Nodes is a library to implement asynchronous dependency graphs for services in Java.
When you write an asynchronous service, like an RPC server, or just a non-blocking library to do some work, you may need to implement an interface like this:
Future<Response> processRequest(Request req)Your logic goes inside processRequest(), you may do some local computation here, call some external services, or even run some code in another thread. The logic dependencies in these code could be complicated. Finagle provides a good paradigm for concurrent programming with Futures. The Future class is a great building block, it decouples the execution logic and the thread scheduling, so the developer can focus on the logic dependencies rather than their actual execution. However, this library was written in Scala and isn't exactly Java friendly. It naturally involves a lot of callbacks and repeated function signatures. When it comes to waiting on multiple Futures, the code gets ugly very fast. Nodes is a Java library that aims to solve these problems, making the asynchronous code easier to read, to maintain and to test in Java.
A node is an asynchronous processing unit. It takes multiple asynchronous input nodes (dependencies), and produces a single output. It will only start executing when all inputs are ready. A node object is also a handle to its output. Like Future<A> in Finagle, Node<A> represents an asynchronously computed data of type A. Actually Node and Future are mutually convertible. The tutorials below will show you how to create nodes and assemble them into a dependency graph.
Another way to understand nodes is to consider them asynchronous functions, every node with input of type A, B, C and return type of X can be thought as a function with signature:
Future<X> process(Future<A> a, Future<B> b, Future<C> c)For quick examples, please see src/main/java/com/twitter/nodes_examples.
If your node has a fixed number (more than 1) of dependencies, the most common way to create a Node is as follows:
// A node that produces an integer as output
public class MyNode extends Node<Integer> {
public enum D { // "D" is a convention
DEP1,
DEP2,
@OptionalDep DEP3 // this dependency is optional, may be omitted
}
@Override
protected Future<Integer> evaluate() throws Exception {
// To get the value of a dependency nodes, use getDep(key)
Type1 value1 = getDep(D.DEP1);
Type2 value2 = getDep(D.DEP2);
Type3 value3 = getDep(D.DEP3, someDefault);
Integer result = compute(value1, value2, value3);
return Future.value(result);
}
}You need write a public class extending Node (or NullableNode, more on this later) with a certain type, inside which you define enums for all your list of dependencies (the name convention is D). When you instantiate the node, you will use these enums to mark out your dependencies.
A dependency is by default required (without any marking), which means throwing an exception in it would fail the depending node directly and cause them not to get executed at all. Optional dependencies are marked as @OptionalDep, allowing them to be omitted or to fail.
Under the hood: inside Node uses an
EnumMapto store all dependency nodes. Even if you don't define your custom enum, there is a default one calledNode.DefaultDependencyEnumthat has 16 entries:DEP0,DEP1...DEP15. If you directly use the constructor of the parent class to specify all your depdnencies, you will need to pass in a list of nodes, which will be mapped to these default enums in the order they appear.
You need to implement the evaluate() method, which will only be called when all dependencies are ready, and called only once. You can get values for each of the dependencies using getDep() and use them in your computation. At the end, you are supposed to return a Future object with desired return type. If you don't have any asynchronous processing inside, you can just wrap your output with Future.value(). However, you can also call other asynchronous services like a remote server, or submit tasks to another thread, and pass back the Future object you acquired from them directly.
There is always a simple way to create a multiple-dependency node inline with Java 8 lambda functions, with support for up to 4 inputs. See Map with Multiple Inputs for more details.
To instantiate your node with enum-based dependencies:
Node<Integer> resultNode = Node.build(
MyNode.class,
MyNode.D.DEP1, node1,
MyNode.D.DEP2, node2,
MyNode.D.DEP3, node3);You can omit the optional dependency without causing any runtime exception.
