Random latency between headers and body of POST - ios

Question

We're using NSUrlConnection sendAsynchronousRequest to send simple POST requests to our Node JS server. Through analyzing tcpdumps, we've noticed that sometimes the request headers and request body are split into 2 separate TCP packets.
NSMutableURLRequest *request = [[NSMutableURLRequest alloc] init];
[request setURL:[NSURL URLWithString:url]];
[request setTimeoutInterval:3];
[request setHTTPMethod:#"POST"];
[request setValue:postLength forHTTPHeaderField:#"Content-Length"];
[request setValue:#"application/json" forHTTPHeaderField:#"Content-type"];
[request setHTTPBody:postData];
[NSURLConnection sendAsynchronousRequest:request queue:[NSOperationQueue mainQueue] completionHandler: ^(NSURLResponse *response, NSData *POSTReply, NSError *error) { }];
The problem is that occasionally the headers are sent to the server, which opens a connection to our API, and then multiple seconds later the body packet is sent. We're seeing a > 1 second delay between headers and body, randomly every hundred requests on the server side. It's the single largest source of latency on our API.
For the majority of requests, headers and body are about the same size (200 bytes each).
Has anyone seen this before?

Answer 1

I don't have concrete insights on the swift APIs (NSUrlConnection etc.) but in general, while the HTTP data is sent in chunks of large size (which can accommodate the 200 bytes in one packet), more granular decisions are taken at the underlying TCP level, which may decide to split these into chunks of lesser sizes.
I have observed this in node (sending side as well as receiving side) on specific platforms (such as AIX) where the packets are split at illogical boundaries by TCP. By TCP specification, applications should not rely on specific order or size of the low level data transport, it guarantees the data integrity at the final stage.
One suspect would be the presence of Nagile's algorithm in the TCP.
I would also recommend to change the client to another platform (such as Linux) and compare the behavior, if it is easy to port the code.
Hope this helps.