Remove apalis + redis => use postgresql instead

.devcontainer/Dockerfile

@@ -15,8 +15,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     cmake \
     pkg-config \
     clang \
-    valkey-server \
-    valkey-tools \
+    postgresql \
     && rm -rf /var/lib/apt/lists/*
 
 RUN groupadd --gid ${USER_GID:-1000} $USERNAME \

.env.example

@@ -22,3 +22,5 @@ METRICS_BIND_ADDR=
 # Sandboxed tool execution
 # DOCKER_HOST=
 SANDBOX_MAX_ITERATIONS=8
+
+DATABASE_URL=

.woodpecker/tests.yml

@@ -2,6 +2,14 @@ when:
     event:
         - push
 
+services:
+    - name: postgres
+      image: postgres:18
+      environment:
+          POSTGRES_USER: postgres
+          POSTGRES_PASSWORD: postgres
+          POSTGRES_DB: steep
+
 steps:
     - name: fmt
       image: rust:1.99
@@ -15,10 +23,17 @@ steps:
           - rustup component add clippy
           - cargo clippy --workspace --all-targets --all-features -- -D warnings
 
+    - name: wait-for-postgres
+      image: postgres:18
+      commands:
+          - until pg_isready -h postgres -U postgres; do sleep 1; done
+
     - name: test
       image: rust:1.99
+      environment:
+          DATABASE_URL: postgresql://postgres:postgres@postgres:5432/steep
       commands:
-          - cargo test --workspace --all-targets
+          - cargo test --workspace --all-targets -- --include-ignored
 
     - name: container-build
       image: quay.io/buildah/stable

Cargo.toml

@@ -28,9 +28,9 @@ hex = "0.4"
 bytes = "1.1"
 metrics = "0.24"
 metrics-exporter-prometheus = { version = "0.18", default-features = false, features = ["http-listener"] }
-apalis = { version = "0.7", features = ["catch-panic", "limit"] }
-apalis-redis = "0.7"
-redis = { version = "0.32", features = ["tokio-comp", "connection-manager"] }
+toasty = { version = "0.11", features = ["postgresql", "serde", "jiff"] }
+toasty-cli = "0.11"
+jiff = "0.2"
 
 [profile.release]
 debug = 1

Containerfile

@@ -10,10 +10,7 @@ RUN cargo build --release --package steep-server
 
 FROM debian:trixie-slim
 
-# git clones the pull request. ca-certificates is what every HTTPS call needs
-# (Gitea, OpenRouter, git). The shared libraries are the non-base ones the binary
-# links against, as reported by `ldd target/release/steep-server`; libc, libm
-# and libgcc_s come from the base image.
+
 RUN apt-get update && apt-get install -y --no-install-recommends \
     ca-certificates \
     git \
@@ -26,6 +23,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 WORKDIR /app
 COPY --from=builder /app/target/release/steep-server ./steep-server
 
-# Exec form, so the binary is PID 1 and receives the SIGTERM it handles to shut
-# down gracefully.
-CMD ["./steep-server"]
+COPY crates/steep-server/toasty ./toasty
+ENV MIGRATIONS_DIR=/app/toasty
+
+CMD ["sh", "-c", "./steep-server migration apply && exec ./steep-server"]

README.md

@@ -38,7 +38,7 @@ Steep reads its configuration from environment variables (a `.env` file is suppo
 | `GITEA_URL` | Base URL of your Gitea instance |
 | `GITEA_TOKEN` | Gitea API token |
 | `GITEA_TIMEOUT` | Gitea API request timeout in seconds |
-| `REDIS_URL` | Redis connection URL used for the durable review queue (e.g. `redis://127.0.0.1:6379`) |
+| `DATABASE_URL` | PostgreSQL connection URL used for the durable review queue (e.g. `postgresql://user:pass@localhost:5432/steep`) |
 | `METRICS_BIND_ADDR` | *(optional)* Bind address for the Prometheus metrics endpoint (e.g. `0.0.0.0:9100`). If unset, the metrics exporter is disabled. |
 | `SENTRY_DSN` | *(optional)* Sentry DSN for error tracking |
 | `RUST_LOG` | *(optional)* Log level, defaults to `info` |
@@ -102,6 +102,26 @@ hooks have run, the container is disconnected from the network for the rest of
 the review. Every container command is bounded by a timeout, and the container,
 network and image are removed when the review ends (including on failure).
 
+## Migrations
+
+The database schema is versioned with [Toasty](https://github.com/tokio-rs/toasty) migrations, stored in `crates/steep-server/toasty/` (commit them).
+
+After changing a model in `crates/steep-server/src/schemas/`, generate a migration:
+
+```sh
+cargo run -p steep-server -- migration generate --name describe_change
+```
+
+Apply the pending migrations **before starting the service** — the server does not migrate at boot:
+
+```sh
+cargo run -p steep-server -- migration apply
+```
+
+The `Containerfile` does this at startup: its entrypoint runs `steep-server migration apply` before exec'ing the server. The CLI reads the migration files from `MIGRATIONS_DIR` (defaults to the crate's `toasty/` directory; the image sets it to `/app/toasty`).
+
+Other subcommands: `migration snapshot`, `migration drop`, `migration reset`. They read `DATABASE_URL`.
+
 ## Development
 
