ironflow_api/

purger.rs

//! Retention-based purging of old runs and their artifacts.
//!
//! The [`RunPurger`] periodically removes terminal runs that exceed the
//! configured retention policy (by age or by per-workflow count) and deletes
//! their artifact blobs from the blob store.
//!
//! This is distinct from the [`Reaper`](crate::reaper::Reaper), which recovers
//! runs whose worker lease expired. The reaper keeps runs alive; the purger
//! removes the ones that are done and old.
//!
//! [`RunPurger::from_config`] wires a purger from a [`ServerConfig`].

use std::sync::Arc;
use std::time::Duration;

use chrono::{TimeDelta, Utc};

use uuid::Uuid;

use crate::config::ServerConfig;
use ironflow_artifacts::blob_store::BlobStore;
use ironflow_store::entities::{PurgePolicy, PurgeReason};
use ironflow_store::store::Store;
use tokio::time::interval;
use tokio_util::sync::CancellationToken;
use tracing::{error, info, warn};

#[cfg(feature = "prometheus")]
use ironflow_core::metric_names::RUNS_PURGED_TOTAL;
#[cfg(feature = "prometheus")]
use metrics::counter;

/// How often the purger runs by default (once per day).
pub const DEFAULT_PURGE_INTERVAL: Duration = Duration::from_secs(86400);

/// How many runs a single tick processes.
pub const DEFAULT_PURGE_BATCH_SIZE: u32 = 100;

/// Default days of Provider Account usage history kept.
pub const DEFAULT_USAGE_RETENTION_DAYS: u32 = 30;

/// Default days received signals are kept.
pub const DEFAULT_SIGNAL_RETENTION_DAYS: u32 = 7;

/// Periodic task that purges terminal runs exceeding the retention policy.
///
/// # Examples
///
/// ```no_run
/// use std::sync::Arc;
/// use std::time::Duration;
/// use ironflow_api::purger::RunPurger;
/// use ironflow_store::entities::PurgePolicy;
/// use ironflow_store::memory::InMemoryStore;
/// use ironflow_store::store::Store;
/// use tokio_util::sync::CancellationToken;
///
/// # async fn example() {
/// let store: Arc<dyn Store> = Arc::new(InMemoryStore::new());
/// let policy = PurgePolicy {
///     max_age_days: 90,
///     max_runs_per_workflow: 1000,
///     dry_run: false,
/// };
///
/// let purger = RunPurger::new(store, policy)
///     .interval(Duration::from_secs(3600));
/// tokio::spawn(purger.run(CancellationToken::new()));
/// # }
/// ```
pub struct RunPurger {
    store: Arc<dyn Store>,
    blob_store: Option<Arc<dyn BlobStore>>,
    pub(crate) policy: PurgePolicy,
    pub(crate) interval: Duration,
    batch_size: u32,
    pub(crate) usage_retention_days: u32,
    pub(crate) signal_retention_days: u32,
}

impl RunPurger {
    /// Create a purger with the default interval and batch size.
    pub fn new(store: Arc<dyn Store>, policy: PurgePolicy) -> Self {
        Self {
            store,
            blob_store: None,
            policy,
            interval: DEFAULT_PURGE_INTERVAL,
            batch_size: DEFAULT_PURGE_BATCH_SIZE,
            usage_retention_days: DEFAULT_USAGE_RETENTION_DAYS,
            signal_retention_days: DEFAULT_SIGNAL_RETENTION_DAYS,
        }
    }

    /// Create a purger from the server configuration.
    ///
    /// Takes the purge policy (`PURGE_MAX_AGE_DAYS`, `PURGE_MAX_RUNS_PER_WORKFLOW`,
    /// `PURGE_DRY_RUN`), the interval (`PURGE_INTERVAL_SECS`) and the retention of
    /// Provider Account usage history and signals
    /// (`PROVIDER_ACCOUNT_USAGE_RETENTION_DAYS`, `SIGNAL_RETENTION_DAYS`).
    /// Chain [`RunPurger::with_blob_store`] to delete artifact blobs too.
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use std::sync::Arc;
    /// use ironflow_api::config::{ConfigError, ServerConfig};
    /// use ironflow_api::purger::RunPurger;
    /// use ironflow_store::memory::InMemoryStore;
    /// use tokio_util::sync::CancellationToken;
    ///
    /// # async fn example() -> Result<(), ConfigError> {
    /// let config = ServerConfig::from_env()?;
    /// let purger = RunPurger::from_config(Arc::new(InMemoryStore::new()), &config);
    /// tokio::spawn(purger.run(CancellationToken::new()));
    /// # Ok(())
    /// # }
    /// ```
    pub fn from_config(store: Arc<dyn Store>, config: &ServerConfig) -> Self {
        let policy = PurgePolicy {
            max_age_days: config.purge_max_age_days,
            max_runs_per_workflow: config.purge_max_runs_per_workflow,
            dry_run: config.purge_dry_run,
        };
        Self::new(store, policy)
            .interval(Duration::from_secs(config.purge_interval_secs))
            .usage_retention_days(config.provider_account_usage_retention_days)
            .signal_retention_days(config.signal_retention_days)
    }

