faucet_cli/serve/

server.rs

//! axum router assembly and the bind / graceful-shutdown serve loop.

use crate::error::{CliError, CliResult};
use crate::serve::config::ServeConfig;
use crate::serve::handlers::{audit, dlq, doctor, health, logs, runs, schemas};
use crate::serve::history::RunHistory;
use crate::serve::state::ServerState;
use crate::serve::{auth, metrics};
use axum::Router;
use axum::routing::{get, post};
use serde_json::Value;
use std::net::SocketAddr;
use std::sync::Arc;
use std::time::Duration;
use tokio_util::sync::CancellationToken;
use tower_http::cors::{AllowOrigin, CorsLayer};
use tower_http::limit::RequestBodyLimitLayer;

/// Build the full router: unauthenticated probes + the bearer-guarded `/v1` API.
pub fn build_router(state: ServerState, config: &ServeConfig) -> Router {
    let public = Router::new()
        .route("/healthz", get(health::healthz))
        .route("/readyz", get(health::readyz))
        .route("/metrics", get(health::metrics));

    // `/v1` routes guarded by the bearer middleware via `route_layer` (only runs
    // for matched routes; OPTIONS preflight is allowed through inside the layer).
    #[cfg_attr(not(any(feature = "triggers", feature = "catalog")), allow(unused_mut))]
    let mut api = Router::new()
        .route("/v1/runs", post(runs::submit_run).get(runs::list_runs))
        .route("/v1/runs/{id}", get(runs::get_run).delete(runs::delete_run))
        .route("/v1/runs/{id}/cancel", post(runs::cancel_run))
        .route("/v1/runs/{id}/logs", get(logs::stream_logs))
        .route("/v1/schemas", get(schemas::list_schemas))
        .route("/v1/schemas/{kind}/{name}", get(schemas::get_schema))
        .route("/v1/doctor", post(doctor::doctor))
        .route("/v1/dlq/inspect", post(dlq::inspect))
        .route("/v1/dlq/replay", post(dlq::replay))
        .route("/v1/dlq/discard", post(dlq::discard))
        .route("/v1/audit", get(audit::list_audit));
    #[cfg(feature = "triggers")]
    {
        api = api.route(
            "/v1/triggers/{name}",
            post(crate::serve::triggers::webhook::handle)
                .put(crate::serve::triggers::webhook::handle),
        );
    }
    #[cfg(feature = "catalog")]
    {
        use crate::serve::handlers::catalog;
        api = api
            .route("/v1/catalog/datasets", get(catalog::list_datasets))
            .route("/v1/catalog/datasets/{id}", get(catalog::get_dataset))
            .route("/v1/catalog/lineage", get(catalog::lineage));
    }
    let api = api.route_layer(axum::middleware::from_fn_with_state(
        state.clone(),
        auth::require_auth,
    ));

    let cors = if config.cors_origins.is_empty() {
        CorsLayer::new()
    } else {
        let origins: Vec<axum::http::HeaderValue> = config
            .cors_origins
            .iter()
            .filter_map(|o| match o.parse() {
                Ok(v) => Some(v),
                Err(e) => {
                    tracing::warn!(origin = %o, error = %e, "ignoring invalid --cors-origin");
                    None
                }
            })
            .collect();
        CorsLayer::new().allow_origin(AllowOrigin::list(origins))
    };

    #[cfg_attr(not(feature = "serve-ui"), allow(unused_mut))]
    let mut router = public.merge(api);

    #[cfg(feature = "serve-ui")]
    if config.ui_enabled {
        use crate::serve::ui_assets;
        router = router
            .route("/", axum::routing::get(ui_assets::index))
            .route("/assets/{*path}", axum::routing::get(ui_assets::asset))
            .fallback(ui_assets::spa_fallback);
    }

    router
        .layer(RequestBodyLimitLayer::new(config.body_limit_bytes))
        .layer(axum::middleware::from_fn(metrics::track_metrics))
        .layer(cors)
        .with_state(state)
}

/// Load the optional `--default-config` once at startup, fully resolved, as a
/// merge base `Value`.
async fn load_default_base(config: &ServeConfig) -> CliResult<Option<Value>> {
    match &config.default_config_path {
        None => Ok(None),
        Some(path) => {
            // `serve` has no --profile flag; honour FAUCET_PROFILE from the environment directly.
            let profile = std::env::var("FAUCET_PROFILE").ok();
            let cfg =
                crate::config::PipelineConfig::from_path_async(path, profile.as_deref()).await?;
            Ok(Some(serde_json::to_value(&cfg).map_err(|e| {
                CliError::Serve(format!("serializing --default-config: {e}"))
            })?))
        }
    }
}

