CitizenCon 2951: Server Meshing & The State Of Persistence
- [Benoit] The concept of server, which has been ingrained since the beginning, has changed to become a mesh of servers. - [Paul] So how do we solve this? The answer should be simple. SERVER MESHING AND THE STATE OF PERSISTENCE Hi, my name is Paul Reindell, and I'm the Director for Online Technology here at CIG. I wanted to take this year's CitizenCon as an opportunity to give you some insight into exciting persistent streaming and server meshing technology.
In this talk, I will cover a quick overview of the current streaming and server architecture, and how we plan to transform the existing tech into what we call persistent streaming and server meshing. I also have Benoit Beauséjour with me, who later in the talk will give you more insight into the graph database that is powering persistent streaming. Hey, Ben. How's it going? - [Benoit] Hi, Paul. Hi, everybody. I'm super excited to share some of the details about what the Game Services team at Turbulent has been working on to support the efforts to build this technology and make it a reality for you guys. - Cool. Let's get started.
UNIVERSE SETUP Before we look into persistent streaming and server meshing, and how this new technology will work, let's have a brief look at entity streaming and how our solar system is set up. Each solar system can be seen as one giant level, containing every single object inside the solar system, from the sun, to the meteor rock, stand as one large map. Since this is a lot of data, the setup is split up into a hierarchy of nested object containers, which can be streamed in and out individually.
If you look at an abstract view of Stanton, it all starts as one solar system route object container. This object container contains the sun, the planets, and the moons around each planet. Each of those locations then has its own object container.
And if you take a closer look at the moon, you will find the entities placed around the moon inside this object container. For example, a space station orbiting the moon. This setup keeps repeating, and the space station could be set up by via multiple rooms, each defined by its own object container. Additionally, to the static hierarchy of object containers, there are also other dynamic entities which bring the universe to life, NPCs, an interactive vending machine, and of course, players and spaceships. Most of these entities are made of a hierarchy as well.
For example, a player has his body, an under suit, and armor attached to it. And they're all shard entities of the player. The streaming system treats these mini hierarchies as streaming groups to make sure that an object like a spaceship is always treat them as one unit.
Loading the entirety of Stanton into memory and simulating every single entity would be very expensive, especially on the client, but also on a single server. STREAMING AND REPLICATION That's why we developed entity bind culling and object container streaming, which allows us to stream object containers and streaming groups individually. When the game server starts, all entities and object containers in the solar system are loaded into local memory of that game server. These entities are not streamed in. We just store the initial state and server memory.
When a player connects, we create a so called streaming bubble around that player. And object containers as well as streaming groups that are visible from the player's point of view are considered inside this bubble. Any object container that is inside the bubble will stream its content, and any streaming group that's in the bubble will also be streamed in on the server and then replicated to the client. Entities are considered inside the stream bubble if their projected screen size on a virtual 1080P plane is larger than five pixel based on the distance of the player.
So while a large object like moon will be considered inside the bubble from far away, a small object like ship will only be considered inside when it's much closer to the player. When the player starts moving across the universe, entities that leave the streaming bubble will become unbound and the replication layer will remove these entities from the client. Entities that enter the streaming bubble will get bound to the client, which cause the network layer to replicate these entities to the client, effectively streaming them in. We call this technique entity bind culling because streaming on the client is driven by the network layer, binding and unbinding entities. If entities are not in any client streaming bubble, so no players in their vicinity, these entities are also streamed out on the server. They go back into a dormant state where they are not simulated.
PROBLEMS TO SOLVE This model works quite well on the client. However, it doesn't scale well on the server. While we do stream entities on the server if no player's close to them, the poor distribution of players will cause the DGS to load most entities. And the more clients we try to match to a given game server, the likelihood of a player being at every single location increases. And that basically nullifies the benefit of servers at streaming. So how do we solve this? The answer should be simple.
Allow multiple instances of the game server to work together so they can split up the work. Well, it's not quite that simple. Let's have a look at the current architecture. ARCHITECTURE As of today, we have a traditional client server architecture.