Node<Integer> resultNode = Node.build(
MyNode.class,
MyNode.D.DEP1, node1,
MyNode.D.DEP2, node2);Or the input can even be a failure, for example:
Node<Integer> resultNode = Node.build(
MyNode.class,
MyNode.D.DEP1, node1,
MyNode.D.DEP2, node2,
MyNode.D.DEP3, Node.fail(new Exception()));Normally a node shouldn't return a null value, an exception would be thrown if it encounters a null output. You should try to use the control flow methods (see below) to manage the execution and make null value unnecessary, or utilize Optional<A> to represent return values that can be null. However, we also provide NullableNode which you can extend from, they can return null without causing any exception. The reason we make default Node class null-unfriendly is to make it easy to reason what went wrong during the execution, and not to confuse an error with a non-existent value, but overall this is a matter of style. The NullableNode is convenient but you should use it with care.
Most of the nodes are just doing local computations. For nodes that calls any asynchronous service, be it an external Thrift RPC server, an HTTP server, or just a in-process scheduler that runs tasks in a thread pool, you can extend from ServiceNode, which provides some convenience for dealing with services. Rather than implementing evaluate(), you need to implement buildRequest() and getService().
The method buildRequst() builds the request for current service call, in the parent class it's called in evaluate() so all dependencies should already be ready.
The method getService() gets a Finagle Service object, with has a method Future<Response> apply(Request req). You can get this object from some factory, some registry, or simply from some static variables.
public class NodeServiceNode extends ServiceNode<Response> {
@Override
public Service<Request, Response> getService() {
...
}
@Override
public Request buildRequest() {
// this is called inside evaluate(), you can get all your dependencies
// the same way, or just get them from your own member variable.
}
}To specify the dependencies, you can either use the same enum-based solution described above, or directly use Node's own constructor (which takes a list of nodes). You can even implement your own constructor if it takes some special inputs more than dependending nodes.
A node (which also denotes a dependency tree rooted at it) doesn't get executed automatically after they are constructed. The instantiation of the node only defines the logic dependency, but nothing has executed yet. To start running, you need to call apply():
Future<A> aFuture = aNode.apply();This creates a Future of the node, which you can wait on. This will trigger the evaluation of all its dependencies and cause their apply() to get called, and in turn trigger the execution recursively. You start from the root of a tree and it eventually reaches all leaves, which are inputs. Only the nodes reachable by walking the dependency links will be executed.
You can just pass this Future object to anywhere it's needed, like inside the processRequest() method of your server implementation in the example at the beginning. If you want to block on it and get its value, you can call:
A a = Await.result(aNode.apply(), Duration.ofSeconds(2));As mentioned above, only the node reachable by the dependency link from an executing node will ever be executed. However, there are situations where we want a node to execute but no one is waiting for its result (like you want to start and forget a logging process after a request has been processed, but you can respond without waiting for that logging work to be done). You can implement this logic with Sinks, a set of nodes you attach to another node which gets executed after it has finished. It's like a reverse dependency.
Node<Integer> resultNode = Node.builder(MyNode.class)
.withSinkNodes(s1, s2, ...)
.withDependencies(
MyNode.D.DEP1, node1,
MyNode.D.DEP2, node2,
MyNode.D.DEP3, node3);
// You can also set the sink after the node is created
resultNode.setSinkNodes(s1, s2, ...);Nodes can be selectively turned on and off by "Deciders", which is just a simple Supplier object that produces a true/false value. You can use this to implement runtime control of your nodes. Having supplier return false would cause node not to execute and return failure.
// You can add deciders and sinks as well
Node<Integer> resultNode = Node.builder(MyNode.class)
.withDeciderSupplier(someDeciderSupplier)
.withDependencies(
MyNode.D.DEP1, node1,
MyNode.D.DEP2, node2,
MyNode.D.DEP3, node3);You node doesn't need to always have a computation class, it can directly wrap a value or a Future object.
// To wrap node around an immediate value, without any computation process:
Node<A> nodeA = Node.value(somethingOfTypeA);
// To wrap node around a Finagle Future of type A
Future<A> futureA = ...;
Node<A> nodeA = Node.wrapFuture(futureA);As we mentioned above, a Node can be converted to a Future by just calling apply():
Node<A> nodeA = ...;
Future<A> futureA = nodeA.apply();A future may fail and return exception, if it's used as a required dependency of another node, it would fail that node. Of course you can make it an optional dependency, but it may not always be feasible when you are calling some existing node code. You can make create a "safe" node masking Future's failure. All failures will simply become null result.