    /// Set the blob store for artifact deletion.
    ///
    /// When `None`, artifact metadata is still removed but no blobs are deleted
    /// (they either do not exist or become orphans).
    pub fn with_blob_store(mut self, blob_store: Option<Arc<dyn BlobStore>>) -> Self {
        self.blob_store = blob_store;
        self
    }

    /// Set how often the purger runs.
    pub fn interval(mut self, interval: Duration) -> Self {
        self.interval = interval;
        self
    }

    /// Set how many runs a single tick processes.
    pub fn batch_size(mut self, batch_size: u32) -> Self {
        self.batch_size = batch_size;
        self
    }

    /// Set how many days of Provider Account usage history to keep.
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use std::sync::Arc;
    /// use ironflow_api::purger::RunPurger;
    /// use ironflow_store::entities::PurgePolicy;
    /// use ironflow_store::memory::InMemoryStore;
    ///
    /// let policy = PurgePolicy { max_age_days: 90, max_runs_per_workflow: 1000, dry_run: false };
    /// let purger = RunPurger::new(Arc::new(InMemoryStore::new()), policy).usage_retention_days(7);
    /// # let _ = purger;
    /// ```
    pub fn usage_retention_days(mut self, days: u32) -> Self {
        self.usage_retention_days = days;
        self
    }

    /// Set how many days received signals are kept.
    ///
    /// A signal older than this can no longer resume a run that opens its
    /// wait step late.
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use std::sync::Arc;
    /// use ironflow_api::purger::RunPurger;
    /// use ironflow_store::entities::PurgePolicy;
    /// use ironflow_store::memory::InMemoryStore;
    ///
    /// let policy = PurgePolicy { max_age_days: 90, max_runs_per_workflow: 1000, dry_run: false };
    /// let purger = RunPurger::new(Arc::new(InMemoryStore::new()), policy).signal_retention_days(3);
    /// # let _ = purger;
    /// ```
    pub fn signal_retention_days(mut self, days: u32) -> Self {
        self.signal_retention_days = days;
        self
    }

    /// Run the purge loop until `shutdown` is cancelled.
    pub async fn run(self, shutdown: CancellationToken) {
        let mut ticker = interval(self.interval);
        ticker.tick().await;

        info!(
            interval_secs = self.interval.as_secs(),
            batch_size = self.batch_size,
            max_age_days = self.policy.max_age_days,
            max_runs_per_workflow = self.policy.max_runs_per_workflow,
            dry_run = self.policy.dry_run,
            usage_retention_days = self.usage_retention_days,
            signal_retention_days = self.signal_retention_days,
            "purger started"
        );

        loop {
            tokio::select! {
                _ = shutdown.cancelled() => {
                    info!("purger stopped");
                    return;
                }
                _ = ticker.tick() => {
                    self.tick().await;
                }
            }
        }
    }

    /// Delete Provider Account usage history older than the retention.
    async fn purge_account_usage(&self) {
        if self.policy.dry_run {
            return;
        }
        let before = Utc::now() - TimeDelta::days(i64::from(self.usage_retention_days));
        match self.store.purge_provider_account_usage(before).await {
            Ok(0) => {}
            Ok(deleted) => info!(deleted, "purged provider account usage history"),
            Err(err) => error!(error = %err, "failed to purge provider account usage history"),
        }
    }

    /// Delete the signals received before the retention window.
    async fn purge_signals(&self) {
        if self.policy.dry_run {
            return;
        }
        let before = Utc::now() - TimeDelta::days(i64::from(self.signal_retention_days));
        match self.store.purge_signals(before).await {
            Ok(0) => {}
            Ok(deleted) => info!(deleted, "purged old signals"),
            Err(err) => error!(error = %err, "failed to purge old signals"),
        }
    }

