Trying to catch up on threads a bit...enjoying the discussion and I hope
I'm following along fully because it's sounding really nice. Letting the
type system be so open in previous versions of TinkerPop has created so
many inconsistencies and inelegant solutions which have only be exaggerated
by Gremlin Language Variants. Anyway, regarding:
>> 5. ComplexTypes don’t go over the wire — a ComplexTypeProxy with
>> appropriately provided toString() is all that leaves the TP4 VM.
>>
> As a tuple, ComplexTypes / ADTs go over the wire. The values of their
> primitive fields should probably go with them. However, the values of
their
> element / entity fields are just references; the attached element doesn't
> go with them.
I think I'd agree with Josh that we'd send these back over the wire,
especially if there is agreement that they are just a tuple form which
means that providers won't need to get into low-level serializer
development for custom types. TinkerPop would just know how to deal with
them for network transport. I guess providers would just have to provide
libraries with the ComplexType/ADT implementations in the programming
languages they wanted to support. In cases where they didn't, a user could
be left to work with a raw TinkerPop ComplexType/ADT instance which could
arguably be a better state than where they are left now which would be
serialization errors.
On Thu, Apr 25, 2019 at 2:07 PM Joshua Shinavier <j...@fortytwo.net> wrote:
> Hi Marko. Responses inline.
>
> On Wed, Apr 24, 2019 at 10:30 AM Marko Rodriguez <okramma...@gmail.com>
> wrote:
>
> > Hi,
> >
> > I think I understand you now. The concept of local and non-local data is
> > what made me go “ah!”
> >
>
> Nice. I also brought this up yesterday in the Property Graph Schema Working
> Group, where there is a discussion going on about whether/how graph
> databases can contain multiple graphs. Can an element belong to multiple
> graphs, can it have different properties in different graphs, etc. If each
> graph element is atomic, referencing other graph elements but not
> containing them, then it is very straightforward to think of a property
> graph as a simple set of elements. Graph relations are just set relations,
> making it easy to pull graphs apart and put graphs together (e.g. when
> building a stream, merging streams, etc.). If you are willing to make the
> open world assumption (e.g. "I know e[7] is a 'knows' edge, but I don't
> know what its out- and in-vertices are"), then you can't even partition a
> graph in such a way that the partitions are not valid graphs.
>
>
> So let me reiterate what I think you are saying.
> >
> > v[1] is guaranteed to have its id data local to it. All other information
> > could be derived via id-based "equi-joins.” Thus, we can’t assume that a
> > vertex will always have its properties and edges co-located with it.
>
>
> Yes indeed. A particular graph vendor may choose to co-locate properties
> with a vertex and edges with out- or in-vertex (or both, e.g. as JanusGraph
> does), but this is an optimization. At a logical level, you can think of an
> element and its dependents as belonging to separate relations.
>
>
>
> > However, we can assume that it knows where to get its property and edge
> > data when requested.
>
>
> Yes; you need to be able to select().
>
>
>
> > Assume the following RDBMS-style data structure that is referenced by
> > com.example.MyGraph.
> >
> > vertex_table
> > id label
> > 1 person
> > 2 person
> > …
>
>
> That is one way to go. I believe this scheme is what Ryan and David would
> call the Grothendieck construction; all relations of a given arity are
> marked with their type and concatenated into a single relation. I am still
> a little sketchy on the Grothendieck construction, so I hope that is a
> correct statement.
>
> However, you can also think of distinct element types (edge labels, vertex
> labels, property keys, hyperedge signatures) as distinct relations. So you
> instead of vertex_table, you would have
>
> person_table
> id
> 1
> 2
>
> Vertices are such trivial relations that they don't need to be stored as a
> tables. Edges are more interesting:
>
> knows_table
> out in
> 1 2
> 1 4
>
> Properties are similar:
>
> name_table
> out out_label out in
> 1 person marko
> 2 person vadas
> 3 project lop
> 4 person josh
> 5 project ripple
> 6 person peter
>
> The property table has a bit of a twist, because its out-label is a
> disjoint union of "person" and "project"; both persons and projects can
> have names, so you tag the out-element with its label/type. This is not
> necessary for "knows" because the out-label is always "person".
>
>
>
> > properties_table
> > id name age
> > 1 marko 29
> > 2 josh 35
> > …
> >
> > edge_table
> > id outV label inV
> > 0 1 knows 2
> > …
> >
>
> Yes, this also works, and is equivalent to what I wrote above, with one
> tweak: if tagged unions are supported (which IMO they should be, so we have
> both a "times" and a "plus" in our type algebra), then property_table
> should also include a "label" column, and edge_table should include
> "outLabel" and "inLabel", i.e.:
>
> properties_table
> id label name age
> 1 person marko 29
> 2 person vadas 27
> …
>
> and
>
> edge_table
> id label outV outLabel inV inLabel
> 0 knows 1 person 2 person
> …
>
> In a tagged union, you mark the type of a field along with the value or
> reference, for the sake of type checking and pattern matching.
>
> Hard to say whether the table-per-relation or the table-per-arity approach
> is better. FWIW, at Uber, we use a table per relation for the sake of
> better data isolation. If you want to take advantage of the physical types
> of the database, you may want multiple properties_tables, one per datatype
> (so you're not storing every property value as a string).
>
>
> If we want to say that the above data structure is a graph, what is
> > required of “ComplexType” such that we can satisfy both Neo4j-style and
> > RDBMS-style graph encodings? Assume ComplexType is defined as:
> >
> > interface ComplexType
> > Iterator<T> adjacents(String label, Object... identifiers)
> >
>
> You can think of a ComplexType as a row in a database. It just has the
> local fields specific to the type. In order to access attached elements,
> you need a select(), and your adjacents() looks pretty close to that. I
> would write:
>
> Iterator<T> adjacents(String label, String field)
>
> So for example, adjacents("knows", "out") from v[1] gives you an iterator
> of "knows" edges for which v[1] is the out-vertex. Btw. adjacents() here is
> the same as from() / comeFrom() in previous emails.
