Struct SubprocessFilter 

pub struct SubprocessFilter { /* private fields */ }

A filter that delegates to a shared PythonProcess via stdin/stdout. Multiple SubprocessFilters can share the same process (Arc<Mutex>).

Implementations

impl SubprocessFilter

pub fn new( process: Arc<Mutex<PythonProcess>>, node_id: String, trainable: bool, config_hash: CacheKey, ) -> Self

A proxy for the filter loaded under node_id in process — shared, so every sibling filter of one plan talks to the same interpreter.

pub fn fallback_config_hash(node_id: &str, pickled_filter: &[u8]) -> CacheKey

Fallback identity for payloads that carry no explicit config hash: hash the pickled filter bytes — any config change changes the pickle, so stale cache hits are still impossible (the pickle is merely less stable across environments than a real config hash).

Trait Implementations

impl Filter for SubprocessFilter

The seam.

Filter is a soma-core trait, so these three return SomaError while everything behind them is typed as a WorkerError. A subprocess that died and a payload that would not decode stay distinguishable right up to here, which is as far as the shared type can carry them.

fn config_hash(&self) -> CacheKey

Compute a hash of this filter’s configuration. Same config must always produce the same hash. Only public constructor parameters contribute.

fn fit(&self, x: &Value, y: Option<&Value>) -> Result<Value>

Learn state from training data. Returns serialized state as a Value.

fn forward(&self, x: &Value, state: &Value) -> Result<Value>

Transform data using learned state. This is the potentially differentiable operation.

fn meta(&self) -> FilterMeta

Metadata for the compiler.

fn composite_fit( &self, peers: &[(String, Arc<dyn Filter>)], x: &Value, y: Option<&Value>, ) -> Option<Result<(Value, HashMap<String, Value>)>>

Execute a composite fit across multiple filters that share an execution context. Used for differentiable filter chains where autograd must flow between them.