Semantics of Remote Procedure Calls

Ideally, Remote Procedure Calls look like local procedure calls and the application programs may be unaware of the existence/need for the network. Like everything else, they suffer from network crashes, lost messages and delays.

Consider what happens when a server crashes. The client stub may :

• Block and await the reply (which will never come).
• Time-out and report a failure, or exception, to the client.
• Time-out and retransmit the request.

Should clients re-issue their requests in the event of a failure? Moreover, should the client's application (manually) or the client's stubs (automatically) re-issue a request?

It is important to understand whether a package supports at-most-once or at-least-once semantics. Remote operations, which may be repeated without consequence, are termed idempotent operations.