Node<A> nodeA = ...; // this may fail
Future<A> futureA = nodeA.toFutureSafe();There are some node literals for your convenience.
Node.TRUE // Node<Boolean> with True value
Node.FALSE // Node<Boolean> with False value
Node.NULL_NODE // typeless Node with a null value
Node.<T>noValue() // typed Node with a null value
// To create a literal failed node
Node.fail(Throwable t)Putting your whole graph in a single file is probably not convenient. Subgraph provides a way to organize your graph. Subgraph is a subset of the graph (duh), its inputs/outputs consist of the union of all inputs/outputs of the nodes at the boundary, so it can have more than 1 input and more than 1 output.
Subgraph doesn't have any effect on the execution of the nodes, it's just a way to organize nodes into a more reusable way. Below is a recommended style to implement Subgraphs.
public class MySubgraph extends Subgraph {
public final Node<Type1> exposedNode1;
public final Node<Type2> exposedNode2;
public MySubgraph(Node<A> inputNode1,
Node<B> inputNode2,
Node<C> inputNode3) {
// ... all the wiring
this.exposedNode1 = ...
this.exposedNode2 = ...
// only needed if you need to generate DOT graph
markExposedNodes();
}
}
// Constructing two copies of the same subgraph, using different inputs:
MySubgraph s1 = new MySubgraph(node1, node2, node3);
MySubgraph s2 = new MySubgraph(node4, node5, node6);
// Access its produced outputs referencing the public members:
// s1.exposedNode1
// s2.exposedNode2The suggested style is:
- Subclass from Subgraph
- All input nodes as constructor arguments
- All output nodes as public member variables
- You can also have other non-node input to configure the subgraph and provide other information.
The existence of subgraphs is transparent to the final dependency graph. You can always flatten a subgraph at its caller side, or split a subgraph into more smaller subgraphs if that helps with your modularity.
The call to the markExposedNodes() method at the end is purely for bookkeeping, this will help you generate better visualization/debug message, but it has no impact on the execution. Even if you forget to call it, it not a big deal.
For nodes with only a single input/dependency, it's cumbersome to create a whole new node class. We provide many syntactic sugar for you to do this kind of transformations easily.
You can convert a node of type A to type B using a map. You only need to provide a function (in the form of a name string and a Java function, or a NamedFunction object). You can also do a flatMap where your function returns a Future<B> rather than straight type B, so you can call other asynchronous process inside.
// To convert the node of one type to another, synchronously
Node<B> bNode = aNode.map(namedFunction); // map with a NamedFunction<A,B>
Node<B> bNode = aNode.mapOnSuccess(func); // only when aNode is successfully done, or emit null
Node<B> bNode = aNode.mapWithDeciderSupplier(ds, func); // only when decider is on
// To convert the node of one type to another, asynchronously
Node<B> bNode = aNode.flatMap(func); // where func is an instance of NamedFunction<A, Future<B>>
Node<B> bNode = aNode.flatMapOnSuccess(func); // only when aNode is successfully done, or emit null
Node<B> bNode = aNode.flatMapWithDeciderSupplier(ds, func); // only when aNode is successfully done, or emit null
// To get a boolean value out of node:
Node<Boolean> booleanNode = aNode.predicate(pred); // where pred is a NamedPredicate<A>
// Java 8 support: you can also use Java 8 function in all above cases, except that you need to provide a name
// as the first argument. The name is only used for tracking and debugging, and ignored in production.
// For example:
Node<B> bNode = aNode.map("name", a -> doSomethingTo(a));
Node<B> bNode = aNode.map("name", a -> { ... });
Node<B> bNode = aNode.map("name", SomeClass::staticMethodOnA);
Node<B> bNode = aNode.map("name", someObject::instanceMethodOnA);
// same applies to .mapOnSuccess(), .flatMap(), etcActually we also support map with multiple inputs, this can be used to implement simple multi-dependency node as well, rather than using enum-based dependency definitions:
Node<X> nodeX = Node.map2(
"map2", nodeA, nodeB, (a, b) -> {...})
Node<X> nodeX = Node.map3(
"map3", nodeA, nodeB, nodeC, (a, b, c) -> {...})
Node<X> nodeX = Node.map4(
"map4", nodeA, nodeB, nodeC, nodeD, (a, b, c, d) -> {...})Also there are flatMap versions, the function passed in should return a Future<X> instead.
