Wednesday Mar 11, 2015

WebSocket vs REST

As controversial or potentially “flame starting” topic this might seem to be, don’t worry. I will approach this purely from pro-WebSocket view and the comparison with REST will be done on the sample, which heavily favours WebSocket ;)

Not that long ago, I had to explain one of my colleague where he should consider using WebSocket protocol and I realised, that lots of people don’t really know about it much. I even heard question whether WebSocket is successor of REST, like REST was/is to SOAP web services.. well, I’ll try to make this little bit clearer in this post.

For starters, WebSocket is NOT REST replacement. These are two technologies, which can coexist very nicely even in single application or webpage. Both are doing similar things and for some applications are even interchangeable. Bold statement, but it’s true. Both approaches have it’s own pros and cons, as with everything else..

Let’s go little back to the history of web services and remember why was WebSocket protocol even created – to allow bi-directional communication with clients, mainly represented by web pages. It was (and still is) possible to achieve the same with plain REST, but there are some issues with it. Let’s name two of them:

  • REST is always Request/Response “stateless” communication,
  • by the nature of the HTTP protocol, lots of information must be sent in each Request and response.

The first one implies simple fact – web server cannot send anything to the webpage without a Request. There are various workarounds (yes, workarounds. First real standard solution is WebSocket protocol) like long polling or JSONP, but they are solving only the communication from server to client, which implies that there needs to be the other channel from client to server. And we are getting to the second item in short list above – efficiency. When the application needs to communicate frequently with the server, the volume of HTTP traffic can be really big. Now try to compute the entropy (how much information is acquired due to observation of the Request/Response) and see how much redundant and unimportant bytes is sent with each HTTP communication. This is already addressed in HTTP/2, but anyway, the overhead still exists. I intentionally skip the HTTP/2 server push implementation – I plan to address that in another blogpost.

Enough with history lesson and plain “code-less” chatter. You might remember Shared collection sample introduced couple of moths ago in Tyrus workspace – the last modification of it was inspired by the discussion with that colleague – I wanted to compare REST and WebSocket implementation of the same thing.

Quick recapitulation – the sample exposes a map-like object in JavaScript and Java and synchronises their changes using WebSocket. Additionally you can register listeners on both sides, so you always know when anyone changes anything.. basically very simple and fragile implementation of coherence/hazelcast/yourFavouriteDistributedFramework.

As you most likely already expect, I implemented the same thing (or extended that implementation) to support REST based transport instead of WebSocket. I must admit I cheated a little – I used Server-Sent Events (SSE), which is not really RESTful, but this feature is implemented as part of Jersey, which is Reference Implementation of JAX-RS: The JavaTM API for RESTful Web Services and also my favourite REST framework (I used to be a contributor, so consider this as another very-impartial fact). SSE also has simple JavaScript API implemented in all modern browsers, so it was an easy decision for me – I have the channel from server to client covered and the other way will be just standard request using XMLHttpRequest (or an ActiveXObject when in M$ Internet Explorer).

Below is the simple scheme with mentioned protocols.

Screen Shot 2015-03-11 at 16.10.45

When you compile and deploy the sample, the standard behaviour is actually quite comparable (I’m on localhost, so that is not that much surprising), both maps receive and send updates and the experience is almost the same. So let’s look under the hood…

When you create or modify a map entry, browsers sends and event. In case of WebSocket, it is short (text) message containing Json “object” + few bytes (let’s say 8) of overhead – WebSocket message “header”. On the other hand, when you do the same action on REST version of the page, HTTP Request is sent (and Response received – it does not contain anything, just status, headers and no entity):

Screen Shot 2015-03-11 at 16.44.18

I don’t even want to know what is the overhead in this case…

You might say that it does not matter, since we are not connected using 33.6 kbps modems.. well, that’s true, but every byte introduces some delay, additional handling in the network and even app servers – they have to read/write that byte, even when it won’t be used. And that’s not everything related to the resources utilisation – imagine, that every such Request created new TCP connection to the server (which does not need to be true when HTTP keep-alive is used). Since I’m still on localhost, I wanted to push the bar little higher and wrote simple performance test: method, which will create 10k updates of single map entry (you can execute it by clicking on [PerfTest] button on the sample page – both Rest and WebSocket version have it). Starting with WebSocket – it does what is expected and I can measure how long it takes to have some comparison with REST.. but.. the problem is that the REST version does not even finish. I did not dig into that that much, but seems like every browser has some kind of limit for JavaScript requests. I was usually able to achieve something around 3-5k Requests, but after that, browser “run out of resources”, most likely to protect itself or the target site from potentially DDoS-y case.

Conclusion? If you need truly bi-directional communication, with high frequency of shorter messages (it is far easier to handle those in JavaScript than big ones), you should consider using WebSocket. On the other side, if you have some application, which already works well and uses REST efficiently, you don’t need to change that! REST is still the best solution for lots of use cases, you should look at these two technologies/protocols as complements, not as competitors.

And please don’t forget, that this was not by any means an attempt for unbiased comparison :-) Any comments/feedback is appreciated!


Thursday Apr 21, 2011

Jersey Client - Apache HTTP Client 4.x integration

Since this week, Jersey users can benefit from Apache HTTP Client 4.x integration. "jersey-apache-http-client4" module has been finalized and provides similar functionality as "jersey-apache-http-client", but make sure you read javadoc - most of settings have changes because of major change in Apache HTTP Client API.

How to start with Jersey and Apache HTTP Client 4.x?

Add following dependency:

    <version>1.7-SNAPSHOT</version> <!-- or 1.7-ea04 and newer, when available -->

And you can instantiate new client for example like described in Javadoc: package com.sun.jersey.client.apache4.


Pavel Bucek-Oracle


« October 2015