 The easiest way to get started is with the provided [Dev Container](https://containers.dev/) (VS Code or Zed with the dev container extension).
@@ -129,7 +149,7 @@ Set `METRICS_BIND_ADDR` (e.g. `0.0.0.0:9100`) to enable it. The metrics are then
 |---|---|---|
 | `steep_webhooks_received_total` | counter | Total webhooks received (label: `event_type`) |
 | `steep_webhooks_duplicate_total` | counter | Webhooks rejected as duplicates (label: `event_type`) |
-| `steep_webhooks_enqueue_failed_total` | counter | Webhooks dropped because enqueueing to Redis failed (label: `event_type`) |
+| `steep_webhooks_enqueue_failed_total` | counter | Webhooks dropped because enqueueing to PostgreSQL failed (label: `event_type`) |
 | `steep_bot_tasks_active` | gauge | Bot tasks currently in progress |
 | `steep_bot_tasks_skipped_total` | counter | Reviews skipped because they were already done (label: `event_type`) |
 | `steep_bot_tasks_completed_total` | counter | Bot tasks completed successfully (label: `event_type`) |

crates/steep-server/Cargo.toml

@@ -26,6 +26,6 @@ metrics = { workspace = true }
 metrics-exporter-prometheus = { workspace = true }
 devcontainer-rs = { path = "../devcontainer-rs" }
 tempfile = "3"
-apalis = { workspace = true }
-apalis-redis = { workspace = true }
-redis = { workspace = true }
+toasty = { workspace = true }
+toasty-cli = { workspace = true }
+jiff = { workspace = true }

crates/steep-server/src/api.rs

@@ -14,10 +14,7 @@ use tracing::{info, instrument};
 
 use tokio_util::sync::CancellationToken;
 
-use apalis::prelude::Storage;
-
 use crate::consts::{GITEA_EVENT_TYPE_HEADER_NAME, GITEA_SIG_HEADER_NAME, MAX_WEBHOOK_BODY_SIZE};
-use crate::dedup::{Dedup, WEBHOOK_TTL};
 use crate::errors::AppError;
 use crate::gitea::WebhookType;
 use crate::metrics;
@@ -64,20 +61,19 @@ async fn webhook(
     metrics::webhook_received(event_type);
 
     let event_id = wb.event_id();
-    let webhook_key = Dedup::webhook_key(event_id);
-    if !app_state
-        .dedup
-        .claim(&webhook_key, WEBHOOK_TTL)
-        .await
-        .map_err(anyhow::Error::from)?
-    {
-        metrics::webhook_duplicate(event_type);
-        return Err(AppError::AlreadyProcessedErr);
+    match app_state.dedup.is_known(event_id).await {
+        Ok(true) => {
+            metrics::webhook_duplicate(event_type);
+            return Err(AppError::AlreadyProcessedErr);
+        }
+        Ok(false) => {}
+        Err(err) => {
+            metrics::webhook_enqueue_failed(event_type);
+            return Err(AppError::EnqueueErr(anyhow::Error::from(err)));
+        }
     }
 
-    let mut storage = app_state.storage.clone();
-    if let Err(err) = storage.push(wb).await {
-        app_state.dedup.release(&webhook_key).await.ok();
+    if let Err(err) = app_state.queue.push(wb).await {
         metrics::webhook_enqueue_failed(event_type);
         return Err(AppError::EnqueueErr(anyhow::Error::from(err)));
     }

crates/steep-server/src/bot.rs

@@ -1,15 +1,21 @@
+use std::sync::Arc;
+
+use futures_util::StreamExt;
+use serde::{Deserialize, Deserializer};
+use tokio::sync::Semaphore;
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info, instrument};
+
 use crate::{
-    dedup::{Dedup, REVIEW_TTL},
+    bot_actions::review::ReviewOutcome,
+    dedup::{Dedup, ReviewRecord},
     gitea::{GiteaAPI, WebhookType},
+    job_queue::JobQueue,
     metrics,
     open_router::OpenRouterClient,
     sandbox::SandboxConfig,
+    schemas::Job,
 };
-use apalis::prelude::{Data, Error, Monitor, WorkerBuilder, WorkerBuilderExt, WorkerFactoryFn};
-use apalis_redis::RedisStorage;
-use serde::{Deserialize, Deserializer};
-use tokio_util::sync::CancellationToken;
-use tracing::{error, info, instrument};
 
 #[derive(Deserialize, Debug)]
 pub struct ReviewResult {
@@ -127,6 +133,7 @@ pub struct Bot {
     open_router_model: String,
     sandbox: SandboxConfig,
     dedup: Dedup,
+    queue: JobQueue,
 }
 
 impl Bot {
@@ -137,6 +144,7 @@ impl Bot {
         open_router_model: String,
         sandbox: SandboxConfig,
         dedup: Dedup,
+        queue: JobQueue,
     ) -> Self {
         Self {
             gitea_api,
@@ -145,40 +153,119 @@ impl Bot {
             open_router_model,
             sandbox,
             dedup,
+            queue,
         }
     }
 
-    /// Consomme la file Redis jusqu'à l'annulation de `shutdown`.
+    /// Consomme la file jusqu'à l'annulation de `shutdown`.
     ///
-    /// Les jobs en cours sont laissés se terminer, et ceux qui n'ont pas encore
-    /// été acquittés restent dans Redis : ils seront repris au prochain démarrage.
-    pub async fn start(
-        &self,
-        storage: RedisStorage<WebhookType>,
-        shutdown: CancellationToken,
-    ) -> anyhow::Result<()> {
+    /// Un sémaphore borne la concurrence ; les jobs en cours sont laissés se
+    /// terminer. Les jobs non terminés restent en base et seront repris au
+    /// prochain démarrage.
+    pub async fn start(&self, shutdown: CancellationToken) -> anyhow::Result<()> {
         info!("Bot started");
 
-        let worker = WorkerBuilder::new("steep-review")
-            .concurrency(self.max_concurrent)
-            .data(self.clone())
-            .backend(storage)
-            .build_fn(handle_review);
+        let semaphore = Arc::new(Semaphore::new(self.max_concurrent));
+        let mut tasks = tokio::task::JoinSet::new();
+        let mut pending = std::pin::pin!(self.queue.pending());
 