/// How often the background maintenance task purges expired history.
///
/// A quarter of the shorter of the two retention windows, clamped to
/// `[60s, 1h]`: frequent enough to bound store growth (and to honour a short
/// `idempotency_retention`) without churning on the multi-day default
/// `retain_terminal_runs`.
fn purge_interval(retain_terminal: Duration, idem_retention: Duration) -> Duration {
    (retain_terminal.min(idem_retention) / 4)
        .clamp(Duration::from_secs(60), Duration::from_secs(3600))
}

/// Background history-maintenance loop: every `period`, drop terminal run
/// records older than `retain` and expired idempotency claims, until `shutdown`
/// fires. Without this, the history store (in-memory `DashMap`s or the SQL
/// `faucet_serve_runs` / `faucet_serve_idem` tables) grows without bound for the
/// life of the process and the `--retain-terminal-runs-secs` /
/// `--idempotency-retention-secs` knobs are inert (audit #146 C4).
pub(crate) async fn maintenance_loop(
    history: Arc<dyn RunHistory>,
    retain: Duration,
    period: Duration,
    shutdown: CancellationToken,
) {
    let mut tick = tokio::time::interval(period);
    tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
    tick.tick().await; // consume the immediate first tick so we don't purge at t=0
    loop {
        tokio::select! {
            _ = shutdown.cancelled() => break,
            _ = tick.tick() => match history.purge_expired(retain).await {
                Ok(n) if n > 0 => {
                    tracing::info!(purged = n, "purged expired run records / idempotency claims")
                }
                Ok(_) => {}
                Err(e) => tracing::warn!(error = %e, "history purge_expired failed"),
            },
        }
    }
}

/// The lease heartbeat / orphan-recovery cadence: one third of the lease TTL
/// (so a run sees ≥2 renewals before its lease could expire), floored at 1s.
fn lease_interval(lease_ttl: Duration) -> Duration {
    (lease_ttl / 3).max(Duration::from_secs(1))
}

