[−][src]Module xaynet::client
Provides client-side functionality to connect to a XayNet service.
This functionality includes:
- Abiding by (the underlying
Participant
's side of) the PET protocol. - Handling the network communication with the XayNet service, including polling of service data.
Participant
In any given round of federated learning, each Participant
of the
protocol is characterised by a role which determines its Task
to carry
out in the round, and which is computed by check_task
.
Participants selected to Update
are responsible for sending masked model
updates in the form of PET messages constructed with
compose_update_message
.
Participants selected to Sum
are responsible for sending ephemeral keys
and global masks in PET messages constructed respectively with
compose_sum_message
and compose_sum2_message
.
Client
A Client
has an intentionally simple API - the idea is that it is
initialised with some settings, and then start()
ed. Currently for
simplicity, clients that have started running will do so indefinitely. It is
therefore the user's responsibility to terminate clients that are no longer
needed. Alternatively, it may be more convenient to run just a single round
(or a known fixed number of rounds). In this case, use during_round()
.
For examples of usage, see the test-drive
scripts.
Note. At present, the Client
implementation is somewhat tightly
coupled with the workings of the C-API SDK, but this may well change in a
future version to be more independently reusable.
Requests via Proxy
There is a Proxy
which a Client
can use to communicate with the
service. To summarise, the proxy:
- Wraps either an in-memory service (for local comms) or a client request object (for remote comms over HTTP).
- In the latter case, deals with logging and wrapping of network errors.
- Deals with deserialization
The client request object is responsible for building the HTTP request and extracting the response body. As an example:
async fn get_sums(&self) -> Result<Option<bytes::Bytes>, reqwest::Error>
issues a GET request for the sum dictionary. The return type reflects the
presence of networking Error
s, but also the situation where the dictionary
is simply just not yet available on the service. That is, the type also
reflects the optionality of the data availability.
Proxy
essentially takes this (deserializing the Bytes
into a SumDict
while handling Error
s into ClientError
s) to expose the overall method
async fn get_sums(&self) -> Result<Option<SumDict>, ClientError>
Structs
Client | A client of the federated learning service |
Participant | A participant in the PET protocol layer. |
Enums
ClientError | Client-side errors |
Proxy | Proxy for communicating with the service. |
Task | Tasks of a participant. |