>
>
>
> > Take this basic Gremlin traversal:
> >
> > g.V(1).out(‘knows’).values(‘name’)
> >
> > I now believe this should compile to the following:
> >
> > [goto,V,1] [goto,outE,knows] [goto,inV] [goto,properties,name]
> >
>
> Polymorphism is cool. Your two-argument goto() appears to be my to(),
> whereas your three-argument goto() appears to be my from(). The minimal
> tweaks I would make to your syntax are:
>
> [goto,V,1] [goto,out,knows] [goto,in] [goto,out,name][goto,in]
>
> I might go a step further and say:
>
> [const,1] [goto,id,person] [goto,out,knows] [goto,in]
> [goto,out,name][goto,in]
>
>
>
> > Given MyGraph/MyVertex/MyEdge all implement ComplexType and there is no
> > local caching of data on these respective objects, then the bytecode
> isn’t
> > rewritten and the following cascade of events occurs:
> >
> > [...]
> >
>
> Looks pretty good.
>
>
>
> > Lets review the ComplexType adjacents()-method:
> >
> > complexType.adjacents(label,identifiers...)
> >
> > complexType must have sufficient information to represent the tail of the
> > relation.
> >
>
> Yes; it need to know what relation type you are matching, and on what field
> (e.g. "out"/"in") in that relation. Note that the table-per-relation
> approach is most appropriate when traversals are always strongly typed.
> E.g. when your step is v.out("knows") as opposed to v.out(). For v.out() to
> be supported efficiently, the monolithic table, or an element-to-type
> table, makes sense.
>
>
>
> > label specifies the relation type (we will always assume that a single
> > String is sufficient)
> >
>
> Exactly. And yeah, I think it is safe to assume that types can be
> identified by strings. Want namespaces? Use a qualified name syntax
> appropriate for your application.
>
>
>
> > identifiers... must contain sufficient information to identify the head
> of
> > the relation.
> >
>
> Yes.
>
>
> The return of the the method adjacents() is then the object(s) on the other
> > side of the relation(s).
> >
>
> Yeah. We're taking an element and then iterating through all of the
> incoming projections, of a given label, to that element. The label is the
> name of the relation together with the name of the field/column.
>
>
>
> > Now, given the way I have my data structure organized, I could beef up
> the
> > MyXXX implementation such that MyStrategy rewrites the base bytecode to:
> >
> > [...]
> > Now, I could really beef up MyStrategy when I realize that no path
> > information is used in the traversal. Thus, the base bytecode compiles
> to:
> >
> > [my:sql,SELECT name FROM properties_table,vertex_table,edge_table WHERE …
> > lots of join equalities]
> >
>
> Something of the kind.
>
>
>
> > [...]
> > To recap.
> >
> > 1. There are primitives.
> >
>
> +1
>
>
> > 2. There are Maps and Lists.
> >
>
> Sure. Lists of primitives, and maps of primitives to primitives.
>
>
>
> > 3. There are ComplexTypes.
> >
>
> I like the fancy term "algebraic data types". They are just tuples in which
> each field is either:
> 1) a primitive value (possibly tagged with a type), or
> 2) an element reference (possibly tagged with a type)
>
> You also need a special "unit" type for optionals.
>
>
>
> > 4. ComplexTypes are adjacent to other objects via relations.
> > - These adjacent objects may be cached locally with the
> > ComplexType instance.
> > - These adjacent objects may require some database
> lookup.
> > - Regardless, TP4 doesn’t care — its up to the provider’s
> > ComplexType instance to decide how to resolve the adjacency.
> >
>
> +1
>
>
> > 5. ComplexTypes don’t go over the wire — a ComplexTypeProxy with
> > appropriately provided toString() is all that leaves the TP4 VM.
> >
>
> As a tuple, ComplexTypes / ADTs go over the wire. The values of their
> primitive fields should probably go with them. However, the values of their
> element / entity fields are just references; the attached element doesn't
> go with them.
>
>
>
> > Finally, to solve the asMap()/asList() problem, we simply have:
> >
> > asMap(’name’,’age’) => complexType.adjacents(‘asMap’,’name’,’age')
> > asList() => complexType.adjacents(‘asList’)
> >
>
> I think I need an example of asList(), but I agree that we can make
> properties into key/value maps. If we want to access metaproperties, then
> we don't use asMap().
>
>
>
> It is up to the complexType to manifest a Map or List accordingly.
> >
> > I see this as basically a big flatmap system. ComplexTypes just map from
> > self to any number of logical neighbors as specified by the relation.
> >
> > Am I getting it?,
> >
>
> Yeah, and I think I am getting how you break down traversals into basic
> instructions. Go go GMachine.
>
> Josh
>
>
>
>
> > Marko.
> >
> > http://rredux.com <http://rredux.com/>
> >
> >
> >
> >
> > > On Apr 24, 2019, at 9:56 AM, Joshua Shinavier <j...@fortytwo.net>
> wrote:
> > >
> > > On Tue, Apr 23, 2019 at 10:28 AM Marko Rodriguez <okramma...@gmail.com
> >
> > > wrote:
> > >
> > >> Hi,
> > >>
> > >> I think we are very close to something useable for TP4 structure/.
> > Solving
> > >> this problem elegantly will open the flood gates on tp4/ development.
> > >>
> > >
> > > Yes, and formality often brings elegance. I don't think we can do much
> > > better than relational algebra and relational calculus in terms of
> > > formality, so to the extent we can reduce the fundamental TP4 traversal
> > > steps to basic relational operations, the floodgates will also be open
> to
> > > applications of query validation and query optimization from the last
> 40+
> > > years of research.
> > >
> > >
> > >
> > >> I still don’t grock your comeFrom().goto() stuff. I don’t get the
> > benefit
> > >> of having two instructions for “pointer chasing” instead of one.
> > >>
> > >
> > > There are just a handful of basic operations in relational algebra.
> > > Projection, selection, union, complement, Cartesian product. Joins, as
> > well
> > > as all other operations, can be derived from these. A lot of graph
> > > traversal can be accomplished using only projection and selection,
> which
> > is
> > > why we were able to get away with only to/goto and from/comeFrom in the
> > > examples above. However, I believe you do need both operations. You can
> > > kind of get away without from() if you assume that each vertex has
> local
> > > inE and outE references to incoming and outgoing edges, but I see that
> > as a
> > > kind of pre-materialized from()/select(). If you think of edges
> strictly
> > as
> > > relations, and represent them in a straightforward way with tables, you
> > > don't need the local inE and outE; whether you have them depends on the
> > > graph back-end.