-        Monitor::new()
-            .register(worker)
-            .run_with_signal(async move {
-                shutdown.cancelled().await;
-                Ok(())
-            })
-            .await?;
+        loop {
+            let job = tokio::select! {
+                biased;
+                _ = shutdown.cancelled() => break,
+                job = pending.next() => match job {
+                    Some(job) => job,
+                    None => break,
+                },
+            };
+
+            while let Some(res) = tasks.try_join_next() {
+                if let Err(err) = res {
+                    error!("Task panicked: {err}");
+                }
+            }
+
+            let permit = semaphore.clone().acquire_owned().await?;
+
+            if let Err(err) = self.queue.mark_processing(job.id).await {
+                error!(%err, "failed to mark job as processing");
+            }
+
+            let bot = self.clone();
+            metrics::increment_task_active();
+
+            tasks.spawn(async move {
+                bot.process(job).await;
+                drop(permit);
+                metrics::decrement_task_active();
+            });
+        }
+
+        tasks.join_all().await;
 
         info!("Bot shutting down complete");
         Ok(())
     }
 
+    async fn process(&self, job: Job) {
+        let event_type = job.payload.event_type_str();
+
+        match self.dedup.is_reviewed(job.event_id).await {
+            Ok(true) => {
+                metrics::task_skipped(event_type);
+                info!("Review already done, skipping");
+
+                if let Err(err) = self.queue.remove(job.id).await {
+                    error!(%err, "failed to remove job");
+                }
+                return;
+            }
+            Ok(false) => {}
+            Err(err) => error!(%err, "Dedup lookup failed, reviewing anyway"),
+        }
+
+        let payload = job.payload.0.clone();
+        let started = std::time::Instant::now();
+        let result = self.exec(payload).await;
+        let duration_ms = started.elapsed().as_millis() as u64;
+
+        let (ok, outcome) = match result {
+            Ok(outcome) => {
+                metrics::task_completed(event_type);
+                info!("Task completed");
+                (true, outcome)
+            }
+            Err(err) => {
+                metrics::task_failed(event_type);
+                error!(%err, "Task error");
+                sentry_anyhow::capture_anyhow(&err);
+                (false, ReviewOutcome::empty())
+            }
+        };
+
+        let record = ReviewRecord {
+            event_id: job.event_id,
+            repo: job.repo.clone(),
+            pr: job.pr,
+            ok,
+            model: self.open_router_model.clone(),
+            markdown: outcome.markdown,
+            cost: outcome.cost,
+            comments: outcome.comments as u64,
+            duration_ms,
+        };
+
+        if let Err(err) = self.dedup.mark_reviewed(record).await {
+            error!(%err, "failed to record review");
+        }
+
+        if let Err(err) = self.queue.remove(job.id).await {
+            error!(%err, "failed to remove job");
+        }
+    }
+
     #[instrument(skip(self, webhook), fields(repo, pr))]
-    pub async fn exec(&self, webhook: WebhookType) -> anyhow::Result<()> {
+    pub async fn exec(&self, webhook: WebhookType) -> anyhow::Result<ReviewOutcome> {
         match &webhook {
             WebhookType::Review(p) => {
                 tracing::Span::current().record("repo", &p.repository.full_name);
@@ -203,40 +290,3 @@ impl Bot {
         }
     }
 }
-
-async fn handle_review(webhook: WebhookType, bot: Data<Bot>) -> Result<(), Error> {
-    let event_type = webhook.event_type_str();
-    let review_key = Dedup::review_key(webhook.event_id());
-
-    match bot.dedup.is_marked(&review_key).await {
-        Ok(true) => {
-            metrics::task_skipped(event_type);
-            info!("Review already done, skipping");
-            return Ok(());
-        }
-        Ok(false) => {}
-        Err(err) => error!(%err, "Dedup lookup failed, reviewing anyway"),
-    }
-
-    metrics::increment_task_active();
-    let result = bot.exec(webhook).await;
-    metrics::decrement_task_active();
-
-    match result {
-        Ok(()) => {
-            metrics::task_completed(event_type);
-            info!("Task completed");
-
-            if let Err(err) = bot.dedup.mark(&review_key, REVIEW_TTL).await {
-                error!(%err, "Failed to mark review as done");
-            }
-        }
-        Err(err) => {
-            metrics::task_failed(event_type);
-            error!(%err, "Task error");
-            sentry_anyhow::capture_anyhow(&err);
-        }
-    }
-
-    Ok(())
-}

crates/steep-server/src/bot_actions/review.rs

@@ -12,6 +12,22 @@ use crate::{
     text::excerpt,
 };
 
+pub struct ReviewOutcome {
+    pub markdown: String,
+    pub cost: Option<f64>,
+    pub comments: usize,
+}
+
+impl ReviewOutcome {
+    pub fn empty() -> Self {
+        Self {
+            markdown: String::new(),
+            cost: None,
+            comments: 0,
+        }
+    }
+}
+
 #[instrument(skip(gitea_api, open_router_client, sandbox_config, tools, review_payload))]
 pub async fn exec_review(
     gitea_api: &GiteaAPI,
@@ -20,7 +36,7 @@ pub async fn exec_review(
     sandbox_config: &SandboxConfig,
     tools: Vec<Tool>,
     review_payload: ReviewPayload,
-) -> anyhow::Result<()> {
+) -> anyhow::Result<ReviewOutcome> {
     tracing::info!(
         repo = %review_payload.repository.full_name,
         pr = review_payload.pull_request.number,
@@ -36,7 +52,7 @@ pub async fn exec_review(
         )
         .await?;
 
-    let bot_result: Result<ReviewResult, anyhow::Error> = async {
+    let bot_result: Result<(ReviewResult, String), anyhow::Error> = async {
         let full_name = &review_payload.repository.full_name;
         let index = review_payload.pull_request.number;
 
@@ -91,15 +107,21 @@ pub async fn exec_review(
             )
             .await?;
 
-        Ok(review_result)
+        Ok((review_result, final_review_markdown))
     }
     .await;
 
     match bot_result {
-        Ok(_) => {
+        Ok((review_result, markdown)) => {
             gitea_api
                 .delete_comment(&review_payload.repository.full_name, new_comment.id)
-                .await
+                .await?;
+
+            Ok(ReviewOutcome {
+                markdown,
+                cost: review_result.cost,
+                comments: review_result.reviews.len(),
+            })
         }
         Err(e) => {
             gitea_api
@@ -108,17 +130,13 @@ pub async fn exec_review(
                     &review_payload.repository.full_name,
                     new_comment.id,
                 )
-                .await
+                .await?;
+
+            Err(e)
         }
     }
 }
 