/// Background lease-maintenance loop (#146 H7). Every `period`:
///
/// 1. **Heartbeat** — renew this instance's own non-terminal runs' leases, so a
///    peer never reclaims a run we are still executing.
/// 2. **Recover** — fail any non-terminal run whose owning instance's lease has
///    expired (a crashed/gone peer), so a survivor eventually cleans up orphans
///    rather than waiting for the next process restart.
///
/// Renew runs *before* recover so this instance's leases are fresh when the
/// expiry scan runs. For the in-memory backend both calls are no-ops.
///
/// In cluster mode (#197) the recover step is replaced by: a membership
/// heartbeat, a live-member refresh (`member_ttl = period * 3`), and a
/// failover `reclaim_orphans` that re-queues an expired-lease peer's runs
/// (capped at `max_attempts`) rather than failing them outright.
pub(crate) async fn lease_loop(state: ServerState, period: Duration, shutdown: CancellationToken) {
    let cluster = state.cluster().clone();
    // Member-liveness window ≈ the real lease TTL (period == lease_ttl/3), so a
    // member must miss ~3 heartbeats before peers treat it as gone.
    let member_ttl = period.saturating_mul(3);
    let mut tick = tokio::time::interval(period);
    tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
    tick.tick().await; // consume the immediate first tick
    loop {
        tokio::select! {
            _ = shutdown.cancelled() => break,
            _ = tick.tick() => {
                if let Err(e) = state.history().renew_leases().await {
                    tracing::warn!(error = %e, "lease heartbeat (renew_leases) failed");
                }
                if cluster.enabled() {
                    // Membership heartbeat.
                    let beat = crate::serve::history::InstanceHeartbeat {
                        started_at: cluster.started_at(),
                        listen: Some(cluster.listen().to_string()),
                        max_concurrent: cluster.max_concurrent(),
                        in_flight: state.registry().in_flight() as u32,
                    };
                    if let Err(e) = state.history().heartbeat_instance(&beat).await {
                        tracing::warn!(error = %e, "cluster: heartbeat_instance failed");
                    }
                    match state.history().live_instances(member_ttl).await {
                        Ok(members) => {
                            cluster.set_members(members.len());
                            crate::serve::metrics::set_cluster_instances(members.len());
                        }
                        Err(e) => tracing::warn!(error = %e, "cluster: live_instances failed"),
                    }
                    // Failover reclaim (re-run orphans).
                    match state.history().reclaim_orphans(cluster.max_attempts()).await {
                        Ok(r) if r.requeued > 0 || r.failed > 0 => {
                            crate::serve::metrics::record_runs_reclaimed(r.requeued, r.failed);
                            tracing::warn!(
                                requeued = r.requeued, failed = r.failed,
                                "cluster: reclaimed orphaned runs from an expired-lease instance"
                            );
                        }
                        Ok(_) => {}
                        Err(e) => tracing::warn!(error = %e, "cluster: reclaim_orphans failed"),
                    }
                    // Mode B (#230): heartbeat this instance's running shards, and
                    // rebalance shards whose owner's lease expired (requeue →
                    // another worker, or poison past max_attempts).
                    if let Err(e) = state.history().renew_shard_leases().await {
                        tracing::warn!(error = %e, "cluster: renew_shard_leases failed");
                    }
                    match state.history().reclaim_shards(cluster.max_attempts()).await {
                        Ok(r) if r.requeued > 0 || r.failed > 0 => {
                            crate::serve::metrics::record_shards_reclaimed(r.requeued, r.failed);
                            tracing::warn!(
                                requeued = r.requeued, failed = r.failed,
                                "cluster: reclaimed orphaned shards from an expired-lease instance"
                            );
                        }
                        Ok(_) => {}
                        Err(e) => tracing::warn!(error = %e, "cluster: reclaim_shards failed"),
                    }
                    // Mode B (#230 / F11): finalize any `sharded` parent whose
                    // shards are all terminal but which no shard task finalized
                    // inline (e.g. the coordinator crashed after the last shard
                    // completed on another instance). Status-fenced + metric
                    // recorded inside the backend, so a parent already finalized by
                    // `maybe_finalize_parent` is not re-finalized or double-counted.
                    match state.history().finalize_completed_sharded_parents().await {
                        Ok(n) if n > 0 => tracing::info!(
                            finalized = n,
                            "cluster: finalized completed sharded parent run(s) via sweep"
                        ),
                        Ok(_) => {}
                        Err(e) => {
                            tracing::warn!(error = %e, "cluster: finalize_completed_sharded_parents failed")
                        }
                    }
                } else {
                    // Single-instance: mark orphans failed (today's behavior).
                    match state.history().recover_orphans().await {
                        Ok(n) if n > 0 => tracing::warn!(
                            recovered = n,
                            "recovered orphaned runs from an expired-lease instance"
                        ),
                        Ok(_) => {}
                        Err(e) => tracing::warn!(error = %e, "orphan recovery failed"),
                    }
                }
            }
        }
    }
}

/// Boot the server: install observability, build state + router, bind, serve
/// until SIGTERM/SIGINT, then drain in-flight runs up to the grace window.
pub async fn serve(config: ServeConfig) -> CliResult<()> {
    let (prom, log_hub) = crate::serve::observability::install(&config.log_level);
    crate::serve::metrics::set_cluster_enabled(config.cluster.enabled);

    // This process's identity for run-ownership leases (#146 H7). A fresh id per
    // process, so a restarted instance recovers its prior incarnation's runs only
    // once their lease expires — never another live instance's heartbeated runs.
    let instance_id = uuid::Uuid::new_v4().to_string();
    tracing::info!(
        instance_id = %instance_id,
        lease_ttl_secs = config.lease_ttl.as_secs(),
        "faucet serve instance id"
    );

    let history = crate::serve::history::connect(
        &config.history,
        config.idempotency_retention,
        config.lease_ttl,
        &instance_id,
    )
    .await?;
    if config.cluster.enabled {
        // Cluster mode: a restarting instance re-queues its prior incarnation's
        // in-flight runs (capped) rather than failing them.
        let report = history
            .reclaim_orphans(config.cluster.max_attempts)
            .await
            .map_err(|e| CliError::Serve(format!("history recovery: {e}")))?;
        if report.requeued > 0 || report.failed > 0 {
            tracing::warn!(
                requeued = report.requeued,
                failed = report.failed,
                "startup reclaim of orphaned runs from an expired-lease instance"
            );
        }
    } else {
        let recovered = history
            .recover_orphans()
            .await
            .map_err(|e| CliError::Serve(format!("history recovery: {e}")))?;
        if recovered > 0 {
            tracing::warn!(
                recovered,
                "marked orphaned non-terminal runs (expired owner lease) as failed"
            );
        }
    }
    let default_base = load_default_base(&config).await?;