> > >
> > >
> > >
> > >> Lets put that aside for now and lets turn to modeling a Vertex. Go
> back
> > to
> > >> my original representation:
> > >>
> > >> vertex.goto(‘label’)
> > >> vertex.goto(‘id’)
> > >>
> > >
> > > Local (in my view). All good.
> > >
> > >
> > >
> > >> vertex.goto(‘outE’)
> > >> vertex.goto(‘inE’)
> > >> vertex.goto(‘properties’)
> > >>
> > >
> > > Non-local (in my view). You can use goto(), but if the goal is to bring
> > the
> > > relational model into the fold, at a lower level you do have a select()
> > > operation. Unless you make projections local to vertices instead of
> > edges,
> > > but then you just have the same problem in reverse. Am I making sense?
> > >
> > >
> > > Any object can be converted into a Map. In TinkerPop3 we convert
> vertices
> > >> into maps via:
> > >>
> > >> g.V().has(‘name’,’marko’).valueMap() => {name:marko,age:29}
> > >> g.V().has(‘name’,’marko’).valueMap(true) =>
> > >> {id:1,label:person,name:marko,age:29}
> > >>
> > >
> > > Maps are A-OK. In the case of properties, I think where we differ is
> that
> > > you see a property like "name" as a key/value pair in a map local to
> the
> > > vertex. I see the property as an element of type "name", with the
> vertex
> > as
> > > a value in its local map, logically if not physically. This allows
> > maximum
> > > flexibility in terms of meta-properties -- exotic beasts which seem to
> be
> > > in a kind of limbo state in TP3, but if we're trying to be as general
> as
> > > possible, some data models we might want to pull in, like GRAKN.AI, do
> > > allow this kind of flexibility.
> > >
> > >
> > >
> > >> In the spirit of instruction reuse, we should have an asMap()
> > instruction
> > >> that works for ANY object. (As a side: this gets back to ONLY sending
> > >> primitives over the wire, no
> > >> Vertex/Edge/Document/Table/Row/XML/ColumnFamily/etc.). Thus, the above
> > is:
> > >>
> > >> g.V().has(‘name’,’marko’).properties().asMap() =>
> > >> {name:marko,age:29}
> > >> g.V().has(‘name’,’marko’).asMap() =>
> > >> {id:1,label:person,properties:{name:marko,age:29}}
> > >>
> > >
> > > Again, no argument here, although I would think of a map as an
> > > optimization. IMO, the fundamental projections from v[1] are id:1 and
> > > label:Person. You could make a map out of these, or just use an offset,
> > > since the keys are always the same. However, you can also build a map
> > > including any key you can turn into a function. properties() is such a
> > key.
> > >
> > >
> > > You might ask, why didn’t it go to outE and inE and map-ify that data?
> > >> Because those are "sibling” references, not “children” references.
> > >>
> > >> goto(‘outE’) is a “sibling” reference. (a vertex does not
> contain
> > >> an edge)
> > >> goto(‘id’) is a “child” reference. (a vertex contains the id)
> > >>
> > >
> > > I agree with both of those statements. A vertex does not contain the
> > edges
> > > incident on it. Again, I am thinking of properties a bit more like
> edges
> > > for maximum generality.
> > >
> > >
> > >
> > >> Where do we find sibling references?
> > >> Graphs: vertices don’t contain each other.
> > >> OO heaps: many objects don’t contain each other.
> > >> RDBMS: rows are linked by joins, but don’t contain each other.
> > >>
> > >
> > > Yep.
> > >
> > >
> > > So, the way in which we structure our references (pointers) determines
> > the
> > >> shape of the data and ultimately how different instructions will
> > behave. We
> > >> can’t assume that asMap() knows anything about
> > >> vertices/edges/documents/rows/tables/etc. It will simply walk all
> > >> child-references and create a map.
> > >>
> > >
> > > Just to play devil's advocate, you *could* include "inE" and "outE" as
> > keys
> > > in the local map of a vertex; it's just a matter of what you choose to
> > do.
> > > inE and outE are perfectly good functions from a vertex to a set of
> > edges.
> > >
> > >
> > > We don’t want TP to get involved in “complex data types.”
> > >
> > >
> > > Well, how do you feel about algebraic data types? They are simple, and
> > > allow you to capture arbitrary relations as elements.
> > >
> > >
> > >
> > >> We don’t care. You can propagate MyDatabaseObject through the TP4 VM
> > >> pipeline and load your object up with methods for optimizations with
> > your
> > >> DB and all that, but for TP4, your object is just needs to implement:
> > >>
> > >> ComplexType
> > >> - Iterator<T> children(String label)
> > >> - Iterator<T> siblings(String label)
> > >> - default Iterator<T> references(String label) {
> > >> IteratorUtils.concat(children(label), siblings(label)) }
> > >> - String toString()
> > >>
> > >
> > > I don't think you need siblings(). I think you need a more generic
> > > select(), but since this is graph traversal, select() only needs the
> > > identifier of a type (e.g. "knows") and the name of a field (e.g.
> "out").
> > >
> > >
> > >
> > >> When a ComplexType goes over the wire to the user, it just represented
> > as
> > >> a ComplexTypeProxy with a toString() like v[1],
> > >> tinkergraph[vertices:10,edges:34], etc. All references are
> disconnected.
> > >> Yes, even children references. We do not want language drivers having
> to
> > >> know about random object types and have to deal with implementing
> > >> serializers and all that non-sense. The TP4 serialization protocol is
> > >> primitives, maps, lists, bytecode, and traversers. Thats it!
> > >>
> > >
> > > No disagreement here. I think the only disconnect is about what keys
> are
> > > local to what elements. Some keys are hard-local, like id and type for
> > all
> > > elements, and "in" and "out" for edges and properties. These *should*
> be
> > > carried over the wire. Properties, incident edges, etc. possibly but
> not
> > > necessarily.
> > >
> > >
> > >
> > >> *** Only Maps and Lists (that don’t contain complex data types)
> maintain
> > >> their child references “over the wire.”
> > >>
> > >
> > > Sure.
