An illustrated browser window containing the curly braces ({}) used in JSON

Use Streaming JSON To Speed Up Your Website


Modern websites are complex and often need to load data in the background in an asynchronous way. JSON streaming is a way to make this happen.

What is JSON streaming?

JSON streaming–or JSONS–is a simple technique we use to decrease search latency for users, particularly anyone on a slow connection. You'll see the most improvement for mobile users on HTTPS connections, but it really benefits everyone. When you’re working with JSONS, you don't need complicated JSON parsers–just send line-delimited JSON. The Twitter streaming API has worked like this for years.

Why You Should Use a JSON Stream

Our social username checker offers a simple example of how this works . In it, we check to see if your search is available on Facebook, Twitter and Pinterest by doing an HTTP HEAD request for the name at each site. Theresults come in random order, and we want to show them as fast as possible. We couldwork around this by sending three HTTPS requests, but that means three TCP connections, three SSL handshakes, and so on. We can avoid this–and probably recycle an existing HTTPS connection–if we just send one request:

The streaming response is analogous to—and relies on—HTTP chunked transfer encoding. The best way to see how this works is to trace the connection with curl:

curl --trace - ""
<= Recv data, 45 bytes (0x2d)
0000: 32 37 0d 0a 7b 22 54 79 70 65 22 3a 22 54 77 69 27..{"Type":"Twi
0010: 74 74 65 72 22 2c 22 49 73 52 65 67 69 73 74 65 tter","IsRegiste
0020: 72 65 64 22 3a 74 72 75 65 7d 0a 0d 0a red":true}...
<= Recv data, 46 bytes (0x2e)
0000: 32 38 0d 0a 7b 22 54 79 70 65 22 3a 22 46 61 63 28..{"Type":"Fac
0010: 65 62 6f 6f 6b 22 2c 22 49 73 52 65 67 69 73 74 ebook","IsRegist
0020: 65 72 65 64 22 3a 74 72 75 65 7d 0a 0d 0a ered":true}...
<= Recv data, 52 bytes (0x34)
0000: 32 39 0d 0a 7b 22 54 79 70 65 22 3a 22 50 69 6e 29..{"Type":"Pin
0010: 74 65 72 65 73 74 22 2c 22 49 73 52 65 67 69 73 terest","IsRegis
0020: 74 65 72 65 64 22 3a 74 72 75 65 7d 0a 0d 0a 30 tered":true}...0
0030: 0d 0a 0d 0a ....
== Info: Curl_http_done: called premature == 0
== Info: Connection #0 to host left intact

You may encounter nerd dogma that each endpoint should return one and only one object. Or that pretty-printed JSON may contain newlines, and what you really need is a streaming parser. Ignore them. Parsing responses is as simple as splitting on newline. The jsonpipe library is a good starting point for your client code.

How To See the Difference JSON Makes

A great way to see what this means for a mobile user is to use rvictl on a Mac to observe how a real cell radio on a real cell network actually talks to your server. One HTTPS connection on my 2-bar Verizon 3G connection looks like this—the first column is time in seconds:

0.000000 DNS 69 Standard query 0x5622 A
2.573090 DNS 85 Standard query response 0x5622 A A

Wait, the DNS query took 2.5 seconds to resolve? Wow. Maybe the 3G modem was in sleep mode. We can finally establish the TCP connection:

2.573822 TCP 64 57709443 [SYN] Seq=0 Win=65535 Len=0 MSS=1388 WS=64 TSval=727045548 TSecr=0 SACK_PERM=1
2.574404 TCP 60 44357709 [SYN, ACK] Seq=0 Ack=1 Win=42540 Len=0 MSS=1430 SACK_PERM=1 TSval=2852312183 TSecr=727045548 WS=128

And begin the SSL/TLS dance:

2.575790 TLSv1.2 285 Client Hello
3.582868 TLSv1.2 1428 [TCP Fast Retransmission] Server Hello
3.585061 TLSv1.2 127 Client Key Exchange
3.585771 TLSv1.2 58 Change Cipher Spec
3.586415 TLSv1.2 97 Encrypted Handshake Message
3.587729 TLSv1.2 103 Change Cipher Spec, Hello Request, Hello Request

Hello! From here, 3.5 seconds later, we’re finally sending/receiving encrypted application data.

JSON Reduces Packets To Increase Speed

Each keystroke on Instant Domain Search queries several data structures to get availability results, marketplace results, suggestions, and so on. We noticed huge performance gains for users once we stopped trying to establish so many HTTPS connections on each keystroke and instead streamed the results back on one connection per query. You can see how that translates into far fewer packets jamming the user’s network connection. For the record, we did try using a web socket but found that per-query HTTPS connections were the most reliable way to deliver fast results to our users.