    // Event-driven triggers (#196): load + validate the file (fail-fast), then
    // build the shared handle (webhook table + health rows).
    #[cfg(feature = "triggers")]
    let triggers = match &config.triggers_path {
        Some(path) => {
            // Register HELP text for the trigger metric family once at startup so
            // the series carry descriptions in `/metrics` (mirrors schedule).
            crate::serve::triggers::metrics::describe();
            Some(crate::serve::triggers::load_triggers(path).await?)
        }
        None => None,
    };
    #[cfg(feature = "triggers")]
    let triggers_handle = match &triggers {
        Some(c) => crate::serve::triggers::health::TriggersHandle::from_compiled(&c.triggers),
        None => crate::serve::triggers::health::TriggersHandle::empty(),
    };
    // A `--triggers` path in a build without the feature is a clear error.
    #[cfg(not(feature = "triggers"))]
    if config.triggers_path.is_some() {
        return Err(CliError::Serve(
            "--triggers requires a build with the `triggers` feature".into(),
        ));
    }

    let shutdown = CancellationToken::new();
    let state = ServerState::new(
        &config,
        prom,
        shutdown.clone(),
        history,
        log_hub,
        default_base,
        #[cfg(feature = "triggers")]
        triggers_handle,
    );
    let app = build_router(state.clone(), &config);

    let listener = tokio::net::TcpListener::bind(config.listen)
        .await
        .map_err(|e| CliError::Serve(format!("failed to bind {}: {e}", config.listen)))?;
    let local = listener
        .local_addr()
        .map_err(|e| CliError::Serve(e.to_string()))?;
    tracing::info!(listen = %local, "faucet serve listening");

    // Background history maintenance: bounds run-record / idempotency-claim
    // growth and makes the retention knobs effective (audit #146 C4).
    let purge_period = purge_interval(config.retain_terminal_runs, config.idempotency_retention);
    tracing::info!(
        interval_secs = purge_period.as_secs(),
        retain_secs = config.retain_terminal_runs.as_secs(),
        "history maintenance task started"
    );
    let maintenance = tokio::spawn(maintenance_loop(
        state.history(),
        config.retain_terminal_runs,
        purge_period,
        shutdown.clone(),
    ));

    // Lease heartbeat + cross-instance orphan recovery (#146 H7). Renews this
    // instance's run leases and reclaims runs whose owning instance's lease has
    // expired. A no-op for the in-memory backend.
    let lease_period = lease_interval(config.lease_ttl);
    let leases = tokio::spawn(lease_loop(state.clone(), lease_period, shutdown.clone()));

    // Cluster claim loop: pulls Pending runs from the shared DB (cluster only).
    let claim = if config.cluster.enabled {
        tracing::info!(
            poll_secs = config.cluster.poll.as_secs(),
            max_attempts = config.cluster.max_attempts,
            "cluster mode enabled; starting claim loop"
        );
        Some(tokio::spawn(crate::serve::cluster::claim_loop(
            state.clone(),
            shutdown.clone(),
        )))
    } else {
        None
    };

    // Event-driven trigger watchers (#196): spawn one supervised task per enabled
    // polling trigger (object_arrival / queue_depth). Webhook triggers are
    // handled by the router — no watcher task needed for them.
    #[cfg(feature = "triggers")]
    let trigger_handles = match &triggers {
        Some(c) => {
            tracing::info!(count = c.triggers.len(), "spawning trigger watchers");
            crate::serve::triggers::spawn_watchers(state.clone(), c, shutdown.clone())
        }
        None => Vec::new(),
    };

    // The HTTP graceful-shutdown future resolves on signal and stops accepting
    // new connections / drains in-flight HTTP — it does NOT cancel run tasks.
    // `into_make_service_with_connect_info` exposes the peer address so the auth
    // layer can record a `source_ip` on audit records (#205).
    axum::serve(
        listener,
        app.into_make_service_with_connect_info::<SocketAddr>(),
    )
    .with_graceful_shutdown(async move {
        wait_for_signal().await;
        tracing::info!("shutdown signal received; draining in-flight runs");
    })
    .await
    .map_err(|e| CliError::Serve(format!("server error: {e}")))?;

    // Stop pulling NEW work the moment we begin draining: abort the claim loop so
    // a shutting-down instance drains its in-flight runs rather than claiming more.
    // The lease loop keeps heartbeating in-flight runs during the drain so peers
    // don't reclaim them mid-shutdown.
    if let Some(claim) = claim {
        claim.abort();
    }