> > >
> > >
> > >
> > >> I don’t get your hypergraph example, so let me try another example:
> > >>
> > >> tp ==member==> marko, josh
> > >>
> > >> TP is a vertex and there is a directed hyperedge with label “member”
> > >> connecting to marko and josh vertices.
> > >>
> > >
> > > That's kind of an unlabeled hyperedge; I am not sure we need to support
> > > those. Look at the GRAKN data model, or at HypergraphDB or earlier
> > > hypergraph data models. A hyperedge is essentially a tuple in which
> each
> > > components has a label ("role", in GRAKN). In other words, it is a
> > relation
> > > in which some of the columns may be foreign keys. In your example,
> rather
> > > than "member" connecting "tp" to a set of vertices, you might have
> > > something like Collaborated{person1:marko, person2:josh, project=tp}.
> > Then
> > > a query like "who did marko collaborate with on tp?" becomes:
> > >
> > > tp.from("Collaborated", "project").restrict("person1",
> > > "marko").to("person2")
> > >
> > > Of course, if you want this relationship to be symmetrical, you can
> > > introduce a constraint.
> > >
> > >
> > >
> > >> tp.goto(“outE”).filter(goto(“label”).is(“member”)).goto(“inV”)
> > >>
> > >> Looks exactly like a property graph query? However, its not because
> > >> goto(“inV”) returns 2 vertices, not 1.
> > >
> > >
> > > I think your example works well for the type of hypergraph you are
> > > referring to. It's just different than the type of hypergraph I am
> > > referring to. I think by now you know that I would rather see a from()
> > > instead of that goto("outE"). I also agree you can make a function out
> of
> > > outE, and expose it using a map, if you really want to. Under the hood,
> > > however, I see this as traversing a projection head to tail rather than
> > > tail to head.
> > >
> > >
> > >
> > >> EdgeVertexFlatmapFunction works for property graphs and hypergraphs.
> It
> > >> doesn’t care — it just follows goto() pointers! That is, it follows
> the
> > >> ComplexType.references(“inV”). Multi-properties are the same as well.
> > >> Likewise for meta-properties. These data model variations are not
> > “special”
> > >> to the TP4 VM. It just walks references whether there are 0,1,2, or N
> of
> > >> them.
> > >>
> > >
> > > At a high level, I agree with what you are saying. We should have a
> > common
> > > data model that unifies traditional property graphs, hypergraphs,
> > > relational databases, and any other data model that can be modeled
> using
> > > algebraic data types with references. We define a small set of basic
> > > operations on this data model which can be combined into more complex
> > > operations that are amenable to static analysis and optimization. We
> can
> > > send graph data over the wire as collections of elements using the bare
> > > minimum of local fields, and reconstruct the graph on the other end. We
> > can
> > > operate on streams of such elements under suitable conditions (elements
> > > sent in an appropriate order). The basic operations are not tied to the
> > > JVM, and should be straightforward to implement in other frameworks.
> > >
> > >
> > >
> > >>
> > >> Thus, what is crucial to all this is the “shape of the data.” Using
> your
> > >> pointers wisely so instructions produce useful results.
> > >>
> > >
> > > +1
> > >
> > >
> > >
> > >> Does any of what I wrote update your comeFrom().goto() stuff?
> > >
> > >
> > > Sadly, no, though I appreciate that you are coming from a slightly
> > > different place w.r.t. properties, hypergraphs, and most importantly,
> the
> > > role of a type system.
> > >
> > >
> > >
> > >> If not, can you please explain to me why comeFrom() is cool — sorry
> for
> > >> being dense (aka “being Kuppitz" — thats right, I said it. boom!).
> > >>
> > >
> > > Let's keep iterating until we reach a fixed point. Maybe Daniel's
> already
> > > there.
> > >
> > > Josh
> > >
> > >
> > >
> > >>
> > >> Thanks,
> > >> Marko.
> > >>
> > >> http://rredux.com <http://rredux.com/>
> > >>
> > >>
> > >>
> > >>
> > >>> On Apr 23, 2019, at 10:25 AM, Joshua Shinavier <j...@fortytwo.net>
> > >> wrote:
> > >>>
> > >>> On Tue, Apr 23, 2019 at 5:14 AM Marko Rodriguez <
> okramma...@gmail.com>
> > >>> wrote:
> > >>>
> > >>>> Hey Josh,
> > >>>>
> > >>>> This gets to the notion I presented in “The Fabled GMachine.”
> > >>>> http://rredux.com/the-fabled-gmachine.html <
> > >>>> http://rredux.com/the-fabled-gmachine.html> (first paragraph of
> > >>>> “Structures, Processes, and Languages” section)
> > >>>>
> > >>>> All that exists are memory addresses that contain either:
> > >>>>
> > >>>> 1. A primitive
> > >>>> 2. A set of labeled references to other references or
> > primitives.
> > >>>>
> > >>>> Using your work and the above, here is a super low-level ‘bytecode'
> > for
> > >>>> property graphs.
> > >>>>
> > >>>> v.goto("id") => 1
> > >>>>
> > >>>
> > >>> LGTM. An id is special because it is uniquely identifying / is a
> > primary
> > >>> key for the element. However, it is also just a field of the element,
> > >> like
> > >>> "in"/"inV" and "out"/"outV" are fields of an edge. As an aside, an id
> > >> would
> > >>> only really need to be unique among other elements of the same type.
> To
> > >> the
> > >>> above, I would add:
> > >>>
> > >>> v.type() => Person
> > >>>
> > >>> ...a special operation which takes you from an element to its type.
> > This
> > >> is
> > >>> important if unions are supported; e.g. "name" in my example can
> apply
> > >>> either to a Person or a Project.
> > >>>
> > >>>
> > >>> v.goto("label") => person
> > >>>>
> > >>>
> > >>> Or that. Like "id", "type"/"label" is special. You can think of it
> as a
> > >>> field; it's just a different sort of field which will have the same
> > value
> > >>> for all elements of any given type.
