Suppose TCP proxy has forwarded request back to the backend server. When it receives reply from the backend server, how does it knows which client to reply. What exact session information does a proxy stores?
Can anyone please throw some light on this
It depends on the protocol, it depends on the proxy, and it depends on whether transparency is a goal. Addressing all of these points exhaustively would take forever, so let's consider a simplistic case.
A network connection in software is usually represented by some sort of handle (whether that's a file descriptor or some other resource). In a C program on a POSIX system, we could simply keep two file descriptors associated with each other:
struct proxy_session {
int client_fd;
int server_fd;
}
This is the bare-minimum requirement.
When a client connects, we allocate one of these structures. There may be a protocol that lets us know what backend we should use, or we may be doing load balancing and picking backends ourselves.
Once we've picked a backend (either by virtue of having parsed the protocol or through having made some form routing decision), we initiate a connection to it. Simplistically, a proxy (as an intermediary) simply forwards packets between a client and a server.
We can use any number of interfaces for tying these two things together. On Linux, for example, epoll(2) allows us to associate a pointer to events on a file descriptor. We can provide it a pointer to our proxy_session structure for both the client and server side. When data comes in either of those file descriptors, we know where to map it.
Lacking such an interface, we necessarily have a means for differentiating connection handles (whether they're file descriptors, pointers, or some other representation). We could then use a structure like a hash table to look up the destination for a handle. The solution is found simply by being able to differentiate connections to each other, and holding some state that "glues" two connections together.