-/// Runs the review inside a sandbox container, letting the model explore the
-/// repository with tools before answering.
-///
-/// The answer of the model is parsed by [`ReviewResult::from_str`], which the
-/// agent loop enforces: an answer that is not a review is sent back to the model
-/// for correction.
 async fn run_sandboxed_review(
     gitea_api: &GiteaAPI,
     open_router_client: &OpenRouterClient,

crates/steep-server/src/db.rs

@@ -0,0 +1,43 @@
+use toasty::Db;
+
+use crate::schemas::{Job, JobStatus};
+
+pub async fn open(url: &str) -> toasty::Result<Db> {
+    Db::builder()
+        .models(toasty::models!(crate::*))
+        .connect(url)
+        .await
+}
+
+pub async fn connect(url: &str) -> toasty::Result<Db> {
+    let mut db = open(url).await?;
+
+    Job::update_by_status(JobStatus::Processing)
+        .status(JobStatus::Pending)
+        .exec(&mut db)
+        .await?;
+
+    Ok(db)
+}
+
+#[cfg(test)]
+pub(crate) async fn test_connect() -> Db {
+    static DB: tokio::sync::OnceCell<Db> = tokio::sync::OnceCell::const_new();
+
+    DB.get_or_init(|| async {
+        let url = std::env::var("DATABASE_URL").expect("DATABASE_URL must be set");
+
+        let migrations = std::path::Path::new(env!("CARGO_MANIFEST_DIR")).join("toasty");
+        let config = toasty_cli::Config::new()
+            .migration(toasty_cli::MigrationConfig::new().path(migrations));
+
+        toasty_cli::ToastyCli::with_config(open(&url).await.expect("open pool"), config)
+            .parse_from(["toasty", "migration", "apply"])
+            .await
+            .expect("apply migrations");
+
+        connect(&url).await.expect("connect to postgres")
+    })
+    .await
+    .clone()
+}

crates/steep-server/src/dedup.rs

@@ -1,66 +1,71 @@
-use std::time::Duration;
+use toasty::Db;
 
-use redis::aio::ConnectionManager;
+use crate::schemas::{Job, Review};
 
-pub const WEBHOOK_TTL: Duration = Duration::from_secs(24 * 60 * 60);
-pub const REVIEW_TTL: Duration = Duration::from_secs(7 * 24 * 60 * 60);
+pub struct ReviewRecord {
+    pub event_id: u64,
+    pub repo: String,
+    pub pr: u64,
+    pub ok: bool,
+    pub model: String,
+    pub markdown: String,
+    pub cost: Option<f64>,
+    pub comments: u64,
+    pub duration_ms: u64,
+}
 
 #[derive(Clone)]
 pub struct Dedup {
-    conn: ConnectionManager,
+    db: Db,
 }
 
 impl Dedup {
-    pub fn new(conn: ConnectionManager) -> Self {
-        Self { conn }
+    pub fn new(db: Db) -> Self {
+        Self { db }
     }
 
-    pub fn webhook_key(event_id: u64) -> String {
-        format!("steep:webhook:{event_id}")
+    pub async fn is_known(&self, event_id: u64) -> toasty::Result<bool> {
+        let mut db = self.db.clone();
+
+        if Job::filter_by_event_id(event_id)
+            .first()
+            .exec(&mut db)
+            .await?
+            .is_some()
+        {
+            return Ok(true);
+        }
+
+        Ok(Review::filter_by_event_id(event_id)
+            .first()
+            .exec(&mut db)
+            .await?
+            .is_some())
     }
 
-    pub fn review_key(event_id: u64) -> String {
-        format!("steep:review:{event_id}")
+    pub async fn is_reviewed(&self, event_id: u64) -> toasty::Result<bool> {
+        let mut db = self.db.clone();
+
+        Ok(Review::filter_by_event_id(event_id)
+            .first()
+            .exec(&mut db)
+            .await?
+            .is_some())
     }
 
-    pub async fn claim(&self, key: &str, ttl: Duration) -> redis::RedisResult<bool> {
-        let set: Option<String> = redis::cmd("SET")
-            .arg(key)
-            .arg(1)
-            .arg("NX")
-            .arg("EX")
-            .arg(ttl.as_secs())
-            .query_async(&mut self.conn.clone())
-            .await?;
+    pub async fn mark_reviewed(&self, record: ReviewRecord) -> toasty::Result<()> {
+        let mut db = self.db.clone();
 
-        Ok(set.is_some())
-    }
-
-    pub async fn mark(&self, key: &str, ttl: Duration) -> redis::RedisResult<()> {
-        let _: String = redis::cmd("SET")
-            .arg(key)
-            .arg(1)
-            .arg("EX")
-            .arg(ttl.as_secs())
-            .query_async(&mut self.conn.clone())
-            .await?;
-
-        Ok(())
-    }
-
-    pub async fn is_marked(&self, key: &str) -> redis::RedisResult<bool> {
-        let exists: i64 = redis::cmd("EXISTS")
-            .arg(key)
-            .query_async(&mut self.conn.clone())
-            .await?;
-
-        Ok(exists > 0)
-    }
-
-    pub async fn release(&self, key: &str) -> redis::RedisResult<()> {
-        let _: i64 = redis::cmd("DEL")
-            .arg(key)
-            .query_async(&mut self.conn.clone())
+        Review::upsert_by_event_id(record.event_id)
+            .repo(record.repo)
+            .pr(record.pr)
+            .ok(record.ok)
+            .model(record.model)
+            .markdown(record.markdown)
+            .cost(record.cost)
+            .comments(record.comments)
+            .duration_ms(record.duration_ms)
+            .exec(&mut db)
             .await?;
 
         Ok(())
@@ -70,51 +75,42 @@ impl Dedup {
 #[cfg(test)]
 mod tests {
     use super::*;
-
-    #[test]
-    fn webhook_key_is_namespaced() {
-        assert_eq!(Dedup::webhook_key(42), "steep:webhook:42");
-    }
-
-    #[test]
-    fn review_key_is_namespaced() {
-        assert_eq!(Dedup::review_key(42), "steep:review:42");
-    }
-
-    #[test]
-    fn the_two_keys_for_the_same_event_differ() {
-        assert_ne!(Dedup::webhook_key(7), Dedup::review_key(7));
-    }
-
-    async fn redis_dedup() -> Dedup {
-        let url = std::env::var("REDIS_URL").expect("REDIS_URL must be set");
-        Dedup::new(apalis_redis::connect(url).await.expect("connect to redis"))
-    }
+    use crate::schemas::Review;
 