    // Now drain run tasks: wait up to the grace window, then cancel the rest.
    let drained =
        tokio::time::timeout(config.shutdown_grace, state.registry().wait_drained()).await;
    if drained.is_err() {
        let remaining = state.registry().in_flight();
        tracing::warn!(remaining, "grace window expired; cancelling in-flight runs");
        shutdown.cancel();
        // Give cancelled tasks a brief moment to write their terminal status.
        let _ = tokio::time::timeout(
            std::time::Duration::from_secs(5),
            state.registry().wait_drained(),
        )
        .await;
    }
    maintenance.abort();
    leases.abort();
    #[cfg(feature = "triggers")]
    for h in trigger_handles {
        h.abort();
    }
    // Flush any buffered OTLP telemetry after in-flight runs drain (no-op without
    // the `otel` feature).
    faucet_core::shutdown_otel();
    tracing::info!("faucet serve stopped");
    Ok(())
}

/// Resolve on SIGTERM (Unix) or Ctrl-C (any platform).
async fn wait_for_signal() {
    #[cfg(unix)]
    {
        use tokio::signal::unix::{SignalKind, signal};
        let mut term = match signal(SignalKind::terminate()) {
            Ok(s) => s,
            Err(_) => {
                let _ = tokio::signal::ctrl_c().await;
                return;
            }
        };
        tokio::select! {
            _ = tokio::signal::ctrl_c() => {}
            _ = term.recv() => {}
        }
    }
    #[cfg(not(unix))]
    {
        let _ = tokio::signal::ctrl_c().await;
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::serve::history::memory::MemoryHistory;
    use crate::serve::history::{RunRecord, RunStatus};
    use chrono::Utc;
    use std::collections::BTreeMap;

    #[test]
    fn lease_interval_is_third_of_ttl_floored_at_one_sec() {
        assert_eq!(
            lease_interval(Duration::from_secs(30)),
            Duration::from_secs(10)
        );
        assert_eq!(
            lease_interval(Duration::from_secs(90)),
            Duration::from_secs(30)
        );
        // Floor: a tiny TTL still heartbeats at least once per second.
        assert_eq!(
            lease_interval(Duration::from_secs(1)),
            Duration::from_secs(1)
        );
        assert_eq!(
            lease_interval(Duration::from_secs(2)),
            Duration::from_secs(1)
        );
    }

    #[test]
    fn purge_interval_is_quarter_of_shorter_window_clamped() {
        // Defaults (retain 7d, idem 1d) → min 1d, /4 = 6h → clamped to the 1h cap.
        assert_eq!(
            purge_interval(Duration::from_secs(604_800), Duration::from_secs(86_400)),
            Duration::from_secs(3600)
        );
        // A short idempotency window drives a faster cadence (but never below 60s).
        assert_eq!(
            purge_interval(Duration::from_secs(604_800), Duration::from_secs(120)),
            Duration::from_secs(60)
        );
        // Both tiny → the 60s floor.
        assert_eq!(
            purge_interval(Duration::from_secs(1), Duration::from_secs(1)),
            Duration::from_secs(60)
        );
        // A 40-minute window lands inside the range: 2400/4 = 600s.
        assert_eq!(
            purge_interval(Duration::from_secs(2400), Duration::from_secs(2400)),
            Duration::from_secs(600)
        );
    }

    #[tokio::test]
    async fn maintenance_loop_purges_expired_terminal_runs() {
        let history: Arc<dyn RunHistory> = Arc::new(MemoryHistory::new(Duration::from_secs(60)));

        // An old terminal record (eligible for purge with retain=0) and a
        // non-terminal one (must be kept).
        let mut old = RunRecord::queued(
            "old".into(),
            None,
            BTreeMap::new(),
            None,
            Utc::now() - chrono::Duration::seconds(10),
        );
        old.status = RunStatus::Completed;
        old.finished_at = Some(Utc::now() - chrono::Duration::seconds(10));
        history.upsert(&old).await.unwrap();
        let live = RunRecord::queued("live".into(), None, BTreeMap::new(), None, Utc::now());
        history.upsert(&live).await.unwrap();

        let shutdown = CancellationToken::new();
        let handle = tokio::spawn(maintenance_loop(
            history.clone(),
            Duration::ZERO,            // retain=0 → every terminal record is expired
            Duration::from_millis(10), // fast tick for the test
            shutdown.clone(),
        ));

        // Allow several ticks (the first is consumed at t=0).
        tokio::time::sleep(Duration::from_millis(80)).await;
        shutdown.cancel();
        let _ = handle.await;

        assert!(
            history.get("old").await.unwrap().is_none(),
            "expired terminal run should have been purged by the maintenance loop"
        );
        assert!(
            history.get("live").await.unwrap().is_some(),
            "non-terminal run must be kept"
        );
    }
}