Web Agents
Swim servers utilize a general purpose distributed object model in which the objects are called Web Agents. Programming with this model feels like typical object-oriented programming with additional key innovations in addressability, statefulness, consistency, boundedness, and composability.
Although this analogy holds very well for the most part, there are two important catches:
- Methods, while still being able to define arbitrarily complicated logic, are not directly invoked. Instead, Web agents come with lifecycle callback functions that are called during specified stages of a Web Agent’s lifetime
- Web Agent instantiation is not accomplished by invoking a constructor (at least from the programmer’s perspective)
Don’t worry if these points feel restrictive through this article; much finer control becomes available once we discuss lanes.
Declaration
Just like with (instantiable) class
declarations in Java, Agent declarations in Swim define the behavior for instances of those Agents.
Declarations alone don’t actually instantiate anything.
To declare a Web Agent, extend the AbstractAgent
class from the swim.api
module:
// swim/basic/UnitAgent.java
package swim.basic;
import swim.api.agent.AbstractAgent;
public class UnitAgent extends AbstractAgent {
}
External Addressability
Every Web Agent has a universal, logical address, in the form of a URI.
The URI of a custom PlanetAgent
might look something like "/planet/Mercury"
.
That of a singleton SunAgent
might just look like "/sun"
.
By decoupling Web Agent’s logical addresses from the network addresses of their host machines, Swim applications become invariant to the infrastructure on which they’re deployed.
Each Web Agent is aware of its own URI, available via its nodeUri()
method.
Let’s add a simple utility method to each UnitAgent
to help us identify the Agent from which a logged message originated.
// swim/basic/UnitAgent.java
package swim.basic;
import swim.api.agent.AbstractAgent;
public class UnitAgent extends AbstractAgent {
private void logMessage(Object msg) {
System.out.println(nodeUri() + ": " + msg);
}
}
Instantiation
For an Agent to know its own identifier is only half of the problem. To address the other half, every Swim server runs a plane that manages the runtime of and provides a shared context for a group of Web Agents.
One of a plane’s many responsibilities is to resolve Agent URIs for requests.
To declare a Web Agent’s URI, simply define a node with URI in the server configuration file.
To declare a dynamic component, we prepend with a colon (:) for the agent id.
# server.recon
basic: @fabric {
@plane(class: "swim.basic.BasicPlane")
# Static Component
# Notice usage of 'uri' keyword for a static component.
@node {
uri: "/unit/foo"
@agent(class: "swim.basic.UnitAgent")
}
# Dynamic Component
# Notice usage of 'pattern' keyword for a dynamic component.
@node {
pattern: "/unit/:id"
@agent(class: "swim.basic.UnitAgent")
}
}
@web(port: 9001) {
space: "basic"
}
Statically defined agents (illustrated using ‘uri’ keyword) are automatically instantiated by the corresponding Swim Plane.
A Dynamic Web Agent is only instantiated when its nodeUri
is invoked for the first time.
With the code we have so far, we can instantiate any number of UnitAgent
s by either defining them in the configuration file or by invoking URIs with the "/unit/"
prefix.
For example, if we invoke "/unit/1"
, "/unit/foo"
, and "/unit/foo_1"
, three UnitAgent
s will be instantiated, one for each URI.
Caution
If you have multiple agent types within a plane, ensure that their URI patterns do not **clash**. This is a stricter requirement than saying that the patterns are identical; for example, "/unit/:id" and "/unit/:foo" clash. Suppose these same patterns annotated different agent types; how would a plane know which type of Agent to seek or instantiate for the request "/unit/1"?
In addition to the nodeUri()
method mentioned in the previous section, every Agent also has access to a Value getProp(String prop)
convenience method.
This returns a swim.structure.Text
object containing the value of the dynamic nodeUri
component with the name prop
, absent()
if it doesn’t exist.
For example, getProp("id").stringValue()
will return either "1"
, "foo"
, or "foo_1"
, depending on which of the above three agents we are currently running in. getProp("foo")
will return absent()
.
Further reading: Planes
Lifecycle Callbacks
Recall that Web Agent methods are not directly invoked. Instead, the Swim runtime schedules and executes callbacks stages of an Agent’s lifecycle. For the most part, you will only care about two:
-
void didStart()
: executed once immediately after this Agent has started -
void willStop()
: executed once immediately before this Agent will stop
// swim/basic/UnitAgent.java
package swim.basic;
import swim.api.agent.AbstractAgent;
public class UnitAgent extends AbstractAgent {
@Override
public void didStart() {
logMessage("didStart");
}
@Override
public void willStop() {
logMessage("willStop");
}
private void logMessage(Object msg) {
System.out.println(nodeUri() + ": " + msg);
}
}
Try It Yourself
A standalone project that combines all of these snippets and handles any remaining boilerplate is available here.
Web Agent Principles
Universally Addressable
Every Web Agent has a universal, logical address, in the form of a URI. By decoupling Web Agent’s logical addresses from the network addresses of their host machines, Swim applications become invariant to the infrastructure on which they’re deployed.
Stateful
Web Agents remember their state in-between operations, eliminating the need for constant database round-trips, and greatly simplifying application development by avoiding object-relational mapping.
Atomic
Each Web Agent executes in a single thread at a time. Though as many distinct Web Agents execute in parallel as you have CPU cores. Combined with a built-in software transactional memory model, Web Agents are naturally atomic, without the overhead of locks or transactions.
Consistent
Together, Web Agents, Lanes, and Links implement a continuous consistency model that’s largely transparent developers. Web Agents applications just stays consistent. Continuously. In network real-time.
Encapsulated
The only way in or out of a Web Agent is through links to its lanes. This gives Web Agents total control over the exposure of sensitive data. There’s no database to compromise.
Persistent
Databases aren’t the only way to store data. Web Agents are internally persistent. By taking persistence off the critical path, the single biggest bottleneck to application performances instantly vanishes. While still letting you keep all the data you have space for.
Bounded
Intrinsic and pervasive backpressure handling automatically adapts the behavior of your application based on network, disk, and CPU availability. And because of its continuous consistency model, developers, for the most part, don’t have to care.
Decentralized
Web Agents inherit the natural decentralization of the World Wide Web. Any Web Agent can link to any other, given its URI, and appropriate permissions.
Composable
Unlink REST applications, which don’t compose well without introducing significantly polling latency, caching overhead, and consistency problems, Web Agents frictionlessly compose, in real-time, at any scale.
As the name implies, Web Agents were designed from first principles to be first class citizens of the World Wide Web. The Web has evolved from a world-wide hypertext library, into the lingua franca of distributed applications. But the technical foundation of the Web, stateless remote procedurce calls over HTTP, is fundamentally incapable of meeting the needs of modern, autonomous, collaborative applications. Web Agents aim to fill that gap.