     #[tokio::test]
-    #[ignore = "requires a running Redis (REDIS_URL)"]
-    async fn claim_then_claim_is_a_duplicate() {
-        let dedup = redis_dedup().await;
-        let key = "steep:test:claim";
-        dedup.release(key).await.unwrap();
+    #[ignore = "requires DATABASE_URL"]
+    async fn a_recorded_review_is_known_and_reviewed() {
+        let db = crate::db::test_connect().await;
+        let dedup = Dedup::new(db.clone());
 
-        assert!(dedup.claim(key, Duration::from_secs(30)).await.unwrap());
-        assert!(!dedup.claim(key, Duration::from_secs(30)).await.unwrap());
+        let event_id = 999_000_101;
+        Review::delete_by_event_id(&mut db.clone(), event_id)
+            .await
+            .ok();
 
-        dedup.release(key).await.unwrap();
-    }
+        assert!(!dedup.is_known(event_id).await.unwrap());
+        assert!(!dedup.is_reviewed(event_id).await.unwrap());
 
-    #[tokio::test]
-    #[ignore = "requires a running Redis (REDIS_URL)"]
-    async fn mark_then_is_marked() {
-        let dedup = redis_dedup().await;
-        let key = "steep:test:mark";
-        dedup.release(key).await.unwrap();
+        dedup
+            .mark_reviewed(ReviewRecord {
+                event_id,
+                repo: "owner/repo".to_string(),
+                pr: 1,
+                ok: true,
+                model: "test/model".to_string(),
+                markdown: "looks good".to_string(),
+                cost: Some(0.01),
+                comments: 2,
+                duration_ms: 5,
+            })
+            .await
+            .unwrap();
 
-        assert!(!dedup.is_marked(key).await.unwrap());
-        dedup.mark(key, Duration::from_secs(30)).await.unwrap();
-        assert!(dedup.is_marked(key).await.unwrap());
+        assert!(dedup.is_reviewed(event_id).await.unwrap());
+        assert!(dedup.is_known(event_id).await.unwrap());
 
-        dedup.release(key).await.unwrap();
+        Review::delete_by_event_id(&mut db.clone(), event_id)
+            .await
+            .unwrap();
     }
 }

crates/steep-server/src/env.rs

@@ -11,7 +11,7 @@ pub struct EnvConfig {
     pub gitea_url: String,
     pub gitea_token: String,
     pub gitea_timeout: u64,
-    pub redis_url: String,
+    pub database_url: String,
     pub metrics_bind_addr: Option<String>,
     pub sandbox_max_iterations: usize,
 }
@@ -26,7 +26,7 @@ pub fn load_config() -> anyhow::Result<EnvConfig> {
     let gitea_url = try_get_env("GITEA_URL")?;
     let gitea_token = try_get_env("GITEA_TOKEN")?;
     let gitea_timeout = try_get_env("GITEA_TIMEOUT")?.parse()?;
-    let redis_url = try_get_env("REDIS_URL")?;
+    let database_url = try_get_env("DATABASE_URL")?;
     let metrics_bind_addr = std::env::var("METRICS_BIND_ADDR").ok();
     let sandbox_max_iterations = std::env::var("SANDBOX_MAX_ITERATIONS")
         .ok()
@@ -43,7 +43,7 @@ pub fn load_config() -> anyhow::Result<EnvConfig> {
         gitea_url,
         gitea_token,
         gitea_timeout,
-        redis_url,
+        database_url,
         metrics_bind_addr,
         sandbox_max_iterations,
     })

crates/steep-server/src/gitea.rs

@@ -281,7 +281,7 @@ async fn read_capped(response: reqwest::Response) -> anyhow::Result<String> {
     Ok(String::from_utf8_lossy(&buf).into_owned())
 }
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Serialize, Deserialize, Clone)]
 pub enum WebhookType {
     Review(ReviewPayload),
 }
@@ -298,9 +298,21 @@ impl WebhookType {
             WebhookType::Review(_) => "review",
         }
     }
+
+    pub fn repository_full_name(&self) -> &str {
+        match self {
+            WebhookType::Review(payload) => &payload.repository.full_name,
+        }
+    }
+
+    pub fn pull_request_number(&self) -> u64 {
+        match self {
+            WebhookType::Review(payload) => payload.pull_request.number,
+        }
+    }
 }
 
-#[derive(Deserialize, Serialize, Debug)]
+#[derive(Deserialize, Serialize, Debug, Clone)]
 pub struct ReviewPayload {
     pub action: String,
     pub pull_request: PullRequest,
@@ -308,13 +320,13 @@ pub struct ReviewPayload {
     pub comment: Comment,
 }
 
-#[derive(Deserialize, Serialize, Debug)]
+#[derive(Deserialize, Serialize, Debug, Clone)]
 pub struct PullRequest {
     pub number: u64,
     pub title: String,
 }
 
-#[derive(Deserialize, Serialize, Debug)]
+#[derive(Deserialize, Serialize, Debug, Clone)]
 pub struct Comment {
     pub id: u64,
     pub body: String,
@@ -325,7 +337,7 @@ pub struct User {
     pub login: String,
 }
 
