Rationale for Orchestrator
We may have scripts that run asynchrounsly, depend on human input or depend on other scripts to complete. We want to be able to implement high-level workflows of rhai scripts.
Design
Direct Acyclic Graphs (DAGs) are a natural fit for representing workflows.
Requirements
- Uses Direct Acyclic Graphs (DAGs) to represent workflows.
- Each step in the workflow defines the script to execute, the inputs to pass to it, and the outputs to expect from it.
- Simplicity: the output cases are binary (success or failure), and params inputted / outputted are simple key-value pairs.
- Multiple steps can depend on the same step.
- Scripts are executed using RhaiDispatcher.
Architecture
The Orchestrator is a simple DAG-based workflow execution system that extends the heromodels flow structures to support workflows with dependencies and distributed script execution.
Core Component
graph TB
subgraph "Orchestrator"
O[Orchestrator] --> RE[RhaiExecutor Trait]
O --> DB[(Database)]
end
subgraph "Executor Implementations"
RE --> RD[RhaiDispatcher]
RE --> WS[WebSocketClient]
RE --> HTTP[HttpClient]
RE --> LOCAL[LocalExecutor]
end
subgraph "Data Models (heromodels)"
F[Flow] --> FS[FlowStep]
FS --> SR[SignatureRequirement]
end
subgraph "Infrastructure"
RD --> RQ[Redis Queues]
RD --> W[Workers]
WS --> WSS[WebSocket Server]
HTTP --> API[REST API]
end
Execution Abstraction
The orchestrator uses a trait-based approach for script execution, allowing different execution backends:
RhaiExecutor Trait
use rhai_dispatcher::{PlayRequestBuilder, RhaiTaskDetails, RhaiDispatcherError};
#[async_trait]
pub trait RhaiExecutor {
async fn call(&self, request: PlayRequestBuilder<'_>) -> Result<RhaiTaskDetails, RhaiDispatcherError>;
}
Executor Implementations
RhaiDispatcher Implementation:
pub struct DispatcherExecutor {
dispatcher: RhaiDispatcher,
}
#[async_trait]
impl RhaiExecutor for DispatcherExecutor {
async fn call(&self, request: PlayRequestBuilder<'_>) -> Result<RhaiTaskDetails, RhaiDispatcherError> {
// Use RhaiDispatcher to execute script via Redis queues
request.await_response().await
}
}
WebSocket Client Implementation:
pub struct WebSocketExecutor {
ws_client: WebSocketClient,
endpoint: String,
}
#[async_trait]
impl RhaiExecutor for WebSocketExecutor {
async fn call(&self, request: PlayRequestBuilder<'_>) -> Result<RhaiTaskDetails, RhaiDispatcherError> {
// Build the PlayRequest and send via WebSocket
let play_request = request.build()?;
// Send script execution request via WebSocket
let ws_message = serde_json::to_string(&play_request)?;
self.ws_client.send(ws_message).await?;
// Wait for response and convert to RhaiTaskDetails
let response = self.ws_client.receive().await?;
serde_json::from_str(&response).map_err(RhaiDispatcherError::from)
}
}
HTTP Client Implementation:
pub struct HttpExecutor {
http_client: reqwest::Client,
base_url: String,
}
#[async_trait]
impl RhaiExecutor for HttpExecutor {
async fn call(&self, request: PlayRequestBuilder<'_>) -> Result<RhaiTaskDetails, RhaiDispatcherError> {
// Build the PlayRequest and send via HTTP
let play_request = request.build()?;
// Send script execution request via HTTP API
let response = self.http_client
.post(&format!("{}/execute", self.base_url))
.json(&play_request)
.send()
.await?;
response.json().await.map_err(RhaiDispatcherError::from)
}
}
Local Executor Implementation:
pub struct LocalExecutor {
engine: Engine,
}
#[async_trait]
impl RhaiExecutor for LocalExecutor {
async fn call(&self, request: PlayRequestBuilder<'_>) -> Result<RhaiTaskDetails, RhaiDispatcherError> {
// Build the PlayRequest and execute locally
let play_request = request.build()?;
// Execute script directly in local Rhai engine
let result = self.engine.eval::<String>(&play_request.script);
// Convert to RhaiTaskDetails format
let task_details = RhaiTaskDetails {
task_id: play_request.id,
script: play_request.script,
status: if result.is_ok() { "completed".to_string() } else { "error".to_string() },
output: result.ok(),
error: result.err().map(|e| e.to_string()),
created_at: chrono::Utc::now(),
updated_at: chrono::Utc::now(),
caller_id: "local".to_string(),
context_id: play_request.context_id,
worker_id: "local".to_string(),
};
Ok(task_details)
}
}
Data Model Extensions
Simple extensions to the existing heromodels flow structures:
Enhanced FlowStep Model
// Extends heromodels::models::flow::FlowStep
pub struct FlowStep {
// ... existing heromodels::models::flow::FlowStep fields
pub script: String, // Rhai script to execute
pub depends_on: Vec<u32>, // IDs of steps this step depends on
pub context_id: String, // Execution context (circle)
pub inputs: HashMap<String, String>, // Input parameters
pub outputs: HashMap<String, String>, // Output results
}
Execution Flow
sequenceDiagram
participant Client as Client
participant O as Orchestrator
participant RE as RhaiExecutor
participant DB as Database
Client->>O: Submit Flow
O->>DB: Store flow and steps
O->>O: Find steps with no dependencies
loop Until all steps complete
O->>RE: Execute ready steps
RE-->>O: Return results
O->>DB: Update step status
O->>O: Find newly ready steps
end
O->>Client: Flow completed
Flexible Orchestrator Implementation
use rhai_dispatcher::{RhaiDispatcher, PlayRequestBuilder};
use std::collections::HashSet;
pub struct Orchestrator<E: RhaiExecutor> {
executor: E,
database: Arc<Database>,
}
impl<E: RhaiExecutor> Orchestrator<E> {
pub fn new(executor: E, database: Arc<Database>) -> Self {
Self { executor, database }
}
pub async fn execute_flow(&self, flow: Flow) -> Result<(), OrchestratorError> {
// 1. Store flow in database
self.database.collection::<Flow>()?.set(&flow)?;
// 2. Find steps with no dependencies (depends_on is empty)
let mut pending_steps: Vec<FlowStep> = flow.steps.clone();
let mut completed_steps: HashSet<u32> = HashSet::new();
while !pending_steps.is_empty() {
// Find ready steps (all dependencies completed)
let ready_steps: Vec<FlowStep> = pending_steps
.iter()
.filter(|step| {
step.depends_on.iter().all(|dep_id| completed_steps.contains(dep_id))
})
.cloned()
.collect();
if ready_steps.is_empty() {
return Err(OrchestratorError::NoReadySteps);
}
// Execute ready steps concurrently
let mut tasks = Vec::new();
for step in ready_steps {
let executor = &self.executor;
let task = async move {
// Create PlayRequestBuilder for this step
let request = RhaiDispatcher::new_play_request()
.script(&step.script)
.context_id(&step.context_id)
.worker_id(&step.worker_id);
// Execute via the trait
let result = executor.call(request).await?;
Ok((step.base_data.id, result))
};
tasks.push(task);
}
// Wait for all ready steps to complete
let results = futures::future::try_join_all(tasks).await?;
// Update step status and mark as completed
for (step_id, task_details) in results {
if task_details.status == "completed" {
completed_steps.insert(step_id);
// Update step status in database
// self.update_step_status(step_id, "completed", task_details.output).await?;
} else {
return Err(OrchestratorError::StepFailed(step_id, task_details.error));
}
}
// Remove completed steps from pending
pending_steps.retain(|step| !completed_steps.contains(&step.base_data.id));
}
Ok(())
}
pub async fn get_flow_status(&self, flow_id: u32) -> Result<FlowStatus, OrchestratorError> {
// Return current status of flow and all its steps
let flow = self.database.collection::<Flow>()?.get(flow_id)?;
// Implementation would check step statuses and return overall flow status
Ok(FlowStatus::Running) // Placeholder
}
}
pub enum OrchestratorError {
DatabaseError(String),
ExecutorError(RhaiDispatcherError),
NoReadySteps,
StepFailed(u32, Option<String>),
}
pub enum FlowStatus {
Pending,
Running,
Completed,
Failed,
}
// Usage examples:
// let orchestrator = Orchestrator::new(DispatcherExecutor::new(dispatcher), db);
// let orchestrator = Orchestrator::new(WebSocketExecutor::new(ws_client), db);
// let orchestrator = Orchestrator::new(HttpExecutor::new(http_client), db);
// let orchestrator = Orchestrator::new(LocalExecutor::new(engine), db);
Key Features
- DAG Validation: Ensures no circular dependencies exist in the
depends_on
relationships - Parallel Execution: Executes independent steps concurrently via multiple workers
- Simple Dependencies: Each step lists the step IDs it depends on
- RhaiDispatcher Integration: Uses existing dispatcher for script execution
- Binary Outcomes: Steps either succeed or fail (keeping it simple as per requirements)
This simple architecture provides DAG-based workflow execution while leveraging the existing rhailib infrastructure and keeping complexity minimal.