> > >>>
> > >>>
> > >>>
> > >>>> v.goto("properties").goto("name") => "marko"
> > >>>>
> > >>>
> > >>> OK, properties. Are properties built-in as a separate kind of thing
> > from
> > >>> edges, or can we treat them the same as vertices and edges here? I
> > think
> > >> we
> > >>> can treat them the same. A property, in the algebraic model I
> described
> > >>> above, is just an element with two fields, the second of which is a
> > >>> primitive value. As I said, I think we need two distinct traversal
> > >>> operations -- projection and selection -- and here is where we can
> use
> > >> the
> > >>> latter. Here, I will call it "comeFrom".
> > >>>
> > >>> v.comeFrom("name", "out").goto("in") => {"marko"}
> > >>>
> > >>> You can think of this comeFrom as a special case of a select()
> function
> > >>> which takes a type -- "name" -- and a set of key/value pairs {("out",
> > >> v)}.
> > >>> It returns all matching elements of the given type. You then project
> to
> > >> the
> > >>> "in" value using your goto. I wrote {"marko"} as a set, because
> > comeFrom
> > >>> can give you multiple properties, depending on whether
> multi-properties
> > >> are
> > >>> supported.
> > >>>
> > >>> Note how similar this is to an edge traversal:
> > >>>
> > >>> v.comeFrom("knows", "out").goto("in") => {v[2], v[4]}
> > >>>
> > >>> Of course, you could define "properties" in such a way that a
> > >>> goto("properties") does exactly this under the hood, but in terms of
> > low
> > >>> level instructions, you need something like comeFrom.
> > >>>
> > >>>
> > >>> v.goto("properties").goto("name").goto(0) => "m"
> > >>>>
> > >>>
> > >>> This is where the notion of optionals becomes handy. You can make
> > >>> array/list indices into fields like this, but IMO you should also
> make
> > >> them
> > >>> safe. E.g. borrowing Haskell syntax for a moment:
> > >>>
> > >>> v.goto("properties").goto("name").goto(0) => Just 'm'
> > >>>
> > >>> v.goto("properties").goto("name").goto(5) => Nothing
> > >>>
> > >>>
> > >>> v.goto("outE").goto("inV") => v[2], v[4]
> > >>>>
> > >>>
> > >>> I am not a big fan of untyped "outE", but you can think of this as a
> > >> union
> > >>> of all v.comeFrom(x, "out").goto("in"), where x is any edge type.
> Only
> > >>> "knows" and "created" are edge types which are applicable to
> "Person",
> > so
> > >>> you will only get {v[2], v[4]}. If you want to get really crazy, you
> > can
> > >>> allow x to be any type. Then you get {v[2], v[4], 29, "marko"}.
> > >>>
> > >>>
> > >>>
> > >>>> g.goto("V").goto(1) => v[1]
> > >>>>
> > >>>
> > >>> That, or you give every element a virtual field called "graph". So:
> > >>>
> > >>> v.goto("graph") => g
> > >>>
> > >>> g.comeFrom("Person", "graph") => {v[1], v[2], v[4], v[6]}
> > >>>
> > >>> g.comeFrom("Person", "graph").restrict("id", 1)
> > >>>
> > >>> ...where restrict() is the relational "sigma" operation as above, not
> > to
> > >> be
> > >>> confused with TinkerPop's select(), filter(), or has() steps. Again,
> I
> > >>> prefer to specify a type in comeFrom (i.e. we're looking specifically
> > >> for a
> > >>> Person with id of 1), but you could also do a comprehension
> > g.comeFrom(x,
> > >>> "graph"), letting x range over all types.
> > >>>
> > >>>
> > >>>
> > >>>> The goto() instruction moves the “memory reference” (traverser) from
> > the
> > >>>> current “memory address” to the “memory address” referenced by the
> > >> goto()
> > >>>> argument.
> > >>>>
> > >>>
> > >>> Agreed, if we also think of primitive values as memory references.
> > >>>
> > >>>
> > >>>
> > >>>> The Gremlin expression:
> > >>>>
> > >>>> g.V().has(‘name’,’marko’).out(‘knows’).drop()
> > >>>>
> > >>>> ..would compile to:
> > >>>>
> > >>>>
> > >>>>
> > >>
> >
> g.goto(“V”).filter(goto(“properties”).goto(“name”).is(“marko”)).goto(“outE”).filter(goto(“label”).is(“knows”)).goto(“inV”).free()
> > >>>>
> > >>>
> > >>>
> > >>> In the alternate universe:
> > >>>
> > >>> g.comeFrom("Person", "graph").comeFrom("name", "out").restrict("in",
> > >>> "marko").goto("out").comeFrom("knows", "out").goto("in").free()
> > >>>
> > >>> I have wimped out on free() and just left it as you had it, but I
> think
> > >> it
> > >>> would be worthwhile to explore a monadic syntax for traversals with
> > >>> side-effects. Different topic.
> > >>>
> > >>> Now, all of this "out", "in" business is getting pretty repetitive,
> > >> right?
> > >>> Well, the field names become more diverse if we allow hyper-edges and
> > >>> generalized ADTs. E.g. in my Trip example, say I want to know all
> > >> drop-off
> > >>> locations for a given rider:
> > >>>
> > >>> u.comeFrom("Trip", "rider").goto("dropoff").goto("place")
> > >>>
> > >>> Done.
> > >>>
> > >>>
> > >>>
> > >>>> If we can get things that “low-level” and still efficient to
> compile,
> > >> then
> > >>>> we can model every data structure. All you are doing is pointer
> > chasing
> > >>>> through a withStructure() data structure. .
> > >>>>
> > >>>
> > >>> Agreed.
> > >>>
> > >>>
> > >>> No one would ever want to write strategies for goto()-based Bytecode.
> > >>>
> > >>>
> > >>> Also agreed.
> > >>>
> > >>>
> > >>>
> > >>>> Thus, perhaps there could be a PropertyGraphDecorationStrategy that
> > >> does:
> > >>>>
> > >>>> [...]