-#[derive(Deserialize, Serialize, Debug)]
+#[derive(Deserialize, Serialize, Debug, Clone)]
 pub struct Repository {
     pub full_name: String,
     pub clone_url: String,

crates/steep-server/src/job_queue.rs

@@ -0,0 +1,196 @@
+use std::{collections::VecDeque, time::Duration};
+
+use futures_util::stream::{self, Stream};
+use toasty::Db;
+use tracing::error;
+
+use crate::{
+    gitea::WebhookType,
+    schemas::{Job, JobStatus},
+};
+
+/// Délai entre deux interrogations de la table quand la file est vide.
+const POLL_INTERVAL: Duration = Duration::from_secs(1);
+
+/// File des reviews à faire.
+#[derive(Clone)]
+pub struct JobQueue {
+    db: Db,
+}
+
+impl JobQueue {
+    pub fn new(db: Db) -> Self {
+        Self { db }
+    }
+
+    /// Enfile une nouvelle review.
+    ///
+    /// L'unicité de `event_id` fait échouer l'insertion d'un doublon.
+    pub async fn push(&self, webhook: WebhookType) -> toasty::Result<()> {
+        let event_id = webhook.event_id();
+        let repo = webhook.repository_full_name().to_string();
+        let pr = webhook.pull_request_number();
+
+        let mut db = self.db.clone();
+
+        toasty::create!(Job {
+            event_id,
+            repo,
+            pr,
+            payload: webhook,
+            status: JobStatus::Pending,
+        })
+        .exec(&mut db)
+        .await?;
+
+        Ok(())
+    }
+
+    /// Flux des jobs en attente.
+    ///
+    /// C'est un `Stream` (et non un `Iterator`) : l'accès à la base est
+    /// asynchrone. Le flux interroge la table en boucle et met en tampon les
+    /// jobs trouvés.
+    pub fn pending(&self) -> impl Stream<Item = Job> {
+        let db = self.db.clone();
+
+        stream::unfold(VecDeque::new(), move |mut buffer| {
+            let db = db.clone();
+
+            async move {
+                let mut db = db;
+
+                loop {
+                    if let Some(job) = buffer.pop_front() {
+                        return Some((job, buffer));
+                    }
+
+                    match Job::filter_by_status(JobStatus::Pending)
+                        .exec(&mut db)
+                        .await
+                    {
+                        Ok(jobs) if jobs.is_empty() => {
+                            tokio::time::sleep(POLL_INTERVAL).await;
+                        }
+                        Ok(jobs) => buffer.extend(jobs),
+                        Err(err) => {
+                            error!(%err, "failed to poll pending jobs");
+                            tokio::time::sleep(POLL_INTERVAL).await;
+                        }
+                    }
+                }
+            }
+        })
+    }
+
+    /// Marque un job comme en cours de traitement.
+    pub async fn mark_processing(&self, id: u64) -> toasty::Result<()> {
+        let mut db = self.db.clone();
+
+        Job::update_by_id(id)
+            .status(JobStatus::Processing)
+            .exec(&mut db)
+            .await?;
+
+        Ok(())
+    }
+
+    /// Retire un job de la file, une fois la review enregistrée.
+    pub async fn remove(&self, id: u64) -> toasty::Result<()> {
+        let mut db = self.db.clone();
+
+        Job::delete_by_id(&mut db, id).await?;
+
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use futures_util::StreamExt;
+
+    use super::*;
+    use crate::gitea::{Comment, PullRequest, Repository, ReviewPayload};
+
+    fn webhook(event_id: u64) -> WebhookType {
+        WebhookType::Review(ReviewPayload {
+            action: "created".to_string(),
+            pull_request: PullRequest {
+                number: 1,
+                title: "Test".to_string(),
+            },
+            repository: Repository {
+                full_name: "owner/repo".to_string(),
+                clone_url: "https://example.com/owner/repo.git".to_string(),
+            },
+            comment: Comment {
+                id: event_id,
+                body: "@bot review".to_string(),
+            },
+        })
+    }
+
+    async fn next_job(queue: &JobQueue, event_id: u64) -> Job {
+        let mut pending = std::pin::pin!(queue.pending());
+
+        tokio::time::timeout(Duration::from_secs(10), async {
+            loop {
+                if let Some(job) = pending.next().await
+                    && job.event_id == event_id
+                {
+                    break job;
+                }
+            }
+        })
+        .await
+        .expect("timed out waiting for the job")
+    }
+
+    #[tokio::test]
+    #[ignore = "requires DATABASE_URL"]
+    async fn a_pushed_job_is_pending_then_removed() {
+        let db = crate::db::test_connect().await;
+        let queue = JobQueue::new(db.clone());
+
+        let event_id = 999_000_201;
+        Job::delete_by_event_id(&mut db.clone(), event_id)
+            .await
+            .ok();
+
+        queue.push(webhook(event_id)).await.unwrap();
+
+        let job = next_job(&queue, event_id).await;
+        queue.mark_processing(job.id).await.unwrap();
+        queue.remove(job.id).await.unwrap();
+
+        assert!(
+            Job::filter_by_event_id(event_id)
+                .first()
+                .exec(&mut db.clone())
+                .await
+                .unwrap()
+                .is_none()
+        );
+    }
+
+    #[tokio::test]
+    #[ignore = "requires DATABASE_URL"]
+    async fn a_duplicate_event_id_is_rejected() {
+        let db = crate::db::test_connect().await;
+        let queue = JobQueue::new(db.clone());
+
+        let event_id = 999_000_202;
+        Job::delete_by_event_id(&mut db.clone(), event_id)
+            .await
+            .ok();
+
+        queue.push(webhook(event_id)).await.unwrap();
+        assert!(queue.push(webhook(event_id)).await.is_err());
+
+        Job::delete_by_event_id(&mut db.clone(), event_id)
+            .await
+            .unwrap();
+    }
+}

crates/steep-server/src/main.rs

@@ -1,12 +1,8 @@
 use std::time::Duration;
 
 use crate::{
-    bot::Bot,
-    dedup::Dedup,
-    gitea::{GiteaAPI, WebhookType},
-    open_router::OpenRouterClient,
-    sandbox::SandboxConfig,
-    state::AppState,
+    bot::Bot, dedup::Dedup, gitea::GiteaAPI, job_queue::JobQueue, open_router::OpenRouterClient,
+    sandbox::SandboxConfig, state::AppState,
 };
 
 use dotenvy::dotenv;
@@ -19,13 +15,16 @@ mod api;
 mod bot;
 mod bot_actions;
 mod consts;
+mod db;
 mod dedup;
 mod env;
 mod errors;
 mod gitea;
+mod job_queue;
 mod metrics;
 mod open_router;
 mod sandbox;
+mod schemas;
 mod state;
 mod text;
 