    /// Purge one batch of eligible runs.
    ///
    /// Exposed for tests and for callers that drive the schedule themselves.
    pub async fn tick(&self) {
        self.purge_account_usage().await;
        self.purge_signals().await;

        let purgeable = match self
            .store
            .list_purgeable_runs(&self.policy, self.batch_size)
            .await
        {
            Ok(p) => p,
            Err(err) => {
                error!(error = %err, "failed to list purgeable runs");
                return;
            }
        };

        if purgeable.is_empty() {
            return;
        }

        if self.policy.dry_run {
            for entry in &purgeable {
                info!(
                    run_id = %entry.run_id,
                    workflow = %entry.workflow_name,
                    reason = %entry.reason,
                    "[dry-run] would purge run"
                );

                #[cfg(feature = "prometheus")]
                counter!(
                    RUNS_PURGED_TOTAL,
                    "workflow" => entry.workflow_name.clone(),
                    "reason" => entry.reason.to_string(),
                    "dry_run" => "true"
                )
                .increment(1);
            }

            return;
        }

        warn!(
            count = purgeable.len(),
            batch_size = self.batch_size,
            "purging old runs"
        );

        for entry in &purgeable {
            self.purge_run(entry.run_id, &entry.workflow_name, &entry.reason)
                .await;
        }
    }

    async fn purge_run(&self, run_id: Uuid, workflow_name: &str, reason: &PurgeReason) {
        match self.store.delete_run(run_id).await {
            Ok(storage_keys) => {
                if let Some(ref blob_store) = self.blob_store {
                    for key in &storage_keys {
                        let refcount = self
                            .store
                            .count_artifacts_by_storage_key(key)
                            .await
                            .unwrap_or(0);
                        if refcount > 0 {
                            continue;
                        }
                        if let Err(err) = blob_store.delete(key).await {
                            error!(
                                run_id = %run_id,
                                storage_key = %key,
                                error = %err,
                                "failed to delete artifact blob"
                            );
                        }
                    }
                }

                info!(
                    run_id = %run_id,
                    workflow = %workflow_name,
                    reason = %reason,
                    "purged run"
                );

                #[cfg(feature = "prometheus")]
                counter!(
                    RUNS_PURGED_TOTAL,
                    "workflow" => workflow_name.to_string(),
                    "reason" => reason.to_string(),
                    "dry_run" => "false"
                )
                .increment(1);
            }
            Err(err) => {
                error!(
                    run_id = %run_id,
                    workflow = %workflow_name,
                    error = %err,
                    "failed to delete run during purge"
                );
            }
        }
    }
}

#[cfg(test)]
mod tests {
    use std::collections::HashMap;
    use std::time::Duration;

    use chrono::{TimeDelta, Utc};
    use ironflow_store::entities::{
        AccountWindowStatus, NewAccountWindow, NewProviderAccount, NewProviderAccountObservation,
        NewRun, NewSignal, PurgePolicy, RunStatus, SignalFilter, TriggerKind,
        provider_account_secret_key,
    };
    use ironflow_store::memory::InMemoryStore;
    use ironflow_store::provider_account_store::ProviderAccountStore;
    use ironflow_store::signal_store::SignalStore;
    use ironflow_store::store::RunStore;
    use serde_json::json;
    use tokio::time::sleep;
    use uuid::Uuid;

    use super::*;

    fn new_run(name: &str) -> NewRun {
        NewRun {
            workflow_name: name.to_string(),
            trigger: TriggerKind::Manual,
            payload: json!({}),
            max_retries: 0,
            handler_version: None,
            labels: HashMap::new(),
            scheduled_at: None,
            created_by: None,
            idempotency_key: None,
            concurrency_key: None,
            max_cost_usd: None,
        }
    }

    async fn create_terminal_run(store: &InMemoryStore, name: &str, status: RunStatus) -> Uuid {
        let run = store.create_run(new_run(name)).await.unwrap().into_run();
        store
            .update_run_status(run.id, RunStatus::Running)
            .await
            .unwrap();
        store.update_run_status(run.id, status).await.unwrap();
        run.id
    }

    async fn backdate_run(store: &InMemoryStore, run_id: Uuid, days: i64) {
        store
            .set_run_created_at(run_id, Utc::now() - TimeDelta::days(days))
            .await;
    }

    fn build(store: Arc<InMemoryStore>, policy: PurgePolicy) -> RunPurger {
        let store_dyn: Arc<dyn Store> = store;
        RunPurger::new(store_dyn, policy)
    }

