The req protocol is one half of a request/reply pattern. In this pattern, a requester sends a message to one replier, who is expected to reply. The request is resent if no reply arrives, until a reply is received or the request times out.
|This protocol is useful in setting up RPC-like services. It is also "reliable", in that a the requester will keep retrying until a reply is received.
|Because requests are resent, it is important that they be idempotent to ensure predictable and repeatable behavior even in the face of duplicated requests, which can occur (for example if a reply message is lost for some reason.)
The requester generally only has one outstanding request at a time unless
in “raw” mode (via
and it will generally attempt to spread work requests to different peer repliers.
This property, when combined with
can help provide a degree of load-balancing.
The req protocol is the requester side, and the rep protocol is the replier side.
nng_req0_open() functions create a requester socket.
This socket may be used to send messages (requests), and then to receive replies.
Generally a reply can only be received after sending a request.
(Attempts to receive a message will result in
NNG_ESTATE if there is no
Furthermore, only a single receive operation may be pending at a time.
Attempts to post more receive operations concurrently will result in
Requests may be canceled by sending a different request. This will cause the requester to discard any reply from the earlier request, but it will not stop a replier from processing a request it has already received or terminate a request that has already been placed on the wire.
Raw mode sockets ignore all these restrictions.
NNG_OPT_REQ_RESENDTIME value may be configured differently
on contexts created this way.
Each context may have at most one outstanding request, and operates independently from the others.
The restrictions for order of operations with sockets apply equally well for contexts, except that each context will be treated as if it were a separate socket.
Only version 0 of this protocol is supported. (At the time of writing, no other versions of this protocol have been defined.)
The following protocol-specific option is available.
nng_duration) When a new request is started, a timer of this duration is also started. If no reply is received before this timer expires, then the request will be resent. (Requests are also automatically resent if the peer to whom the original request was sent disconnects, or if a peer becomes available while the requester is waiting for an available peer.)
This protocol uses a backtrace in the header. This form uses a “stack” of 32-bit big-endian identifiers. There must be at least one identifier, the request ID, which will be the last element in the array, and must have the most significant bit set.
There may be additional peer IDs preceding the request ID. These will be distinguishable from the request ID by having their most significant bit clear.
When a request message is received by a forwarding node (see
nng_device()), the forwarding node prepends a
32-bit peer ID (which must have the most significant bit clear),
which is the forwarder’s way of identifying the directly connected
peer from which it received the message.
(This peer ID, except for the
most significant bit, has meaning only to the forwarding node itself.)
It may help to think of prepending a peer ID as “pushing” a peer ID onto the front of the stack of headers for the message. (It will use the peer ID it popped from the front to determine the next intermediate destination for the reply.)
When a reply message is created, it is created using the same headers that the request contained.
A forwarding node can “pop” the peer ID it originally pushed on the message, stripping it from the front of the message as it does so.
When the reply finally arrives back at the initiating requester, it should have only a single element in the message, which will be the request ID it originally used for the request.