@@ -38,6 +37,11 @@ const SHUTDOWN_TIMEOUT: Duration = Duration::from_secs(60);
 fn main() -> anyhow::Result<()> {
     dotenv().ok();
 
+    // Sous-commande de migration : `steep-server migration <generate|apply|...>`.
+    if std::env::args().nth(1).as_deref() == Some("migration") {
+        return tokio::runtime::Runtime::new()?.block_on(run_migrations());
+    }
+
     tracing_subscriber::registry()
         .with(fmt::layer())
         .with(
@@ -63,6 +67,24 @@ fn main() -> anyhow::Result<()> {
     tokio::runtime::Runtime::new()?.block_on(run())
 }
 
+async fn run_migrations() -> anyhow::Result<()> {
+    let database_url = env::try_get_env("DATABASE_URL")?;
+    let db = db::open(&database_url).await?;
+
+    let migrations = match std::env::var("MIGRATIONS_DIR") {
+        Ok(dir) => std::path::PathBuf::from(dir),
+        Err(_) => std::path::Path::new(env!("CARGO_MANIFEST_DIR")).join("toasty"),
+    };
+    let config =
+        toasty_cli::Config::new().migration(toasty_cli::MigrationConfig::new().path(migrations));
+
+    toasty_cli::ToastyCli::with_config(db, config)
+        .parse_and_run()
+        .await?;
+
+    Ok(())
+}
+
 async fn run() -> anyhow::Result<()> {
     let config = env::load_config()?;
 
@@ -101,10 +123,9 @@ async fn run() -> anyhow::Result<()> {
         );
     }
 
-    let redis_conn = apalis_redis::connect(config.redis_url.clone()).await?;
-    let storage: apalis_redis::RedisStorage<WebhookType> =
-        apalis_redis::RedisStorage::new(redis_conn.clone());
-    let dedup = Dedup::new(redis_conn);
+    let db = db::connect(&config.database_url).await?;
+    let dedup = Dedup::new(db.clone());
+    let queue = JobQueue::new(db);
 
     let bot = Bot::new(
         gitea_api,
@@ -113,12 +134,13 @@ async fn run() -> anyhow::Result<()> {
         config.open_router_model.clone(),
         sandbox,
         dedup.clone(),
+        queue.clone(),
     );
 
     let app_state = AppState {
-        storage: storage.clone(),
-        bot_name: gitea_user.login.clone(),
+        queue,
         dedup,
+        bot_name: gitea_user.login.clone(),
         config,
     };
 
@@ -143,7 +165,7 @@ async fn run() -> anyhow::Result<()> {
     tokio::select! {
         result = async {
             tokio::try_join!(
-                bot.start(storage, shutdown.clone()),
+                bot.start(shutdown.clone()),
                 api::start(app_state, shutdown.clone()),
                 signal
             )

crates/steep-server/src/metrics.rs

@@ -54,7 +54,7 @@ pub fn describe() {
     describe_counter!(
         "steep_webhooks_enqueue_failed_total",
         Unit::Count,
-        "Webhooks dropped because enqueueing to Redis failed"
+        "Webhooks dropped because enqueueing to PostgreSQL failed"
     );
     describe_gauge!(
         "steep_bot_tasks_active",

crates/steep-server/src/schemas/job.rs

@@ -0,0 +1,31 @@
+use jiff::Timestamp;
+
+use crate::gitea::WebhookType;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, toasty::Embed)]
+pub enum JobStatus {
+    Pending,
+    Processing,
+}
+
+#[derive(Debug, toasty::Model)]
+pub struct Job {
+    #[key]
+    #[auto]
+    pub id: u64,
+
+    #[unique]
+    pub event_id: u64,
+
+    pub repo: String,
+    pub pr: u64,
+
+    #[column(type = jsonb)]
+    pub payload: toasty::Json<WebhookType>,
+
+    #[index]
+    pub status: JobStatus,
+
+    #[auto]
+    pub created_at: Timestamp,
+}

crates/steep-server/src/schemas/mod.rs

@@ -0,0 +1,5 @@
+pub mod job;
+pub mod review;
+
+pub use job::{Job, JobStatus};
+pub use review::Review;

crates/steep-server/src/schemas/review.rs

@@ -0,0 +1,35 @@
+use jiff::Timestamp;
+
+#[derive(Debug, toasty::Model)]
+pub struct Review {
+    #[key]
+    #[auto]
+    pub id: u64,
+
+    #[unique]
+    pub event_id: u64,
+
+    pub repo: String,
+    pub pr: u64,
+
+    /// Vrai si la review a abouti.
+    pub ok: bool,
+
+    /// Modèle utilisé.
+    pub model: String,
+
+    /// Markdown posté sur la pull request (vide en cas d'échec).
+    pub markdown: String,
+
+    /// Coût OpenRouter en USD, si connu.
+    pub cost: Option<f64>,
+
+    /// Nombre de commentaires produits.
+    pub comments: u64,
+
+    /// Durée du traitement en millisecondes.
+    pub duration_ms: u64,
+
+    #[default(jiff::Timestamp::now())]
+    pub completed_at: Timestamp,
+}

crates/steep-server/src/state.rs

@@ -1,11 +1,9 @@
-use apalis_redis::RedisStorage;
-
-use crate::{dedup::Dedup, env::EnvConfig, gitea::WebhookType};
+use crate::{dedup::Dedup, env::EnvConfig, job_queue::JobQueue};
 