One instance of a dedicated game server serves up to 50 clients. This is called instance, as the dedicated game server has its own instance view of the persistent universe. Once a server is full, we start a new server instance, which then serves additional 50 players. As we've seen before, when a DGS instance is created, it loads a unique version of the Stanton System into its local server's memory. Therefore, each dedicated game server instance has a unique copy of every single object that's part of Stanton.
As these entities only exist in the memory of the game server, when instances shut down, these entities are deleted. REPLICATION LAYER The goal of server meshing is to allow multiple DGS instances to work together and divide simulation costs between each server and the mesh. In the best case, we can scale to infinity by adding more nodes to the mesh. As we saw earlier, each server node stores the state of entities locally. If you want to mesh these servers together, we need to find an efficient way to synchronize state between each server. With our current architect, depending on the vision of the simulation and the overlap, this would require a lot of synchronization points between each node.
It's an exponential problem, as in the worst case, each node would need to talk to each other node in the mesh, severely limiting our ability to scale it. To solve this issue, we are separating simulation and replication. Instead of just meshing multiple dedicated game servers together, and have them synchronize state between each other, we are introducing a new layer called replication layer. The replication layer has two major functions. It holds the state of every entity in memory and replicates the state to clients, but also to server nodes.
I said server nodes because in this set up, the traditional dedicated game server becomes a game server node. This server node connects to the replication layer, very similar to a client, and only a subset of entities are replicated to that server node. Replication to server nodes is controlled by the network bind culling algorithm that we saw earlier. And it's driven by streaming bubbles.
And it works very similar to how it works on clients. The server node has certain streaming bubbles assigned to it, which will cause the replication here to replicate entities from these streaming bubbles to the server node. Contrary to a player's client, a server node has the additional responsibility to execute server side authoritative code for those entities, controlling AI, doing damage calculations, etc., etc. The result of the simulation is then written back from the server node to the replication layer. And from there, it's replicated to all connected clients and other server nodes. Since streaming bubbles can overlap, entities may be replicated to multiple server nodes, exactly the same way how they are currently replicated to multiple clients of players at the same location.
To avoid two server nodes trying to simulate the same entity, only one server node can have authority over any given entity. And only that server's allowed to write entity state back to the replication layer. This is usually the first server node to replicate the entity and other server nodes will only run client code on those entities.
Basically, you can see a game server node as a client with authority, to write the result of its local simulation back to the replication layer. Authority can transfer between server nodes. For example, if an entity leaves the streaming bubble of the current authoritative server, it is then transferred to the next server node that has this entity currently streamed in.
Further authority can be transferred between server nodes on demand, in order to load balance the mesh. SHARDS AND PERSISTENT STREAMING Since we've now meshed multiple server instances together to simulate a shared state of the universe, we no longer call this instance, but instead, we call it shard. A shard is still a unique version of the universe, and we still have multiple shards running in parallel. However, the server mesh will lift our current hard limit of 50 players and it will enable us to steadily increase the number of players we can support within one shard. It will take some time, and in our first version of server meshing, we will still have a very similar situation as we have today, there's quite a few shards running in parallel.
However, this technology is going to enable us to start scaling the universe to become a true MMO experience. There are some fundamental differences between a shard and an instance. And for this, we need to take a closer look at the replication layer and talk a little bit about persistent streaming. Previously, the entity state was held entirely in memory on the dedicated game server.
And besides some selected persistent play items, all that state would be lost when the servers shut down or crashed. The replication layer is fundamentally different as the entire state of the universe is stored in a graph database. We call this entity graph and it's an evolution of the original iCache.
When we create a new shard, the initial entity state of the universe is seeded into this database. This happens offline before we let player join the shard. When the shard comes online, the replication mesh caches the state from the entity graph. As player connect to the shard and as we start to spin up new server nodes, simulation begins and alter the state of the universe. The replication layer does not only replicate the state changes to connect players and server nodes, it also replicates the state into the entity graph.
Since the entity graph is a persistent database, the state of the shard is never lost. And even if the shard is shut down, the state persistent can be resumed at a later time. Benoit is going to show you some more technical details about the entity graph. - Thank you, Paul.
Get ready for a deep dive into the entity graph persistent database. PERSISTENCE The entity graph is our approach to persisting the game world. This is fundamentally different to what is happening today in the game, where only items you own are actually stored.