    #[tokio::test]
    async fn purger_purges_old_provider_account_usage() {
        let store = Arc::new(InMemoryStore::new());
        let id = Uuid::now_v7();
        store
            .create_provider_account(NewProviderAccount {
                id,
                name: "perso".to_string(),
                display_name: "Perso".to_string(),
                kind: "claude_subscription".to_string(),
                secret_key: provider_account_secret_key(id),
                enabled: true,
                priority: 100,
                tags: Vec::new(),
                max_concurrency: None,
                alert_threshold: 0.8,
                expires_at: Utc::now() + TimeDelta::days(365),
                plan: None,
                created_by: None,
            })
            .await
            .unwrap();
        let window = |days_ago: i64| NewAccountWindow {
            window: "five_hour".to_string(),
            utilization: 0.5,
            resets_at: None,
            status: AccountWindowStatus::Allowed,
            model_scope: None,
            observed_at: Utc::now() - TimeDelta::days(days_ago),
        };
        store
            .record_provider_account_observation(
                id,
                NewProviderAccountObservation {
                    windows: vec![window(40), window(1)],
                    auth_failed: false,
                },
            )
            .await
            .unwrap();

        let policy = PurgePolicy {
            max_age_days: 90,
            max_runs_per_workflow: 10000,
            dry_run: false,
        };
        build(store.clone(), policy)
            .usage_retention_days(30)
            .tick()
            .await;

        let history = store
            .list_provider_account_usage(id, Utc::now() - TimeDelta::days(90))
            .await
            .unwrap();
        assert_eq!(history.len(), 1);
    }

    #[tokio::test]
    async fn tick_purges_runs_older_than_max_age() {
        let store = Arc::new(InMemoryStore::new());
        let old_id = create_terminal_run(&store, "deploy", RunStatus::Completed).await;
        backdate_run(&store, old_id, 100).await;
        let recent_id = create_terminal_run(&store, "deploy", RunStatus::Completed).await;

        let policy = PurgePolicy {
            max_age_days: 90,
            max_runs_per_workflow: 10000,
            dry_run: false,
        };
        let purger = build(store.clone(), policy);
        purger.tick().await;

        assert!(store.get_run(old_id).await.unwrap().is_none());
        assert!(store.get_run(recent_id).await.unwrap().is_some());
    }

    #[tokio::test]
    async fn tick_in_dry_run_does_not_delete() {
        let store = Arc::new(InMemoryStore::new());
        let old_id = create_terminal_run(&store, "deploy", RunStatus::Completed).await;
        backdate_run(&store, old_id, 100).await;

        let policy = PurgePolicy {
            max_age_days: 90,
            max_runs_per_workflow: 10000,
            dry_run: true,
        };
        let purger = build(store.clone(), policy);
        purger.tick().await;

        assert!(store.get_run(old_id).await.unwrap().is_some());
    }

    #[tokio::test]
    async fn tick_does_not_purge_non_terminal_runs() {
        let store = Arc::new(InMemoryStore::new());
        let pending = store
            .create_run(new_run("deploy"))
            .await
            .unwrap()
            .into_run();
        backdate_run(&store, pending.id, 200).await;

        let running = store
            .create_run(new_run("deploy"))
            .await
            .unwrap()
            .into_run();
        store
            .update_run_status(running.id, RunStatus::Running)
            .await
            .unwrap();
        backdate_run(&store, running.id, 200).await;

        let policy = PurgePolicy {
            max_age_days: 90,
            max_runs_per_workflow: 10000,
            dry_run: false,
        };
        let purger = build(store.clone(), policy);
        purger.tick().await;

        assert!(store.get_run(pending.id).await.unwrap().is_some());
        assert!(store.get_run(running.id).await.unwrap().is_some());
    }

    #[tokio::test]
    async fn purge_does_not_delete_blob_when_shared_by_another_artifact() {
        use ironflow_artifacts::local::LocalBlobStore;
        use ironflow_artifacts::stream_from_bytes;
        use ironflow_store::artifact_store::ArtifactStore;
        use ironflow_store::entities::{NewArtifact, NewStep, StepKind, step_trace_id};
        use ironflow_store::store::Store;
        use tempfile::TempDir;

        let dir = TempDir::new().expect("temp dir");
        let store = Arc::new(InMemoryStore::new());
        let blob: Arc<dyn BlobStore> = Arc::new(LocalBlobStore::new(dir.path()));