 #[derive(Clone)]
 pub struct AppState {
-    pub storage: RedisStorage<WebhookType>,
-    pub bot_name: String,
+    pub queue: JobQueue,
     pub dedup: Dedup,
+    pub bot_name: String,
     pub config: EnvConfig,
 }

crates/steep-server/toasty/history.toml

@@ -0,0 +1,6 @@
+version = 1
+
+[[migrations]]
+id = 3569579561409902287
+name = "0000_init.sql"
+snapshot_name = "0000_snapshot.toml"

crates/steep-server/toasty/migrations/0000_init.sql

@@ -0,0 +1,28 @@
+CREATE TYPE "job_status" AS ENUM ('pending', 'processing');
+CREATE TABLE "jobs" (
+    "id" BIGINT NOT NULL GENERATED BY DEFAULT AS IDENTITY,
+    "event_id" BIGINT NOT NULL,
+    "repo" TEXT NOT NULL,
+    "pr" BIGINT NOT NULL,
+    "payload" JSONB NOT NULL,
+    "status" "job_status" NOT NULL,
+    "created_at" TIMESTAMPTZ(6) NOT NULL,
+    PRIMARY KEY ("id")
+);
+CREATE UNIQUE INDEX "index_jobs_by_event_id" ON "jobs" ("event_id");
+CREATE INDEX "index_jobs_by_status" ON "jobs" ("status");
+CREATE TABLE "reviews" (
+    "id" BIGINT NOT NULL GENERATED BY DEFAULT AS IDENTITY,
+    "event_id" BIGINT NOT NULL,
+    "repo" TEXT NOT NULL,
+    "pr" BIGINT NOT NULL,
+    "ok" BOOLEAN NOT NULL,
+    "model" TEXT NOT NULL,
+    "markdown" TEXT NOT NULL,
+    "cost" DOUBLE PRECISION,
+    "comments" BIGINT NOT NULL,
+    "duration_ms" BIGINT NOT NULL,
+    "completed_at" TIMESTAMPTZ(6) NOT NULL,
+    PRIMARY KEY ("id")
+);
+CREATE UNIQUE INDEX "index_reviews_by_event_id" ON "reviews" ("event_id");

crates/steep-server/toasty/snapshots/0000_snapshot.toml

@@ -0,0 +1,160 @@
+version = 1
+
+[schema]
+
+[[schema.tables]]
+id = 0
+name = "jobs"
+primary_key = { columns = [{ table = 0, index = 0 }], index = { table = 0, index = 0 } }
+
+[[schema.tables.columns]]
+id = { table = 0, index = 0 }
+name = "id"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+auto_increment = true
+
+[[schema.tables.columns]]
+id = { table = 0, index = 1 }
+name = "event_id"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+
+[[schema.tables.columns]]
+id = { table = 0, index = 2 }
+name = "repo"
+ty = "String"
+storage_ty = "Text"
+
+[[schema.tables.columns]]
+id = { table = 0, index = 3 }
+name = "pr"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+
+[[schema.tables.columns]]
+id = { table = 0, index = 4 }
+name = "payload"
+ty = "String"
+storage_ty = "Jsonb"
+
+[[schema.tables.columns]]
+id = { table = 0, index = 5 }
+name = "status"
+ty = "String"
+storage_ty = { Enum = { name = "job_status", variants = [{ name = "pending" }, { name = "processing" }] } }
+
+[[schema.tables.columns]]
+id = { table = 0, index = 6 }
+name = "created_at"
+ty = "Timestamp"
+storage_ty = { Timestamp = 6 }
+
+[[schema.tables.indices]]
+id = { table = 0, index = 0 }
+name = "index_jobs_by_id"
+on = 0
+columns = [{ column = { table = 0, index = 0 }, op = "Eq", scope = "Partition" }]
+unique = true
+primary_key = true
+
+[[schema.tables.indices]]
+id = { table = 0, index = 1 }
+name = "index_jobs_by_event_id"
+on = 0
+columns = [{ column = { table = 0, index = 1 }, op = "Eq", scope = "Partition" }]
+unique = true
+
+[[schema.tables.indices]]
+id = { table = 0, index = 2 }
+name = "index_jobs_by_status"
+on = 0
+columns = [{ column = { table = 0, index = 5 }, op = "Eq", scope = "Partition" }]
+
+[[schema.tables]]
+id = 1
+name = "reviews"
+primary_key = { columns = [{ table = 1, index = 0 }], index = { table = 1, index = 0 } }
+
+[[schema.tables.columns]]
+id = { table = 1, index = 0 }
+name = "id"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+auto_increment = true
+
+[[schema.tables.columns]]
+id = { table = 1, index = 1 }
+name = "event_id"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+
+[[schema.tables.columns]]
+id = { table = 1, index = 2 }
+name = "repo"
+ty = "String"
+storage_ty = "Text"
+
+[[schema.tables.columns]]
+id = { table = 1, index = 3 }
+name = "pr"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+
+[[schema.tables.columns]]
+id = { table = 1, index = 4 }
+name = "ok"
+ty = "Bool"
+storage_ty = "Boolean"
+
+[[schema.tables.columns]]
+id = { table = 1, index = 5 }
+name = "model"
+ty = "String"
+storage_ty = "Text"
+
+[[schema.tables.columns]]
+id = { table = 1, index = 6 }
+name = "markdown"
+ty = "String"
+storage_ty = "Text"
+
+[[schema.tables.columns]]
+id = { table = 1, index = 7 }
+name = "cost"
+ty = "F64"
+storage_ty = { Float = 8 }
+nullable = true
+
+[[schema.tables.columns]]
+id = { table = 1, index = 8 }
+name = "comments"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+
+[[schema.tables.columns]]
+id = { table = 1, index = 9 }
+name = "duration_ms"
+ty = "U64"
+storage_ty = { UnsignedInteger = 8 }
+
+[[schema.tables.columns]]
+id = { table = 1, index = 10 }
+name = "completed_at"
+ty = "Timestamp"
+storage_ty = { Timestamp = 6 }
+
+[[schema.tables.indices]]
+id = { table = 1, index = 0 }
+name = "index_reviews_by_id"
+on = 1
+columns = [{ column = { table = 1, index = 0 }, op = "Eq", scope = "Partition" }]
+unique = true
+primary_key = true
+
+[[schema.tables.indices]]
+id = { table = 1, index = 1 }
+name = "index_reviews_by_event_id"
+on = 1
+columns = [{ column = { table = 1, index = 1 }, op = "Eq", scope = "Partition" }]
+unique = true