Our objective is to be able to save the state of the replication layer, which includes all entities in a given universe shard, in order to provide a truly persistent world, where actions you take as a player can influence environments in the game world, permanently. The entity graph, as the name implies, stores game data as a graph. This representation is native for the game engine because it is how internally, those data structures the game uses, are addressed and manipulated.
Using a graph also has several advantages. We're basically storing and retrieving from a gigantic indexed list of nodes and edges. And those edges between those nodes are optimized in a sharded database. But in order to properly explain the system, we must first view the game world as the game engine sees it. Game objects are constructed of several game entities linked together in hierarchical structure.
You can picture this as a tree, which is a specialized kind of graph. This is how the game engine holds and simulates the elements on screen as it is running the simulation. In a server meshing world, this is also how replicants hold the entities in memory for each of their assigned territories. For example, a ship is made up of several entities that make up different parts of the entire playable vehicle. Each part is parented to another entity until the root of the ship is reached.
Each of these entity nodes holds properties with regard to what the entity represents in the game. That class of object it is, the item type, its legal owner, orientation, and of course, the very precise physical location within the game world. Each edge in our graph qualifies the relationship to the parent. In the case of the vehicle, our edges store properties that tell the system which port is being used to attach the entity into the parent, and what kind of attachment it is, an important attachment, a zone attachment, many others. In a constellation, for example, the different major sections of the hall are entity nodes with edges to the ship route. We call the small graph of item an aggregate, because it is a whole movable unit.
The ship route in this case is called the aggregate route because it sits at the top of a logical object. You can think of what you normally call an item as an aggregate, with the aggregate route being the actual item you're talking about. For example, a first person weapon with attached scope, mag clips, and laser sights, is a small hierarchy of entities. We distinguish the aggregate route from other nodes by giving it a label. Labels allow us to distinguish and rapidly look up and find nodes of a specific type, either when we retrieve parts of the graph, or when we look up specific nodes.
Those labels exist to allow correct reversal of the graph data when we query for specific things in the game world. Other labels include streaming groups, universe route, star system route. This allows us to really look up and index those types. But the tree depth doesn't stop there.
Additional information is required for a fully functional ship. The insides of each of those structural entities have to be flushed out. Object containers are the building blocks of how space division is achieved, which aggregates what part of the hierarchy they're in.
In fact, most major areas of ships are represented as OC entities attached to the ship route or another OC. The shape of this data actually takes in reality is driven by designers. Of course then, each of those object containers also contain entity hierarchies as well, expanding to have the common static and dynamic entities you're used to playing with, like elevators, beds, guns, seeds, gimbals, and others. In addition to object containers that make up the structure of the ship, other aggregates can also be attached within the hierarchy of our ship.
For example, a rover parked in the cargo bay of our colony will be attached as a sub-aggregate attached to your ship's cargo grid. Same goes for turrets, which are changeable, and themselves expose item ports, allowing guns to be mounted. For each of those entity nodes, a snapshot document is also stored. This document contains all the runtime values the game components attach to those entities.
This data is the dynamic part of the model where game developers can persist variables on any entity in the game world, according to the rules of a game component. For example, damage state and health data are stored within the snapshot document of those ship entities. Storing and retrieving data in graph form really have some awesome properties. It's a native structure of the game engine, so it gets loaded rapidly. It's very simple and effective to serialize and transport, because it's just a list of nodes and edges.
There are optimized databases that we can use that allow us to fetch entities recursively in a traversal rapidly. And the data set can be sharded across multiple database instance reliably. One key element here and one big advantage of having a graph data is also that we can reduce writes. So, in order to reduce that, all the hierarchical changes that we need to do can be minimized. For example, if we want a detach mutation command, will detach an entity from the hierarchy. In this case, a single edge must be erased, which is a very inexpensive operation.
A nice side effect of this is also that it's the same operation, whether we're detaching a single entity or an entire aggregate. In both cases, the single edge must be erased in order to perform the detach. Rolling away in your stowed rover, detaching a gun for a replacement, or selling a turret becomes really a cheap operation to persist. That is good because it happens very frequently. Compared to a columnar approach where index columns must be maintained for every write, linking all object to the aggregate route, this is a really great performance improvement.