> > >>>
> > >>>
> > >>> No argument here, though the alternate-universe "bytecode" would look
> > >>> slightly different. And the high-level syntax should also be able to
> > deal
> > >>> with generalized relations / data types gracefully. As a thought
> > >>> experiment, suppose we were to define the steps to() as your goto(),
> > and
> > >>> from() as my comeFrom(). Then traversals like:
> > >>>
> > >>> u.from("Trip", "rider").to("dropoff").to("time")
> > >>>
> > >>> ...look pretty good as-is, and are not too low-level. However,
> ordinary
> > >>> edge traversals like:
> > >>>
> > >>> v.from("knows", "out").to("in")
> > >>>
> > >>> ...do look a little Assembly-like. So in/out/both etc. remain as they
> > >> are,
> > >>> but are shorthand for from() and to() steps using "out" or "in":
> > >>>
> > >>> v.out("knows") === v.outE("knows").inV() === v.from("knows",
> > >> "out").to("in")
> > >>>
> > >>>
> > >>> [I AM NOW GOING OFF THE RAILS]
> > >>>> [sniiiiip]
> > >>>>
> > >>>
> > >>> Sure. Again, I like the idea of wrapping side-effects in monads. What
> > >> would
> > >>> that look like in a Gremlinesque fluent syntax? I don't quite know,
> but
> > >> if
> > >>> we think of the dot as a monadic bind operation like Haskell's >>=,
> > then
> > >>> perhaps the monadic expressions look pretty similar to what you have
> > just
> > >>> sketched out. Might have to be careful about what it means to nest
> > >>> operations as in your addEdge examples.
> > >>>
> > >>>
> > >>>
> > >>> [I AM NOW BACK ON THE RAILS]
> > >>>>
> > >>>> Its as if “properties”, “outE”, “label”, “inV”, etc. references mean
> > >>>> something to property graph providers and they can do more
> intelligent
> > >>>> stuff than what MongoDB would do with such information. However,
> > >> someone,
> > >>>> of course, can create a MongoDBPropertyGraphStrategy that would make
> > >>>> documents look like vertices and edges and then use O(log(n))
> lookups
> > on
> > >>>> ids to walk the graph. However, if that didn’t exist, it would still
> > do
> > >>>> something that works even if its horribly inefficient as every
> > database
> > >> can
> > >>>> make primitives with references between them!
> > >>>>
> > >>>
> > >>> I'm on the same same pair of rails.
> > >>>
> > >>>
> > >>>
> > >>>> Anywho @Josh, I believe goto() is what you are doing with
> > >> multi-references
> > >>>> off an object. How do we make it all clean, easy, and universal?
> > >>>>
> > >>>
> > >>> Let me know what you think of the above.
> > >>>
> > >>> Josh
> > >>>
> > >>>
> > >>>
> > >>>>
> > >>>> Marko.
> > >>>>
> > >>>> http://rredux.com <http://rredux.com/>
> > >>>>
> > >>>>
> > >>>>
> > >>>>
> > >>>>> On Apr 22, 2019, at 6:42 PM, Joshua Shinavier <j...@fortytwo.net>
> > >> wrote:
> > >>>>>
> > >>>>> Ah, glad you asked. It's all in the pictures. I have nowhere to put
> > >> them
> > >>>> online at the moment... maybe this attachment will go through to the
> > >> list?
> > >>>>>
> > >>>>> Btw. David Spivak gave his talk today at Uber; it was great. Juan
> > >>>> Sequeda (relational <--> RDF mapping guy) was also here, and Ryan
> > joined
> > >>>> remotely. Really interesting discussion about databases vs. graphs,
> > and
> > >>>> what category theory brings to the table.
> > >>>>>
> > >>>>>
> > >>>>> On Mon, Apr 22, 2019 at 1:45 PM Marko Rodriguez <
> > okramma...@gmail.com
> > >>>> <mailto:okramma...@gmail.com>> wrote:
> > >>>>> Hey Josh,
> > >>>>>
> > >>>>> I’m digging what you are saying, but the pictures didn’t come
> through
> > >>>> for me ? … Can you provide them again (or if dev@ is filtering
> them,
> > >> can
> > >>>> you give me URLs to them)?
> > >>>>>
> > >>>>> Thanks,
> > >>>>> Marko.
> > >>>>>
> > >>>>>
> > >>>>>> On Apr 21, 2019, at 12:58 PM, Joshua Shinavier <j...@fortytwo.net
> > >>>> <mailto:j...@fortytwo.net>> wrote:
> > >>>>>>
> > >>>>>> On the subject of "reified joins", maybe be a picture will be
> worth
> > a
> > >>>> few words. As I said in the thread <
> > >>>>
> > https://groups.google.com/d/msg/gremlin-users/_s_DuKW90gc/Xhp5HMfjAQAJ
> > >> <
> > >>>>
> > https://groups.google.com/d/msg/gremlin-users/_s_DuKW90gc/Xhp5HMfjAQAJ
> > >>>>
> > >>>> on property graph standardization, if you think of vertex labels,
> edge
> > >>>> labels, and property keys as types, each with projections to two
> other
> > >>>> types, there is a nice analogy with relations of two columns, and
> this
> > >>>> analogy can be easily extended to hyper-edges. Here is what the
> schema
> > >> of
> > >>>> the TinkerPop classic graph looks like if you make each type (e.g.
> > >> Person,
> > >>>> Project, knows, name) into a relation:
> > >>>>>>
> > >>>>>>
> > >>>>>>
> > >>>>>> I have made the vertex types salmon-colored, the edge types
> yellow,
> > >>>> the property types green, and the data types blue. The "o" and "I"
> > >> columns
> > >>>> represent the out-type (e.g. out-vertex type of Person) and in-type
> > >> (e.g.
> > >>>> property value type of String) of each relation. More than two
> arrows
> > >> from
> > >>>> a column represent a coproduct, e.g. the out-type of "name" is
> Person
> > OR
> > >>>> Project. Now you can think of out() and in() as joins of two tables
> > on a
> > >>>> primary and foreign key.
> > >>>>>>
> > >>>>>> We are not limited to "out" and "in", however. Here is the ternary
> > >>>> relationship (hyper-edge) from hyper-edge slide <
> > >>>>
> > >>
> >
> https://www.slideshare.net/joshsh/a-graph-is-a-graph-is-a-graph-equivalence-transformation-and-composition-of-graph-data-models-129403012/49
> > >>>> <
> > >>>>
> > >>
> >
> https://www.slideshare.net/joshsh/a-graph-is-a-graph-is-a-graph-equivalence-transformation-and-composition-of-graph-data-models-129403012/49
> > >>>>
> > >>>> of my Graph Day preso, which has three columns/roles/projections:
> > >>>>>>
> > >>>>>>
> > >>>>>>
> > >>>>>> I have drawn Says in light blue to indicate that it is a
> generalized
> > >>>> element; it has projections other than "out" and "in". Now the line
> > >> between
> > >>>> relations and edges begins to blur. E.g. in the following, is
> > >> PlaceEvent a
> > >>>> vertex or a property?
