LMDVVYD, a Union-Schema contracting database Part 2

Posted on November 11, 2012

As detailed in the last post, I plan to define roughly how Events work and the Logical layout of the server. I plan to defer How commands are constructed and Return Formats for now.

Service Layout: Server

I plan for both the server and client to be libraries, and the server as a stand-alone executable will really be a wrapper that links to it and broadcasts ipc or localhost or *.

The client will have the same interface as the server for the following kinds of actions

  • Get all entities in a property set
  • Get Raw Entity
  • Create Entity
  • Destroy Entity
  • Create Property
  • Define Property
  • Add Property Set to Entity
  • Remove Property set from Entity, optional destruction of entity if no properties remain
  • Some form of serialization for saving or replicating

I want these actions to be present on both the client and the server, because I want the possibility of the client having two paths of communication, one for when the server is local, and embedded, and the other for when the server may be local, or it may be anywhere. This requirement fits well with the proxy pattern.

For the first mode, the client calls the server methods directly. However when the server is remote, the client uses ZMQ to communicate with the server, which has a worker that will do the processing. The workers following the thread pool pattern, otherwise called a worker crew model.

Regarding workers, there is only a finite amount available at a given time, it may be dynamic, but it needs to be configurable. This is one of the parameters that should be set when the server serves over the network, whether or not it is embedded, since remote clients may want access. Essentially in an embedded environment though, a client method caller is their own worker.

Events

As the actions / methods listed above are executed, events will be emitted, such as

  • Entities Changed
  • Entity Joined Property Set
  • Entity Left Property Set

There is no need for destruction, since it is not relevant to listeners, and it can be contained in the Left Property Set event.

The events will go over ZMQ, which the client context will provide when polled for events.

I’ll introduce the client context in a moment

There would also be some utility functions available, since both the server and the client should know the contracts, and for safety, I think there should be a Conform Entity to Property Set function, which should be used when an entity joins or leaves a property set.

Service Layout: Client

Within whatever programs that interact with the server, the client library offers what the server offers, plus a few things like subscribing to specific property sets for events, polling these events.

The client library offers what is called a client context, which has it’s own subscribing socket(s), and will automatically filter to avoid feedback. The context is alive as long as the component that uses it is alive. Having a context is essentially RAII. It is up to the programmer to make sure that the components get their event messages pumped, otherwise stuff will likely stay and collect in memory.

Ideally, I want the components to rarely depend on the shared server-client functions, only to be used in special circumstances like creating new entities, or destroying. The getting should only be used on initialization. Additionally, the Get All Entities in Property Set function should have a serialization debug function, but also one that returns effectively nothing, but will cause events to be emitted, however directed at only this context, which will be handled in the form of Entity joined property set, which will be processed at a later time.

This makes for reusable event driven programming, where initialization and awareness are handled in one place.

Multiple Client resolution

I was considering for a while the problem of avoiding conflicts with different actors, otherwise known as completely unique and separate clients.

One resolution was to have a local database for things, and that we have an indirection so that something in the middle resolves what maps from local to shared, and shared to local. This mapping can be costly since I anticipate it happening all the time.

Another thought was to have everything on the database’s side, and the database automatically updates their session entities from the shared entities if the master updates them, but then this loses the entire purpose of having the communication be local and fast.

Then I thought… Well, if we do follow the event driven pattern, then I can essentially fake having a local database entirely if I throw one more event into the design, Request Entity Knowledge For Property Sets.

This kind of function would essentially say to every component to emit its version of the local copy, and then eventually the client context, as the information comes back in the form of Entity Joined Property Set, the message will have a whitelist, containing whoever requested the entity knowledge. The client context will handle the merging once it has a group of the same entity ID which fit within the property set. All entities which have part of the entity set, but is a non-empty subset of the specified set, which seem to never collect to have the full set will likely have some sort of expiration date for which the information is discarded.

Note: The event that an Entity Joined a Property Set has either a whitelist for initializations, or a blacklist to avoid feedback. Also note, that the blacklist must be per client instance, not global, since other clients with the same component need to be aware of the change! The entry of the blacklist or whitelist will be described at the end.

Although this kind of feature is important to having a consistent knowledge for when components are dynamically initialized, there is one huge benefit.

When the master dies, in other words, the game host dies, it is now possible to have an automated reconstruction for a new game host from such fragmented data! And, it might be possible to even run the game in a crippled mode like this until the master comes back, or a new master is ready.

Blacklists and Whitelists

When it comes to some events, we need a way to feed information to all but one, or only one. The way we can identify that is through an ID of sorts. We can’t depend on a sequence ID, since this may cause conflict as components may be initalized nondeterministically. To resolve this, each context should have it’s own UUID, which is a mostly random (except and embedded version number), 128-bit identifier. According to wikipedia..

Only after generating 1 billion UUIDs every second for the next 100 years, the probability of creating just one duplicate would be about 50%.

Since There are 340,282,366,920,938,463,463,374,607,431,768,211,456 possible UUIDs, and the chance of collision is low, this is a rather dependable means to have distributed instances. In fact, this is what many databases, like mongoDB and CouchDB use to have unique entries created on multiple servers and still be eventually consistent.

Initial Prototype

I will probably use YAML as a configuration for a stand alone server, though I might also try out the things POCO provides for configuration.

The service on initialization will take a parameter specifying which engine to use, since there might be many eventually, such as a disk based one which may be ACIDly robust, or simply efficient for non-critical data, or even implementations riding on other databases like mongoDB or even OrientDB which is a graph database.

The first kind will not be super scalable, it will depend on read-write locks, and use the STL for storage, such as the red black tree map.

It also seems that all classes, such as what defines an entity, property, and contracts will need to be implemented for each data engine individually.

I do not anticipate anything like running multiple engines within the same server.


So, for the next post, I’ll describe more about what the Return Formats are for the events, Commands if any, and possibly more details of an initial implementation.

My Current Todo list now that a lot of other things are depending on ride on making a template class for representing sets with the following features:

  • Add single entry
  • Minus single entry
  • Union with another set
  • Difference of another set
  • Symmetric difference with another set
  • Intersection of another set
  • Powerset of this set

I will likely use a vector<T> for storage, and use some of the existing standard functions in <algorithm>.