:Foo>>addOneTo: aNumberNode� �^GraphManager instance�� � � �getCalcNode: #addOneTo:� � � �for: self�� � � �withArgs: #( aNumberNode )� � � �block: [ :node |� � � � � node value + 1� � � �]� -�pretend this goes and asks the identity cache and either gets an existing leaf node for the addOneTo: for this object for a particular argument, or if none exist it creates it and attaches the block as the calculation to perform when #value is sent to the node� - let's say 'foo' is an instance of Foo class, then 'foo addOneTo: baz' refers to the calc node, and '(foo addOneTo: baz) value' refers to the value of the calculation- What I was thinking here was that I could use method wrappers to hide all the junk, like this:Foo>>addOneTo: aNode� �<calcNode>� �aNode value + 1- I listen to all method add/update/remove announcements, and when a method that is marked with the <calcNode> pragma is involved, what I do is create a MethodWrapper that takes care of all the uggly stuff to get/create the graph node via the identity cache and as #value method for the node uses the method implementation the user originally provided (and replaces the entry in the method dictionary)Does that sound doable? The browsers etc wont go nuts?If I do that, will system browsers still show the original source for the method even though I've replaced it via the method dictionary and MethodWrapper?I'm assuming stack traces and debugging etc will be just fine, will just have the "hidden" node logic via the MethodWrappers included in the call stack.�I'll give that type of setup a go, just wanted to throw it out there and hear if it's an obviously terrible idea first.P.S. for leaf nodes (or any node not requiring an argument to be provided) I think I could use custom slots instead of method wrappers and do the same kind of logic as the method wrapper would do in the slot instead. I'll worry about that later.�On Jul 6, 2012 1:16 PM, "Patrik Sundberg" wrote:On Fri, Jul 6, 2012 at 12:38 PM, S Krish��wrote:You should read as much as you can about Kapital of JPMC. It does more or less what you describe for esoteric Prime Interest Derivatives and very scalably ..�I've heard of that one. Never worked at JPM so haven't seen it first hand though. I'll see what I can find. In terms of concepts I'm fully on top of it (have a background with other companies using similar ideas, but not in smalltalk but their own in-house language). So the good part is that I know exactly what I want conceptually, but I'm not stuck into smalltalk enough to know the natural way to implement the concepts. It may sound like a huge�endeavour but I've done a lot of it in ruby and even as a side project to my real job we're talking about a couple of months to have something useful.��I doubt if there is much in the public domain, but the fact that Kapital is the golden standard in this product line talks highly of why Smalltalk is THE platform for this kind of product.�I wouldn't say it's the golden standard, that's taking it a bit far, there are others that I rate higher bit I'm also biased :) But I do agree that smalltalk is a reasonably natural platform for these types of concepts, hence why I'm exploring it.�It would be nice to have a Pharo based derivative product that can easily beat Ruby/ Java / .Net at this dynamic visualization using standard browsers, where Nautilus too can serve your purpose.. or a home grown tree based browser.../ Grids..That's my hunch, that the tooling work would be much cut down. In my current ruby mockup I'm kind of creating a poor man's image like experience - it's jruby on the JVM, vim + vim-slime to send code fragments to the "runtime", git, and most likely homegrown GUI to navigate the graph and play with it (not done anything on that yet). I'll be keeping jruby as a dear tool as right now it's a fantastic glue tool in terms of interacting with anything out there in the world (i.e. I can easily work with the bloomberg java API), but I'm very keen to explore pharo for my graph + tooling around that.�(slight tangent - coming from outside the smalltalk world it'd be HUGE to me to have git backend for packages and use of github for collaboration. seen that mentioned in other threads. I think it'd be HUGE for community's ability to attract more people).�Once I have completed my run at a PharoTabletIDE/ PharoMorphicView framework, I would love to collaborate on your endeavour to see if it can be modelled on it and exposed..Cool. Early days yet and we'll see where I end up. Not familiar with those projects, I shall take a look.
PatrikOn Fri, Jul 6, 2012 at 4:08 PM, Patrik Sundberg wrote:Hi,
First of all, sorry for the long email as my first email to the list.
Hopefully some people will find it an interesting discussion :)
I'm a long time programmer that has been studying smalltalk and pharo over
the last year. It's a beautiful language I'm liking the image style of
development a whole lot. I've been looking for a good test case to try out
in pharo and I've got something in mind. I've used ruby since around 2000
and given how much it has been inspired by smalltalk it's not a big leap.
My real job is being a commodities trader but I build my own tools and
something I'm always working on is improved risk and pricing tools. Lots of
that comes down to 2 things:
1. evaluating a "function" (in a general sense) with inputs that are picked
depending on time. e.g. I want to price a security and I want to pick up
prices for time X as inputs to the pricing model.
