How does WebSocket compress messages?

You could use Unishox compression technique for compressing text sent over Websockets. However this is not integrated with the spec and the implementor is responsible for compressing one one side and decompressing on the other side.

Unishox achieves compression by assigning fixed prefix-free codes for each letter in the given Character Set based on known frequency of popular symbols (entropy coding). It also encodes repeating letter sets separately (dictionary coding). For Unicode characters, delta coding is used. More information is available in this article.

So far it has been implemented in C and Javascript.

Disclaimer: I am the developer of Unishox.

WebSocket is, at its heart, just a set of framing for TEXT or BINARY data.

It performs no compression on its own.

However, the WebSocket spec allows for Extensions, and there have been a variety of compression extensions in the wild (the formalized specs for one of these is finalized).

As of today (August 2018) the accepted compression spec is permessage-deflate.

Some of the extensions seen in the wild:

• permessage-deflate - the name of the formalized spec for using deflate to compress entire messages, regardless of the number of websocket frames.
• x-webkit-deflate-frame - an early proposed compression that compresses each raw websocket data frame. Seen in use by Chrome and Safari. (now deprecated in Chrome and Safari)
• perframe-deflate - a renamed version of the above compression. Seen in use by various websocket server implementations, and also briefly showed up in various WebKit based clients. (Completely deprecated in modern browsers, but does still show up in various WebSocket client libraries)

Of note, the permessage-deflate extension is the first in a line of PMCE (Per-Message Compression Extensions) that will eventually include other compression schemes (ones being discussed are permessage-bzip2, permessage-lz4, and permessage-snappy)

Does websocket compress its JSONs before sending to stream?

The short answer is: sometimes, but you can't depend on it.

As stated aptly by Joakim Erdfelt, Websocket connections support both Text and Binary messages.

JSON is just one way to transport data, with the advantage of versatility and ease of use (while being wasteful as far as space is concerned).

You can easily transfer binary data using the Websocket API, eliminating bandwidth overhead at the expense of other concerns (such as endieness, word length, parsing, etc').

Many browsers also support Websocket message compression as an extension to the Websocket protocol (although the server might not support the extension).

Extensions are negotiated using the Sec-WebSocket-Extensions HTTP header. The negotiations are usually implemented by the client / server without offering a public API to control them.

Right up until 2015, there were a number of approaches and implementations in the wild, but since December 2015 RFC 7692 is the only real contender for message compression and things are much clearer.

RFC 7692 compresses the whole message before wrapping (and perhaps fragmenting) it in Websocket "packets", making it easier to implement than some previous compression schemes.

The current draft offers a permessage-foo compression negotiation scheme (where foo is the requested / supported compression).

I only experienced the permessage-deflate extension myself.

Do notice that extension negotiations are optional, meaning that potential network clients are usually allowed to negotiate connections without compression, even if your server supports the extension.

Moreover, RFC 7692 supports selective compression, meaning some messages might be compressed while others aren't compressed...

... for example, [123456789,123456789] might be sent as is, since it's length indicates it isn't likely to be worth the compression effort.

Support for permessage-deflate (RFC 7692), August 8th, 2017:

This is a combination of the information in the comments, last updated August 8th, 2017.

If I missed anything, add them here and update the date.

Known Browser Support

• Chrome since 32
• Mozilla Firefox since 37
• Safari - unsupported (10.1.2 seems to use x-webkit-deflate-frame)
• Microsoft Edge - unsupported as of version 40

Known Server Support

• Crossbar.io (Autobahn)
• wspy
• Tornado
• Faye on node.js and Ruby
• SwiftWebSocket (read only?)
• WS-RS (Rust)