> > >>>>>>
> > >>>>>>
> > >>>>>>
> > >>>>>> With the right type system, we can just speak of graph elements,
> and
> > >>>> use "vertex", "edge", "property" when it is convenient. In the
> > >> relational
> > >>>> model, they are relations. If you materialize them in a relational
> > >>>> database, they are rows. In any case, you need two basic graph
> > traversal
> > >>>> operations:
> > >>>>>> project() -- forward traversal of the arrows in the above
> diagrams.
> > >>>> Takes you from an element to a component like in-vertex.
> > >>>>>> select() -- reverse traversal of the arrows. Allows you to answer
> > >>>> questions like "in which Trips is John Doe the rider?"
> > >>>>>>
> > >>>>>> Josh
> > >>>>>>
> > >>>>>>
> > >>>>>> On Fri, Apr 19, 2019 at 10:03 AM Marko Rodriguez <
> > >> okramma...@gmail.com
> > >>>> <mailto:okramma...@gmail.com> <mailto:okramma...@gmail.com <mailto:
> > >>>> okramma...@gmail.com>>> wrote:
> > >>>>>> Hello,
> > >>>>>>
> > >>>>>> I agree with everything you say. Here is my question:
> > >>>>>>
> > >>>>>> Relational database — join: Table x Table x equality
> function
> > >>>> -> Table
> > >>>>>> Graph database — traverser: Vertex x edge label -> Vertex
> > >>>>>>
> > >>>>>> I want a single function that does both. The only think was to
> > >>>> represent traverser() in terms of join():
> > >>>>>>
> > >>>>>> Graph database — traverser: Vertices x Vertex x equality
> > >>>> function -> Vertices
> > >>>>>>
> > >>>>>> For example,
> > >>>>>>
> > >>>>>> V().out(‘address’)
> > >>>>>>
> > >>>>>> ==>
> > >>>>>>
> > >>>>>> g.join(V().hasLabel(‘person’).as(‘a’)
> > >>>>>> V().hasLabel(‘addresses’).as(‘b’)).
> > >>>>>> by(‘name’).select(?address vertex?)
> > >>>>>>
> > >>>>>> That is, join the vertices with themselves based on some predicate
> > to
> > >>>> go from vertices to vertices.
> > >>>>>>
> > >>>>>> However, I would like instead to transform the relational database
> > >>>> join() concept into a traverser() concept. Kuppitz and I were
> talking
> > >> the
> > >>>> other day about a link() type operator that says: “try and link to
> > this
> > >>>> thing in some specified way.” .. ?? The problem we ran into is
> again,
> > >> “link
> > >>>> it to what?”
> > >>>>>>
> > >>>>>> - in graph, the ‘to what’ is hardcoded so you don’t need to
> > >>>> specify anything.
> > >>>>>> - in rdbms, the ’to what’ is some other specified table.
> > >>>>>>
> > >>>>>> So what does the link() operator look like?
> > >>>>>>
> > >>>>>> ——
> > >>>>>>
> > >>>>>> Some other random thoughts….
> > >>>>>>
> > >>>>>> Relational databases join on the table (the whole collection)
> > >>>>>> Graph databases traverser on the vertex (an element of the whole
> > >>>> collection)
> > >>>>>>
> > >>>>>> We can make a relational database join on single row (by
> providing a
> > >>>> filter to a particular primary key). This is the same as a table
> with
> > >> one
> > >>>> row. Likewise, for graph in the join() context above:
> > >>>>>>
> > >>>>>> V(1).out(‘address’)
> > >>>>>>
> > >>>>>> ==>
> > >>>>>>
> > >>>>>> g.join(V(1).as(‘a’)
> > >>>>>> V().hasLabel(‘addresses’).as(‘b’)).
> > >>>>>> by(‘name’).select(?address vertex?)
> > >>>>>>
> > >>>>>> More thoughts please….
> > >>>>>>
> > >>>>>> Marko.
> > >>>>>>
> > >>>>>> http://rredux.com <http://rredux.com/> <http://rredux.com/ <
> > >>>> http://rredux.com/>> <http://rredux.com/ <http://rredux.com/> <
> > >>>> http://rredux.com/ <http://rredux.com/>>>
> > >>>>>>
> > >>>>>>
> > >>>>>>
> > >>>>>>
> > >>>>>>> On Apr 19, 2019, at 4:20 AM, pieter martin <
> > pieter.mar...@gmail.com
> > >>>> <mailto:pieter.mar...@gmail.com> <mailto:pieter.mar...@gmail.com
> > >> <mailto:
> > >>>> pieter.mar...@gmail.com>>> wrote:
> > >>>>>>>
> > >>>>>>> Hi,
> > >>>>>>> The way I saw it is that the big difference is that graph's have
> > >>>>>>> reified joins. This is both a blessing and a curse.
> > >>>>>>> A blessing because its much easier (less text to type, less
> > mistakes,
> > >>>>>>> clearer semantics...) to traverse an edge than to construct a
> > manual
> > >>>>>>> join.A curse because there are almost always far more ways to
> > >>>> traverse
> > >>>>>>> a data set than just by the edges some architect might have
> > >>>> considered
> > >>>>>>> when creating the data set. Often the architect is not the domain
> > >>>>>>> expert and the edges are a hardcoded layout of the dataset, which
> > >>>>>>> almost certainly won't survive the real world's demands. In
> graphs,
> > >>>> if
> > >>>>>>> their are no edges then the data is not reachable, except via
> > indexed
> > >>>>>>> lookups. This is the standard engineering problem of database
> > design,
> > >>>>>>> but it is important and useful that data can be traversed,
> joined,
> > >>>>>>> without having reified edges.