        // Create two runs, each with a step
        let run_a = store.create_run(new_run("wf")).await.unwrap().into_run();
        store
            .update_run_status(run_a.id, RunStatus::Running)
            .await
            .unwrap();
        store
            .update_run_status(run_a.id, RunStatus::Completed)
            .await
            .unwrap();
        backdate_run(&store, run_a.id, 100).await;

        let step_a = store
            .create_step(NewStep {
                run_id: run_a.id,
                trace_id: step_trace_id(run_a.id, "build", 0),
                name: "build".to_string(),
                kind: StepKind::Shell,
                position: 0,
                input: None,
                is_error_handler: false,
            })
            .await
            .unwrap();

        let run_b = store.create_run(new_run("wf")).await.unwrap().into_run();
        let step_b = store
            .create_step(NewStep {
                run_id: run_b.id,
                trace_id: step_trace_id(run_b.id, "build", 0),
                name: "build".to_string(),
                kind: StepKind::Shell,
                position: 0,
                input: None,
                is_error_handler: false,
            })
            .await
            .unwrap();

        // Upload one blob, share the storage_key
        let shared_key = "artifacts/shared/blob";
        blob.put(shared_key, stream_from_bytes(b"shared".to_vec()))
            .await
            .unwrap();

        let id_a = Uuid::now_v7();
        store
            .create_artifact(NewArtifact {
                id: id_a,
                run_id: run_a.id,
                step_id: step_a.id,
                name: "report.txt".to_string(),
                storage_key: shared_key.to_string(),
                content_type: "text/plain".to_string(),
                size_bytes: 6,
                sha256: "abc".repeat(21),
            })
            .await
            .unwrap();

        let id_b = Uuid::now_v7();
        store
            .create_artifact(NewArtifact {
                id: id_b,
                run_id: run_b.id,
                step_id: step_b.id,
                name: "report.txt".to_string(),
                storage_key: shared_key.to_string(),
                content_type: "text/plain".to_string(),
                size_bytes: 6,
                sha256: "abc".repeat(21),
            })
            .await
            .unwrap();

        // Purge run_a (old) -- the blob should survive because run_b still references it
        let policy = PurgePolicy {
            max_age_days: 90,
            max_runs_per_workflow: 10000,
            dry_run: false,
        };
        let store_dyn: Arc<dyn Store> = store.clone();
        let purger = RunPurger::new(store_dyn, policy).with_blob_store(Some(blob.clone()));
        purger.tick().await;

        assert!(
            store.get_run(run_a.id).await.unwrap().is_none(),
            "run_a purged"
        );
        assert!(
            store.get_run(run_b.id).await.unwrap().is_some(),
            "run_b kept"
        );

        // The blob should still exist
        let get_result = blob.get(shared_key).await;
        assert!(get_result.is_ok(), "shared blob should not be deleted");
    }

    #[tokio::test]
    async fn purger_tick_purges_old_signals() {
        let store = Arc::new(InMemoryStore::new());
        store
            .insert_signal(NewSignal {
                name: "demo.done".to_string(),
                key: "k1".to_string(),
                payload: json!({}),
                idempotency_id: None,
            })
            .await
            .unwrap();
        sleep(Duration::from_millis(5)).await;
        let policy = PurgePolicy {
            max_age_days: 90,
            max_runs_per_workflow: 1000,
            dry_run: false,
        };

        // A one-day window keeps the fresh signal.
        build(store.clone(), policy.clone())
            .signal_retention_days(1)
            .tick()
            .await;
        let kept = store
            .list_signals(SignalFilter::default(), 1, 20)
            .await
            .unwrap();
        assert_eq!(kept.total, 1);

        // A zero-day window drops everything received before now.
        build(store.clone(), policy)
            .signal_retention_days(0)
            .tick()
            .await;
        let purged = store
            .list_signals(SignalFilter::default(), 1, 20)
            .await
            .unwrap();
        assert_eq!(purged.total, 0);
    }

    #[tokio::test]
    async fn run_stops_on_shutdown() {
        let store = Arc::new(InMemoryStore::new());
        let policy = PurgePolicy {
            max_age_days: 90,
            max_runs_per_workflow: 1000,
            dry_run: false,
        };
        let purger = build(store, policy);
        let shutdown = CancellationToken::new();
        shutdown.cancel();

        tokio::time::timeout(Duration::from_secs(5), purger.run(shutdown))
            .await
            .expect("purger stopped");
    }
}