DevFlow/crates/df-workflow/src/executor.rs

//! DAG 执行器 — 按拓扑层级调度执行

use std::collections::HashMap;

use df_core::events::WorkflowEvent;
use df_core::types::NodeId;

use crate::dag::Dag;
use crate::eventbus::EventBus;
use crate::node::{NodeContext, NodeOutput};
use crate::state::StateMachine;

/// DAG 执行器
pub struct DagExecutor {
    /// 事件总线
    event_bus: EventBus,
    /// 节点状态机
    state_machine: StateMachine,
    /// 工作流执行 ID（由调用方传入，下沉到每个 NodeContext）
    execution_id: String,
}

impl DagExecutor {
    /// 创建执行器
    ///
    /// `execution_id` 为本次工作流执行的唯一标识，会下沉到每个节点的 NodeContext，
    /// 用于节点内的事件关联、审计追踪等。
    pub fn new(event_bus: EventBus, execution_id: String) -> Self {
        Self {
            event_bus,
            state_machine: StateMachine::new(),
            execution_id,
        }
    }

    /// 执行 DAG 工作流
    ///
    /// 同一拓扑层内的节点并发执行，层与层之间串行；
    /// 本层全部节点完成后统一更新状态，任一失败则中止后续层。
    pub async fn run(
        &mut self,
        dag: &Dag,
        initial_config: serde_json::Value,
    ) -> anyhow::Result<HashMap<NodeId, NodeOutput>> {
        let layers = dag.topological_layers()?;
        let mut outputs: HashMap<NodeId, NodeOutput> = HashMap::new();
        let start = std::time::Instant::now();

        tracing::info!("DAG 执行开始，共 {} 层", layers.len());

        for (layer_idx, layer) in layers.iter().enumerate() {
            tracing::info!("执行第 {} 层，共 {} 个节点", layer_idx, layer.len());

            // 阶段一：逐节点发 NodeStarted、置为运行中，并构建执行 future
            let mut node_futures = Vec::with_capacity(layer.len());
            for node_id in layer {
                let node = dag.nodes.get(node_id).ok_or_else(|| {
                    anyhow::anyhow!("节点 {} 不存在于 DAG 中", node_id)
                })?;

                // 发送 NodeStarted 事件
                self.event_bus
                    .send(WorkflowEvent::NodeStarted {
                        node_id: node_id.clone(),
                    })
                    .await;

                self.state_machine.set_running(node_id.clone())?;

                // 构建节点上下文
                let mut inputs = HashMap::new();
                for pred_id in dag.predecessors(node_id) {
                    if let Some(out) = outputs.get(&pred_id) {
                        inputs.insert(pred_id, out.clone());
                    }
                }

                let ctx = NodeContext {
                    node_id: node_id.clone(),
                    inputs,
                    config: initial_config.clone(),
                    execution_id: self.execution_id.clone(),
                    event_bus: self.event_bus.clone(),
                    // 共享执行器状态机，使节点（如 HumanNode）能读取真实状态而非空状态机
                    node_status: self.state_machine.clone(),
                };

                // 仅捕获节点共享引用与所有权上下文，避免与 self 借用冲突
                let id = node_id.clone();
                node_futures.push(async move {
                    let node_start = std::time::Instant::now();
                    let result = node.execute(ctx).await;
                    (id, result, node_start.elapsed().as_millis() as u64)
                });
            }

            // 阶段二：同层节点并发执行，等待全部完成
            let results = futures::future::join_all(node_futures).await;