You can use these mappers/flatMappers to implement simple nodes with no optional dependencies. You can test them by just testing the normal function passed in.
We provide some convenient methods for manipulating nodes wiht boolean values:
// Boolean nodes
Node<Boolean> booleaNode = AndNode.create(b1Node, b2Node, b3Node, ...);
Node<Boolean> booleaNode = AndNode.createLazy(b1Node, b2Node, b3Node, ...); // short-circuiting, will not evaluate later nodes
Node<Boolean> booleaNode = OrNode.create(b1Node, b2Node, b3Node, ...);
Node<Boolean> booleaNode = OrNode.createLazy(b1Node, b2Node, b3Node, ...); // short-circuiting, ditto
Node<Boolean> booleaNode = NotNode.create(b1Node);Note that the "lazy" version would short-circuit the execution, not all input nodes are to be executed, only the first input is its real dependency.
We also have some other transformations that deal with 2 nodes.
Node<A> nodeA = ...;
Node<B> nodeB = ...;
Node<List<A>> listNode = NodeUtils.asList(nodeA);
Node<Pair<A, B>> pairNode = NodeUtils.asPair(nodeA, nodeB);You can also convert a list of nodes (of the same type) to a node with the list type. This is handy when you need to call unknown number of async processes and handle all of their responses together.
List<Node<A>> listOfNodeA = ...
Node<List<A>> node = Node.collect(listOfNodeA);Similarly, there is a Map based version:
Map<A, Node<B>> nodesByKey = ...
Node<Map<A, B>> node = Node.collect(nodesByKey);You can also apply some transformation to each entry in a list under a node, and collect all results:
Node<List<A>> alistNode = ...
Node<List<B>> blistNode = Node.splitAndCollect(
alistNode, "functionname", a -> { ... return something Node<B> ... });You can take a look at their implementation and you can also build your own fancy node transformations.
You can implement if-then-else logic with Nodes. There is also syntactic sugar to represent some other common condition checks, like whether a node has finished running successfully (more on "successful nodes" later).
Simple if-then-else control flow
// An if-else switch between two nodes:
Node<C> cNode = Node.ifThenElse(booleanNode, cNode1, cNode2);
Node<C> cNode = Node.ifThen(booleanNode, cNode1); // equivalent to cNode2 being Node.noValue()
// Simpler if-else
Node<C> cNode = xNode.when(booleanNode); // gets xNode if booleanNode emits true, or Node.noValue()
Node<C> cNode = xNode.unless(booleanNode); // gets xNode if booleanNode emits false, or Node.noValue()Success-based control flow
// On success
Node<C> cNode = xNode.whenSuccess(anyNode); // gets xNode if anyNode is success, or Node.noValue();
Node<C> cNode = xNode.orElse(yNode); // gets xNode if xNode is a success, otherwise gets yNodeNodes provides a simple framework for collecting debug messages in the asychrounous execution of the dependency graph. You can use the DebugManager class with a flexibly-scoped thread-local DebugMessageBuilder inside. For each execution of a graph, you can set a new DebugMessageBuilder and all debug messages will be collected over there.
DebugManager.update(new DebugMessageBuilder(DebugLevel.DEBUG_DETAILED));You can choose from multiple DebugLevels, When you produce the debug message, only those no higher than the set level would be collected.
Anywhere in your node's evaluate() method or other non-static methods, you can call following member methods to append a message to the current DebugMessageBuilder. It supports String.format() style formatters. A line break will be added automatically after each call.
basic("current score: %d", score);
detailed("input type = %, parameters = %s", type, params);
verbose("server full response: %s", response);
verbose2("detailed debug info: %s", stuff);
verbose3("super detailed debug info: %s", stuff);If you are not in the scope of a Node subclass, you can still append message by calling similar methods in DebugManager.
