Trait Transport 

pub trait Transport: Send + Sync {
    // Required methods
    fn execute(
        &self,
        plan: &ExecutionPlan,
        filters: &NodeCatalog,
        input: &Value,
        mode: &RunMode,
        seed: Option<i64>,
    ) -> Result<(Value, HashMap<String, Value>)>;
    fn get_state(&self, node_ids: &[String]) -> Result<HashMap<String, Value>>;
    fn set_state(&self, states: &HashMap<String, Value>) -> Result<()>;
    fn get_gradients(
        &self,
        node_ids: &[String],
    ) -> Result<HashMap<String, Value>>;
    fn apply_gradients(&self, gradients: &HashMap<String, Value>) -> Result<()>;

    // Provided method
    fn execute_node(
        &self,
        node_id: &str,
        input: Option<&Value>,
    ) -> Result<Value> { ... }
}

Abstraction for communicating with remote workers. Implemented by WsTransport (WebSocket), but could be HTTP, gRPC, etc.

Required Methods

fn execute( &self, plan: &ExecutionPlan, filters: &NodeCatalog, input: &Value, mode: &RunMode, seed: Option<i64>, ) -> Result<(Value, HashMap<String, Value>)>

Send a plan for execution and receive the output + trained states.

mode says what to do with the nodes, and carries the labels when there are any. It replaced a fit_mode: bool sitting beside an y: Option<&Value> — a flag selecting between two operations with differently shaped results, and a parameter that meant nothing unless the flag was set. It is the same RunMode the local executor reads, so the two paths cannot disagree about what a fit is.

seed is the run’s experiment seed, and it is a parameter rather than something the transport digs out because the transport has no RunContext to dig in. Without it the worker salts nothing, and a five-seed sweep run remotely shares one cache line across all five — the worker protocol’s SerializedPlan::seed documents that as the bug it exists to close, and this path was still passing None.

fn get_state(&self, node_ids: &[String]) -> Result<HashMap<String, Value>>

Request trained states from the remote worker.

fn set_state(&self, states: &HashMap<String, Value>) -> Result<()>

Load states on the remote worker.

fn get_gradients(&self, node_ids: &[String]) -> Result<HashMap<String, Value>>

Request gradients from the remote worker.

fn apply_gradients(&self, gradients: &HashMap<String, Value>) -> Result<()>

Apply aggregated gradients on the remote worker.

Provided Methods

fn execute_node(&self, node_id: &str, input: Option<&Value>) -> Result<Value>

Convenience: execute a single node remotely (used by the plan executor).

Unseeded, and it has to be: this takes a node id and nothing else, so there is no run to take a seed from. Callers that have a RunContext should go through Transport::execute with ctx.seed instead of reaching for this.

Implementors