The same properties apply with the attach command, which will only have to create a single edge when rejoining items to the hierarchy, be it via attachment or parking another vehicle on a docking hub. The attach and detach commands are two of the many semantic commands that the entity graph API proposes, allowing to express a mutation to the graph. Other example of the different commands are, create, possess, transfer, stack, unstack, change location, change snapshot, bury, stow, and unstow.
One important change, an addition that comes about with the entity graph is also how mutations are applied to the database. Each mutation is composed of multiple commands which are executed in sequence, but committed transactionally to the database. They either succeed together or fail and rollback together.
This ensures that the changes to the graph are always consistent, and no lost writes or errors can cause data corruption. For example, a mutation consisting of detached transfer and attach commands would succeed only if all three commands are applied successfully. The system retrieves a constant ordered streams of mutation from the replicant scribes that are part of the replication layer and are enqueued in durable queues, to ensure that no message is lost even if the service is unavailable or paused. It's important to understand that the graph does not only cover your ships and items, but the entire game world is made up this way.
Your ship is actually attached to the zone host location you travel in. Your playing character is attached to your ship seed when you are piloting it, just like planets are attached to their star system routes. SEEDING The game world, though, must exist in persistence, before it can be replicated and mutated.
This is part of a process called seeding, where a new database is created by the replication layer. It is during that process that millions of entities are initially created in a sort of a big bang. At this stage, every object container, every minor or major entity, from planets to doorknobs, are inserted into the entity graph in their default state.
That is the state that the designers decided was the initial state of the world. This process goes down from the universe route to the star system route, and into the different areas and planets, into their landing zone, their buildings, their rooms, down to the smallest possible entity. There are multiple types of entities that are created during this process.
First, are unstreamable entities, which make up the skeleton of the universe. Those are entities you do not get to see but are part of and are always present on every worker node in the mesh. It is by looking up unstreamable entities that the game world is able to stream in the other types of entities into your client and into the server mesh. It is from those entities that other entities bloom. Static entities make up the game world that you cannot interact with.
Most map objects that makeup buildings like the Hearthstone tower, rooms and walls of hospitals, or the bar G-Loc, are all made up of static entities. And the last type is dynamic entities. These are entities that you as a player can manipulate, a bottle on a bar, a door and a level, a ship component, everything you interact with when you're playing the game. Of course, during seeding, all object containers are also seeded as part of this hierarchy and inform the shape of the loadable sub-graphs. The seeding process takes a couple of minutes to complete. Once created, this newly seeded database represents a full dimension of the universe, and will now persist as it is modified by players.
As you play the game and go about with your ship, your playing character entity moves from location to location, getting attached to new zones as you travel. Your player aggregate is itself part of the [inaudible 00:21:15] graph, and your location and state are persisted by the replication layer scribes to the entity graph of your given territory. When you interact with dynamic objects and their properties change, the state of that entity will not persist until this instance of the database is undeployed.
There are in fact multiple copies of the universe that are seeded at a given time. We call those shards. Each shard is a unique copy of the game world, complete with all of its entities and unique states. Think of it as an alternate universe.
Dynamic entities that have been modified in each dimension will have different states. The bottle on the bar was moved or the door was destroyed, might not be in the same state between shards. This technique is a way to gain scalability as our player base grows. A single shard can grow to host multiple millions of entities.
Even if each shard database is itself clustered and can grow substantially past a single machine, there is a point where multiple clusters are needed. As you join the persistent universe ,the matchmaking system is getting retooled in order to select the correct universe shard for you to play on, using multiple data points, like your friends' location, your active party, your last game session, and/or which shards still have items on it that you own. This is to ensure, as much as possible, that you end up on the same shard you expect to be as a player.
GLOBAL VS SHARD DATA In order to provide a seamless game experience, it would be terrible if you lost items you used when you were in a given shard versus another, or if your character was bound to a shard forever. To alleviate this, the system includes the concept of stowed and unstowed entities. An item is considered unstowed when it is currently active in a shard database and being actively simulated on by the worker nodes of the replication layer. Stowed entities are player owned entities that are stowed in inventory containers or location inventories.