> > >>>>>>> In Sqlg at least, but I suspect it generalizes, I want to create
> > the
> > >>>>>>> notion of a "virtual edge". Which in meta data describes the join
> > and
> > >>>>>>> then the standard to(direction, "virtualEdgeName") will work.
> > >>>>>>> In a way this is precisely to keep the graphy nature of gremlin,
> > i.e.
> > >>>>>>> traversing edges, and avoid using the manual join syntax you
> > >>>> described.
> > >>>>>>> CheersPieter
> > >>>>>>>
> > >>>>>>> On Thu, 2019-04-18 at 14:15 -0600, Marko Rodriguez wrote:
> > >>>>>>>> Hi,
> > >>>>>>>> *** This is mainly for Kuppitz, but if others care.
> > >>>>>>>> Was thinking last night about relational data and Gremlin. The
> T()
> > >>>>>>>> step returns all the tables in the withStructure() RDBMS
> database.
> > >>>>>>>> Tables are ‘complex values’ so they can't leave the VM (only a
> > >>>> simple
> > >>>>>>>> ‘toString’).
> > >>>>>>>> Below is a fake Gremlin session. (and these are just ideas…)
> > tables
> > >>>>>>>> -> a ListLike of rows rows -> a MapLike of primitives
> > >>>>>>>> gremlin> g.T()==>t[people]==>t[addresses]gremlin>
> > >>>>>>>> g.T(‘people’)==>t[people]gremlin>
> > >>>>>>>>
> > >>>>
> g.T(‘people’).values()==>r[people:1]==>r[people:2]==>r[people:3]greml
> > >>>>>>>> in>
> > >>>>>>>>
> > >>>>
> g.T(‘people’).values().asMap()==>{name:marko,age:29}==>{name:kuppitz,
> > >>>>>>>> age:10}==>{name:josh,age:35}gremlin>
> > >>>>>>>>
> > >>>>
> g.T(‘people’).values().has(‘age’,gt(20))==>r[people:1]==>r[people:3]g
> > >>>>>>>> remlin>
> > >>>>>>>>
> > >>>>
> g.T(‘people’).values().has(‘age’,gt(20)).values(‘name’)==>marko==>jos
> > >>>>>>>> h
> > >>>>>>>> Makes sense. Nice that values() and has() generally apply to all
> > >>>>>>>> ListLike and MapLike structures. Also, note how asMap() is the
> > >>>>>>>> valueMap() of TP4, but generalizes to anything that is MapLike
> so
> > it
> > >>>>>>>> can be turned into a primitive form as a data-rich result from
> the
> > >>>>>>>> VM.
> > >>>>>>>> gremlin> g.T()==>t[people]==>t[addresses]gremlin>
> > >>>>>>>>
> > >>>>
> g.T(‘addresses’).values().asMap()==>{name:marko,city:santafe}==>{name
> > >>>>>>>> :kuppitz,city:tucson}==>{name:josh,city:desertisland}gremlin>
> > >>>>>>>> g.join(T(‘people’).as(‘a’),T(‘addresses’).as(‘b’)).
> > >>>> by(se
> > >>>>>>>> lect(‘a’).value(’name’).is(eq(select(‘b’).value(’name’))).
> > >>>>
> > >>>>>>>> values().asMap()==>{a.name:marko,a.age:29,b.name:
> > >>>> marko,b.city:santafe
> > >>>>>>>> }==>{a.name:kuppitz,a.age:10,b.name:kuppitz,b.city:tucson}==>{
> > >>>> a.name <http://a.name/> <http://a.name/ <http://a.name/>>:
> > >>>>>>>> josh,a.age:35,b.name:josh,b.city:desertisland}gremlin>
> > >>>>>>>> g.join(T(‘people’).as(‘a’),T(‘addresses’).as(‘b’)).
> > >>>> by(’n
> > >>>>>>>> ame’). // shorthand for equijoin on name
> > >>>>>>>> column/key values().asMap()==>{a.name:marko,a.age:29,
> > >>>> b.name <http://b.name/> <http://b.name/ <http://b.name/>>
> > >>>>>>>> :marko,b.city:santafe}==>{a.name:kuppitz,a.age:10,b.name:
> kuppitz,
> > >>>> b.ci <http://b.ci/> <http://b.ci/ <http://b.ci/>>
> > >>>>>>>> ty:tucson}==>{a.name:josh,a.age:35,b.name:
> > >>>> josh,b.city:desertisland}gr
> > >>>>>>>> emlin>
> > >>>>>>>> g.join(T(‘people’).as(‘a’),T(‘addresses’).as(‘b’)).
> > >>>> by(’n
> > >>>>>>>> ame’)==>t[people<-name->addresses] // without asMap(), just the
> > >>>>>>>> complex value ‘toString'gremlin>
> > >>>>>>>> And of course, all of this is strategized into a SQL call so its
> > >>>>>>>> joins aren’t necessarily computed using TP4-VM resources.
> > >>>>>>>> Anywho — what I hope to realize is the relationship between
> > “links”
> > >>>>>>>> (graph) and “joins” (tables). How can we make (bytecode-wise at
> > >>>>>>>> least) RDBMS join operations and graph traversal operations ‘the
> > >>>>>>>> same.’?
> > >>>>>>>> Singleton: Integer, String, Float, Double, etc. Collection:
> > >>>>>>>> List, Map (Vertex, Table, Document) Linkable: Vertex, Table
> > >>>>>>>> Vertices and Tables can be “linked.” Unlike Collections, they
> > don’t
> > >>>>>>>> maintain a “parent/child” relationship with the objects they
> > >>>>>>>> reference. What does this mean……….?
> > >>>>>>>> Take care,Marko.
> > >>>>>>>> http://rredux.com <http://rredux.com/> <http://rredux.com/ <
> > >>>> http://rredux.com/>> <http://rredux.com/ <http://rredux.com/> <
> > >>>> http://rredux.com/ <http://rredux.com/>>> <http://rredux.com/ <
> > >>>> http://rredux.com/> <http://rredux.com/ <http://rredux.com/>> <
> > >>>> http://rredux.com/ <http://rredux.com/> <http://rredux.com/ <
> > >>>> http://rredux.com/>>>>
> > >>>>>
> > >>>>> <diagrams.zip>
> > >>>>
> > >>>>
> > >>
> > >>
> >
> >
>
