Archive

The Dulin Report

Browsable archive from the WordPress export.

Results (28)

Stop Shakespearizing Sep 16, 2022 Using GNU Make with JavaScript and Node.js to build AWS Lambda functions Sep 4, 2022 Monolithic repository vs a monolith Aug 23, 2022 TypeScript is a productivity problem in and of itself Apr 20, 2022 Node.js and Lambda deployment size restrictions Mar 1, 2021 What programming language to use for a brand new project? Feb 18, 2020 The religion of JavaScript Nov 26, 2018 Let’s talk cloud neutrality Sep 17, 2018 TypeScript starts where JavaScript leaves off Aug 2, 2017 Node.js is a perfect enterprise application platform Jul 30, 2017 Design patterns in TypeScript: Chain of Responsibility Jul 22, 2017 Singletons in TypeScript Jul 16, 2017 Collaborative work in the cloud: what I learned teaching my daughter how to code Dec 10, 2016 JavaScript as the language of the cloud Feb 20, 2016 Operations costs are the Achille's heel of NoSQL Nov 23, 2015 Ten Questions to Consider Before Choosing Cassandra Aug 8, 2015 The Three Myths About JavaScript Simplicity Jul 10, 2015 Big Data is not all about Hadoop May 30, 2015 Smart IT Departments Own Their Business API and Take Ownership of Data Governance May 13, 2015 We Need a Cloud Version of Cassandra May 7, 2015 Apple is (or was) the Biggest User of Apache Cassandra Apr 23, 2015 Building a Supercomputer in AWS: Is it even worth it ? Apr 13, 2015 Ordered Sets and Logs in Cassandra vs SQL Apr 8, 2015 Where AWS Elastic BeanStalk Could be Better Mar 3, 2015 Trying to Replace Cassandra with DynamoDB ? Not so fast Feb 2, 2015 Why I am Tempted to Replace Cassandra With DynamoDB Nov 13, 2014 Cassandra: Lessons Learned Jun 6, 2014 Best way to start writing an XSLT Jun 25, 2006

Design patterns in TypeScript: Chain of Responsibility

July 22, 2017

In event-driven systems messages produced by one object can be handled by one or more other objects. None of the objects need to know of one another – all they need to share is a common mechanism for distributing messages. Messages are sent from one object to another making them part of a chain. This pattern is called “Chain of Responsibility.”

Chain of responsibility in TypeScript


Node.js – and TypeScript by association – have idiomatic support for asynchronous event-driven programming. The best way to implement this pattern is by utilizing the EventEmitter class.

EventEmitter acts as a message bus. Events can be emitted or sent and can be subscribed to. Objects emitting events are independent of objects subscribing to events. This message bus needs to be a Singleton since it is shared between event publishers and event subscribers.

Source: ChainOfResponsibility.ts


import {EventEmitter} from 'events';

var messageBus=new EventEmitter();

function messageConsumer(msg:Object) {
console.log(JSON.stringify(msg));
}

messageBus.on("topic", (msg) => {
messageConsumer(msg);
});

messageBus.emit("topic", {
payload: "Hello world"
});

Case study: equities trading system backend


Events are useful constructs that represent real-world business workflows. In an equities trading system, there is typically a front-end application that allows traders to enter orders. The orders are published on a message bus, and there are multiple consumers:

  1. Order Management System (OMS) whose purpose is to capture and track orders and executions of those orders. Large orders are rarely executed all in one shot – they are typically executed as some number of smaller parts.

  2. Compliance and reporting system whose purpose is to track everything that is happening and analyze data for potential patterns of fraud.


There can be many more participants in the ecosystems – market data, reporting, etc. For this article we’ll stick to a simple simulation of OMS and compliance.

Source: HFT.ts


import {EventEmitter} from 'events';

var messageBus=new EventEmitter();

interface Order {
orderId: number,
side:"Buy"|"Sell",
symbol:string,
quantity:number
}

interface Execution {
orderId: number,
executionId: number,
symbol: string,
quantity: number,
price: number
}

class OrderManagementSystem {


constructor() {
messageBus.on("order", (orderMessage) => {
this.processOrder(orderMessage as Order);
});
messageBus.on("execution", (executionMessage) => {
this.processExecution(executionMessage as Execution);
});
}

processOrder(order:Order) {
console.log(new Date().getTime()+":OMS:order:"+JSON.stringify(order));
var numberOfExecutions=10;
var quantityPerExecution=order.quantity/numberOfExecutions;
for (var i=0;i<numberOfExecutions;i++) {
messageBus.emit("execution", {
orderId:order.orderId,
executionId: i,
symbol: order.symbol,
quantity: quantityPerExecution,
price: 101.5

});
}
}

processExecution(execution:Execution) {
console.log(new Date().getTime()+":OMS:execution:"+JSON.stringify(execution));
}
}

class ComplianceSystem {
constructor() {
messageBus.on("order", (orderMessage) => {
this.trackOrder(orderMessage as Order);
});
messageBus.on("execution", (executionMessage) => {
this.trackExecution(executionMessage as Execution);
})
}

trackOrder(order:Order) {
console.log(new Date().getTime()+":COMPLIANCE:order:"+JSON.stringify(order));
}

trackExecution(execution:Execution) {
console.log(new Date().getTime()+":COMPLIANCE:execution:"+JSON.stringify(execution));
}
}

var oms=new OrderManagementSystem();
var compliance=new ComplianceSystem();


/**
* Simulate orders coming in from the front end
**/
for (var orderId=0;orderId<10;orderId++) {
messageBus.emit("order", {
orderId: orderId,
symbol: "MSFT",
quantity: 1000,
side: "Buy"
} as Order);
}

Node events implications for concurrency and parallelism


Even though Node supports asynchronous event processing, it is important to remember that there is a single CPU thread per Node process. That means that asynchronous processing of events does not necessarily happen in parallel.

To achieve true parallel event processing, Node.js offers cluster framework. This framework supports multiple forked Node processes running on the same physical machine to operate in parallel and take advantage of multi-core CPUs.

Event processing at scale


While cluster framework helps with taking advantage of multi-core CPUs, that is often not enough. In large enterprise systems, it is not uncommon to have both event publishers and event subscribers existing independently of one another. In a cloud environment, it is also possible to auto-scale event processors based on workload requirements.

Message queues are crucial for scalable event processing. While it is beyond the scope of this article to get into the details of message queues, here are some examples: