Currently Google-Engineers plan moving towards HTTP/3. As HTTP/2 still is not used commonly worldwide and major implementation flaws and performance issues exist we try to build a much simpler and secure solution: HTTP/1.2.
It will provide the following features:
Also the part "HyperText" or even a complete renaming could be thought over. Its representation is by far not more appropriate these days.
We will publish detailed RFP / RFC in the near future on http://der-it-pruefer.de.
See BUILD.md.
HTTP/3 should be implemented using UDP or even UDPLite. Phew, also it should be HTTP downwards compatible.
How could this be achieved by switching from a Connection-Oriented (TCP) to a much much simpler Connection-Less-Protocol (UDP)? Maybe engineering department has contracts with Harry Potter?
Take a look at the following topics:
All logic which has been removed from transparent, 100% working Layer3 protocols must now be re-implemented "By Hand" including packet retransmission into Layer7. Phew: every Browser, every Server, every HTTP-related program has to implement this by themselves.
Much fun implementing! See more details on http://der-it-pruefer.de.
HTTP/2 adds too much complexity. HTTP/1.1 simplicity has been lost, especially OpenSource libraries tend to be confusing and ununderstandable even using libevent.
Its evertything packed in one really big box. Logical separation? Security? All going unfiltered over a single TCP port. Ah, just use your IPS correctly! Bam: IPS process killed by 0day inside TCP packet.
Also adding TLS / SSL handling inside the protocol is not appropriate anymore. Let this handle a separated component like nginx (reverse proxy), ingress-nginx on Kubernetes, stunnel or a Hardware-Loadbalancer.
Do not duplicate things you are not familiar with especially when other products exist which do the same for decades much smarter.
HTTP/2 fixes the HTTP/1.1 Pipelined Connections problem. Due to a bug in asynchronously handling the responses (wrong Request / Reply order) the complete feature was removed from ALL browsers worldwide.
Our HTTP/1.2 implementation also fixes the problem by extending the HTTP/1.1 protocol by just some lines of code. A unique identifier (UUID) will be added for each Request so Response-Ordering will be obsolete.
With fixing the Pipelined Connections problem, HTTP/2 should also fix speeding up dynamic WebServices Request / Response times. In theory the HTTP/2 specs look quite good, real live shows different behaviour.
This could be caused by:
The provided Proof Of Concept Code includes the working "FalconAS" HTTP/1.1 and HTTP/1.2 Server including those components:
The Proxy-Server component has been engineered and will be published later on.
Detailed draft and specifications can be found under /draft and /specs subdirs.
Currently Coroutines is the word of interest in the Coding-Community. Libevent uses Coroutines. Multiple Server Products use Libevent for client connection handling.
But beware: Coroutines could be contraproductive for scaling aspects. If you are not familiar with Kernel Socket Programming and implement the socket in a blocking way: Each client filedescriptor will be monitored by the kernel separetely. Welcome to the 90s.
Concrete: the main loop in user space iterates over ALL current connection filedescriptors. If we have 10000 active or sleeping Keep-Alive connections: 10000 syscalls (context switches between user and kernel space) will be done every main loop iteration. The kernel will additionally loop on 10000 filedescriptors checking if there is waiting incoming data. This will kill the kernels responsiveness.
Unlike (also nginX) using epoll (also we use epoll) a single syscall will tell user space which filedescriptors have waiting data. With optimized 64bit server CPU and good server code: Scaling, bingo!
In times of Flatpak and Snap a desktop-application will be delivered in one single packed FILE, e.g. application1.flatpak. Also this has been widely used in Debian Linux as packaging system since 1996.
The common problems of HTTP, especially what HTTP/3 should solve: a huge amount of single requests served with less overhead to a CCDN (Cloud Content Distribution Network).
But think much more simple! The following workflow demonstrates: if you pack your browser-application (or subsections) you only have to transfer 1 single file 1 time and again when the application changes. A 304 not-modified request will keep your static application-server farm or CCDN idle.
Also with modern CPU and 100Mbit+ DSL uplink 1. request 2. response 3. depacking 4. app rendering should be < 1 second if not too much media data included. We will provide Proof Of Concept in the near future.
Somehow big Images or Videos still can be referenced dynamically as a link and still the technique is practicable.
Min:Sec
00:00 Client -------------- HTTP/1.2 GET /app1.tar.bz2 -------------------> Server
00:00 Client <------------- HTTP/1.2 app1.tar.bz2 ------------------------- Server
00:01 Client ::depack()
00:01 Client ::render()Min:Sec
00:00 Client -------------- HTTP/1.2 GET /app1.tar.bz2 -------------------> Server
00:00 Client <------------- HTTP/1.2 304 not modified --------------------- Server
00:00 Client ::depack()
00:00 Client ::render()While collecting experience in HTTP/1.2 and Linux Socket Interface processing and testing at low level we came to the conclusion: The Linux Kernel / Socket Interface drastically needs a rewrite!
Detailed RFP / how to fix multiple logical processing and performance bugs will follow.