DebugManager.basic(...)
DebugManager.detailed(...)
DebugManager.verbose(...)
DebugManager.verbose2(...)
DebugManager.verbose3(...)All these messages will be properly prefixed with the node name and timestamp. Node itself also produces some debug messages to mark the beginning and end of its execution, storing the time spent in the node, etc.
To get the debug message after the node has finish executing, just call DebugManager.getDebugMessage().
If you run the sample code in `src/main/java/com/twitter/nodes_examples/search/SearchExampleMain.java', you will get debug messages like:
[5184] NODE [hasQuery]: Start
[5185] NODE [hasQuery]: End (52/0 ms)
[5185] NODE [IF::hasQuery(BuildResponseNode, null)]: Start
[5196] NODE [hasUserId]: Start
[5196] NODE [getQuery]: Start
[5196] NODE [getNumResults]: Start
[5196] NODE [hasUserId]: End (2/0 ms)
[5196] NODE [getQuery]: End (1/0 ms)
[5197] NODE [getNumResults]: End (2/0 ms)
[5197] NODE [IF::hasUserId(UserScoreServiceNode[?], defaultUserScore)]: Start
[5197] NODE [SearchIndexNode]: Start
[5206] NODE [getUserId]: Start
[5206] NODE [getUserId]: End (1/0 ms)
[5207] NODE [UserScoreServiceNode[?]]: Start
[5207] NODE [UserScoreServiceNode[?]]: built service request: 777
[5854] NODE [SearchIndexNode]: End (659/8 ms)
...
The number at the beginning of each line is the last 4 digits of current time in milliseconds, so you have a sense when each step was run.
There are some naming conventions for node code we'd like to suggest:
- Start your node class and subgraph class name with a verb, like
CreateRecordNodeorResolveGeoLocationSubgraph. - Name your node object as a noun and always end with
Node, describing the product rather than the process, likerecordNode.
It's easy to write tests for nodes. Just instantiate them and test their output with specific inputs.
- if you have any optional input, make sure you have some tests without these inputs.
- for subgraphs, try to test the wiring, or the control flow implemented in the graph, rather than the logic of nodes inside.
You can visualize your node by generating a DOT file.
// for nodes
String dot = mynode.toDotGraph();
// for subgraphs
String dot = mygraph.toDotGraph();
// You can save these strings to a file to be rendered by other tools.
Files.write(mygraph.toDotGraph().getBytes(), new File("graph.dot"));You can render the generated file using Graphviz (download here, you may also need X11). Check out the DOT language reference if you want to understand the generated file, or read NodeDotGraphGenerator.java. A rendered graph looks like this:
The type of the node is indicated by its shape, while the optionality of dependencies are represented by the edges connecting them.
For nodes:
- pink box: value nodes
- green-yellow trapezoid: transform node
- green-blue inverted trapezoid: predicate switch node (with "condition" and true/false branch), created by
Node.ifThen()orNode.ifThenElse() - gray double-edged box: service node
- white square box: other normal nodes
For edges (dependencies):
- solid: required dependency
- dashed: optional dependency
- label: dependency name, useful especially if you use enum style suggested above.
To make your graph look nice, always remember naming TransformNode (give it method-like name) and ValueNode (give it variable like name).
You can download the code by cloning this repo:
git clone https://github.com/twitter/nodes
Go into the newly created nodes directory, to compile and run tests:
mvn compile
mvn test
You may see many exception stacktrace in the output, ignore them, they are expected.
To create a .jar file for nodes, run:
mvn package
and you shall find the .jar file at dist/lib/nodes-1.0.0.jar.
This library depends on following Twitter libraries:
- Twitter util (util-core, util-logging)
- Finagle (finagle-core)
Please see pom.xml for other dependencies.
You can join our mailing list twitter-nodes@googlegroups.com on Google Groups, you can check out the archived threads on web.
Copyright 2016 Twitter, Inc and other contributors
Licensed under the Apache License, Version 2.0: http://www.apache.org/licenses/LICENSE-2.0