2. calculating a derivative of the "function" above with respect to inputs
What I tend to do in order to keep things consistent by design, plus add
natural caching, is create a calculation graph. I define all the
dependencies and create an in-memory graph where each node is unique using
an identity map. I then perform a calculation where the calculations trickle
down the graph starting from the node I'm interested in down all it's
dependencies. To calculate the derivative I then find the node representing
the input I want the derivative with respect to and use finite differences,
e.g. move it's value, and recalculate the top level node. On the second
valuation of the graph only the parts that are affected by me changing that
1 value will need to be evaluated, the rest is cached an unaffected. It
makes it very easy to ask questions like "How much will I be affected by X
changing by Y?" since I can take the top level node of the graph, search for
X, and if found change it by Y and recalculate.
How the graph is constructed in memory depends on time (and a few other
things). I always store all previous states of the world so that I can
recreate any calculation I did in the past. Very useful for diagnosing
problem. Hence if I for example change what model I use to model something,
that change is recorded in a way such that if I set my "time" to before the
change happened the graph will get created as it would have with the old
model, and for a time after the change it'll create a different graph
reflecting the new model. Hence the graph can only be known at runtime when
the "time" etc is known.
I currently have a ruby mockup of this. It's a DSL that looks like this:
----------- EXAMPLE
class ExampleObject
� include GraphEntity # includes a module with functionality for objects
participating in the graph, including #property and #calc used below
� property :foo do
� � 123 # this is the default value for this property which will be used
before a value has been set and saved
� end
� calc :bar do |arg|
� � foo.value + arg � � # take the value of the arg property and add the
argument given to the calculation node
� end
end
o = ExampleObject.new
# this will kick of the graph being built in memory and setup the dependency
between bar and foo nodes
o.bar(1).value �# -> 124
# this will be a "cache lookup" since nothing in the graph that bar depends
on has changed (i.e. the "expensive" calculation is not performed again
o.bar(1).value �# -> 124
o.foo.set_value(2)
o.bar(1).value �# -> 125, realizes that foo changed and performs a recalc
--------------------------------------------------------
To accomplish this I use dynamic code generation like the below for
#property and similar for other types of nodes:
� � � � # NOTE: it's a class level method, hence how I can call it when
defining the class in the example
� � � � def property(name, time = CdrEnv.instance.time, &block)
� � � � � clear_property(name)
� � � � � getter_method_body = <<-EOF
� � � � � � def #{name}
� � � � � � � init_block = self.class.property_init_block_for(:#{name})
� � � � � � � time_for_property =
self.class.property_time_dependency_for(:#{name})
� � � � � � � if init_block.nil?
� � � � � � � � CdrEnv.instance.get_property_node(self, :#{name},
time_for_property)
� � � � � � � else
� � � � � � � � CdrEnv.instance.get_property_node(self, :#{name},
time_for_property, &init_block)
� � � � � � � end
� � � � � � end
� � � � � EOF
� � � � � setter_method_body = <<-EOF
� � � � � � def #{name.to_s}=(value)
� � � � � � � #{name.to_s}.mutate(value)
� � � � � � end
� � � � � EOF
� � � � � class_eval getter_method_body
� � � � � class_eval setter_method_body
� � � � � register_property_node(name, time, &block)
� � � � end
Don't worry about the details or all the unfamiliar ruby, the main point is
that it creates an instance method that uses the singleton CdrEnv.instance
to either get or create the node representing the property from the graph
identity cache depending on if it exists or not.
>From a tooling point of view I think I'd love to work with this kind of
thing in pharo. Building my own browsers for inspecting and debugging my
dynamic graph should be very good fit. However, I'd appreciate some pointers
to idiomatic smalltalk to attack this kind of problem in terms of
implementing the graph itself - I obviously want the user (even if it's me)
to just have to focus on the domain model and hide as much as possible of
the graph bits under the covers the same way I've done with the code
generation stuff in my ruby example.
Any input into this would be much appreciated,
Patrik
P.S.
Btw, the persistence backend for this is the neo4j graph database, but it's
fronted by a service slotting into my own service framework built using
ZeroMQ and services exchanging messages in protobuf format. One can use it
from any language as long as one can send protobuf messages over zeromq. I
see there's a zeromq ffi library available on SS that I'll check out, but
I'm not finding a protobuf implementation. It'd be easy enough for me to
port a ruby protobuf implementation but I may as well ask if someone has
already done any work on protobuf for smalltalk?
--
View this message in context: http://forum.world.st/Dynamic-graph-exploration-seeking-input-on-idiomatic-smalltalk-LONG-tp4638702.html
Sent from the Pharo Smalltalk Users mailing list archive at Nabble.com.