            // 阶段三：统一更新状态并发送事件，任一失败则整体返回 Err
            let mut first_err: Option<anyhow::Error> = None;
            for (node_id, result, duration) in results {
                match result {
                    Ok(output) => {
                        self.state_machine.set_completed(node_id.clone())?;
                        self.event_bus
                            .send(WorkflowEvent::NodeCompleted {
                                node_id: node_id.clone(),
                                duration_ms: duration,
                            })
                            .await;
                        outputs.insert(node_id, output);
                    }
                    Err(e) => {
                        self.state_machine.set_failed(node_id.clone())?;
                        self.event_bus
                            .send(WorkflowEvent::NodeFailed {
                                node_id: node_id.clone(),
                                error: e.to_string(),
                            })
                            .await;
                        // 同层多个失败时只报告第一个
                        if first_err.is_none() {
                            first_err = Some(e.context(format!("节点 {} 执行失败", node_id)));
                        }
                    }
                }
            }

            if let Some(e) = first_err {
                return Err(e);
            }
        }

        let total = start.elapsed().as_millis() as u64;
        self.event_bus
            .send(WorkflowEvent::WorkflowCompleted {
                total_duration_ms: total,
            })
            .await;

        tracing::info!("DAG 执行完成，耗时 {}ms", total);
        Ok(outputs)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::node::{Node, NodeResult, NodeSchema};
    use async_trait::async_trait;
    use std::time::Duration;

    /// 测试节点：sleep 指定毫秒后返回空输出
    struct SleepNode {
        sleep_ms: u64,
    }

    #[async_trait]
    impl Node for SleepNode {
        async fn execute(&self, _ctx: NodeContext) -> NodeResult {
            tokio::time::sleep(Duration::from_millis(self.sleep_ms)).await;
            Ok(NodeOutput::empty())
        }

        fn schema(&self) -> NodeSchema {
            NodeSchema {
                params: serde_json::Value::Null,
                output: serde_json::Value::Null,
            }
        }

        fn node_type(&self) -> &str {
            "sleep"
        }
    }

    /// 测试节点：直接返回错误
    struct FailNode;

    #[async_trait]
    impl Node for FailNode {
        async fn execute(&self, _ctx: NodeContext) -> NodeResult {
            Err(anyhow::anyhow!("故意失败"))
        }

        fn schema(&self) -> NodeSchema {
            NodeSchema {
                params: serde_json::Value::Null,
                output: serde_json::Value::Null,
            }
        }

        fn node_type(&self) -> &str {
            "fail"
        }
    }

    #[tokio::test]
    async fn test_same_layer_runs_in_parallel() {
        // 两个无依赖节点位于同一层，各 sleep 100ms
        let mut dag = Dag::new();
        dag.add_node("a".to_string(), Box::new(SleepNode { sleep_ms: 100 }));
        dag.add_node("b".to_string(), Box::new(SleepNode { sleep_ms: 100 }));

        let mut executor = DagExecutor::new(EventBus::new(), "test-exec".to_string());
        let start = std::time::Instant::now();
        let outputs = executor.run(&dag, serde_json::Value::Null).await.unwrap();
        let elapsed = start.elapsed();

        assert_eq!(outputs.len(), 2);
        // 串行需要约 200ms，并行应明显小于 180ms
        assert!(
            elapsed < Duration::from_millis(180),
            "同层节点应并行执行，实际耗时 {:?}",
            elapsed
        );
    }

    #[tokio::test]
    async fn test_layer_failure_aborts_following_layers() {
        // a(失败) 与 b(成功) 同层，c 依赖 a，失败后 c 不应执行
        let mut dag = Dag::new();
        dag.add_node("a".to_string(), Box::new(FailNode));
        dag.add_node("b".to_string(), Box::new(SleepNode { sleep_ms: 10 }));
        dag.add_node("c".to_string(), Box::new(SleepNode { sleep_ms: 10 }));
        dag.add_edge("a".to_string(), "c".to_string());

        let mut executor = DagExecutor::new(EventBus::new(), "test-exec".to_string());
        let err = executor
            .run(&dag, serde_json::Value::Null)
            .await
            .unwrap_err();
        assert!(err.to_string().contains("节点 a 执行失败"));

        // 同层成功节点状态正常更新，下游节点保持 Pending
        use df_core::types::NodeStatus;
        assert_eq!(executor.state_machine.get(&"a".to_string()), NodeStatus::Failed);
        assert_eq!(executor.state_machine.get(&"b".to_string()), NodeStatus::Completed);
        assert_eq!(executor.state_machine.get(&"c".to_string()), NodeStatus::Pending);
    }
}