Those entities live in another database entirely called the global database, a large cluster database that spans all shards. Aggregate routes in that shard are stored and linked with edges to inventory nodes. Any entity in a shard can have an inventory node in global for stowing things in it. For example, a box entity that is unstowed in a shard would have an inventory node in the global database to store its content. This allows to keep unstimulated entities in a non-shard specific database, while keeping the live aggregate within the shard. As you transition between shards, your playing character gets unstowed into the selected chart.
This process effectively moves your player aggregate data from the global database to the shard database. Your player entity now gets simulated by game worker nodes and is being updated at a regular rate as you play and move around the game world. Accessing items that are stowed like a ship from the Aesop terminal is basically reading the inventory contents of the global graph at your location. Same goes for personal inventories, or cargo inventories. When you request the ship to be spawn, the system will unstow the ship onto a landing pad by submitting a shard mutation.
Alternatively, when a ship gets despawn during parking at a rest stop, the ship gets stowed back into the global database, making it available for unstow in any shard. The global database is where all of your stowed items will live. Hero items will always be available for unstow in any shard, if they are not already in use. The process of stowing and unstowing also helps to alleviate problems related to entity authority so that only game worker nodes in the right shard can update unstowed entities in that shard.
This brings about a nice property of the server meshing and persistent streaming architecture, in that the state of the entities are being persisted transactionally during play, be it in a shard or a global database through stowing. A single server crash or 30K should no longer result in item loss. This model also has a real scalability benefit that stems from the separation of the read intensive workloads that are isolated to the global DB, from the write intensive workloads that are handled by the individual shard database. The global graph exists to provide seamless access to your belongings no matter what shard or alternate universe you're currently playing in. BENEFITS AND CHALLENGES Okay, let's go back to Paul to learn about some of the benefits of the server meshing architecture. - Yeah.
The first benefit is obviously the advantage that we don't have this issue of synchronization between different server nodes. Each server node has one single connection to the replication layer, which is used to push and get updates for entities of interest. The second advantage is that the same streaming and replication logic that we already use for clients can be applied to servers. And that server nodes will only stream in a small area, which will greatly increase performance.
It also allows us to increase resilience down the road. As long as the client is connected to the replication layer, the client stays connected even if the server node crashes. In this case, the simulation for an entity may be stopped for a moment, but as soon as a new server comes online, the simulation will just continue. While the underlying tech is close to completion, there are some upcoming challenges that we need to solve before we can get that into your hands.
The first version of this technology will contain a static server mesh. Instead of the fully dynamic mesh that we saw earlier, the static mesh assigns server nodes to predefined sections of the solar system. This will reduce the amount of authority transfer that game code has to address in this first release. There will also be a lot of challenges for the game services and game feature teams. Maybe Benoit can give a little bit more.
- Yeah. There are many parts of the game that are affected by this new server meshing architecture. So any game play features that has to rely on the concept of a server, right? Currently, when you connect to a game server, we know what match you're in. So to send messages and update to that server, we simply locate your active match, and then send those messages out there.
That concept needs to change because we now have a mesh to deal with. And so there are multiple game servers that need to receive this information. They need to be able to subscribe dynamically to it or unsubscribe dynamically to players transitioning through them. So you can imagine that this will affect things like missions that currently are spawned locally on the game server. These now need to be spawned globally within the shard and also persist their state. So all services that are attached to missions, whether it's the quantum system in the back end, or the quasar tools, need to now know about the concept of a shard.
This also goes deep into, like, things that are mechanical. Like, you know, getting global chat to work on a server. That concept now needs to be extended to the shard where this will probably push us to implement this as a location based chat, for example. And so many teams in the company now need to change their feature to take into account the meshing technology that's behind it, because the concept of server which has been ingrained since the beginning has changed to become a mesh of servers. - Yeah, exactly. So I also want to shout out to the network team who's working on the replication layer and the bind culling, as well as the persistent tech team who's working on the entity and object container streaming.
And as Benoit said also, all the other teams that work on game play features or game play services that are affected by this new technology. There are a lot of devs working on this and we're very excited to push on this new technology. Thank you for your time.
♪ [music] ♪