Computation Graph

A Computation Graph is an in-process, deterministic, event-driven DAG. Unlike a Workflow (where the task is the unit and state is persisted), a computation graph runs as a single traversal in memory, in response to an event. It is built from Nodes and is fired either by a Reactor or invoked directly from a workflow task.

Mental model

A graph is a set of nodes wired by a topology (a -> b).
Entry nodes read named sources from an InputCache; processing nodes take their upstream node’s output; the terminal node (no downstream) produces the result.
It compiles to a callable function (<name>_compiled in Rust) that the reactor invokes; the Boundary events an Accumulator emits are what populate the cache.

Interfaces

A two-node pricing graph triggered by a reactor:

Rust

The #[computation_graph] macro declares the topology and generates the compiled function; each pub async fn in the module is a node:

#[cloacina_macros::computation_graph(
    trigger = reactor("pricing_reactor"),
    graph = {
        ingest(orderbook) -> format_output,
    }
)]
pub mod pricing {
    use super::*;
    pub async fn ingest(orderbook: Option<&OrderBookSnapshot>) -> SpreadSignal { /* ... */ }
    pub async fn format_output(input: &SpreadSignal) -> FormattedOutput { /* ... */ }
}
// generates pricing_compiled(&InputCache) -> GraphResult

Python

ComputationGraphBuilder is a context manager; nodes are @cloaca.node functions inside it, and the topology is a dict:

import cloaca

@cloaca.reactor(name="pricing_reactor", accumulators=["orderbook"], mode="when_any")
class PricingReactor:
    pass

with cloaca.ComputationGraphBuilder(
    "pricing", reactor=PricingReactor,
    graph={
        "ingest": {"inputs": ["orderbook"], "next": "format_output"},
        "format_output": {},   # no `next` → terminal
    },
) as g:
    @cloaca.node
    def ingest(orderbook):
        return spread_signal(orderbook)

    @cloaca.node
    def format_output(ingest):
        return formatted(ingest)

reactor= takes a @cloaca.reactor-decorated class, and each graph entry is a dict (inputs list, plus next / routes, or empty for terminal). See the topology dict schema for the full format. Python node functions receive owned values (the boundary copies them) where Rust takes references.

Key facts

Reactor-triggered or trigger-less. A graph either subscribes to a reactor (trigger = reactor("…")) or is trigger-less and invoked inline from a workflow task.
Data is serialized. Values that cross node boundaries implement Serialize/Deserialize; the cache stores bytes.
Deterministic. Given the same cache, the traversal is deterministic.

Build one

Embed it → Embed · Tutorials
Ship it to a server → Run the Service · Tutorials

Computation Graph

Computation Graph

Mental model

Interfaces

Key facts

Build one

See also