improve message reliability

rust/src/api/messages/incoming/messages.rs

@@ -26,6 +26,39 @@ use std::sync::LazyLock;
 static ALREADY_QUEUED_RECEIPTS: LazyLock<Mutex<HashMap<String, std::time::Instant>>> =
     LazyLock::new(|| Mutex::new(HashMap::new()));
 
+/// How long a receipt the server has already taken is kept out of the queue.
+#[cfg(not(debug_assertions))]
+const RESEND_SUPPRESSION: std::time::Duration = std::time::Duration::from_secs(120);
+
+/// Whether the server accepted this receipt inside the suppression window.
+///
+/// Every inbound message marks a peer's whole outbox for retry, so a chat busy
+/// enough to keep flushing would otherwise re-send the same envelope on every
+/// message that arrives.
+#[cfg(not(debug_assertions))]
+fn recently_sent(receipt_id: &str) -> bool {
+    let mut sent = ALREADY_QUEUED_RECEIPTS
+        .lock()
+        .unwrap_or_else(PoisonError::into_inner);
+    sent.retain(|_, time| time.elapsed() < RESEND_SUPPRESSION);
+    sent.contains_key(receipt_id)
+}
+
+/// Starts the suppression window, once the server has actually taken the
+/// envelope.
+///
+/// Recorded on success rather than on the attempt: a send that fails has not
+/// reached anyone, and locking it out for two minutes turned a transient
+/// failure -- one dropped socket, one timed-out request -- into an outbox that
+/// stayed stuck long after the connection came back.
+#[cfg(not(debug_assertions))]
+fn note_sent(receipt_id: &str) {
+    ALREADY_QUEUED_RECEIPTS
+        .lock()
+        .unwrap_or_else(PoisonError::into_inner)
+        .insert(receipt_id.to_owned(), std::time::Instant::now());
+}
+
 /// Coalesces overlapping flushes of one account's receipt queue.
 ///
 /// Every committed inbound message asks for a flush, and a flush scans the
@@ -502,19 +535,12 @@ pub(crate) async fn prepare_queued_receipt(
 
 pub(crate) async fn send_queued_receipt(ctx: &Arc<Context>, receipt_id: &str) -> Result<()> {
     #[cfg(not(debug_assertions))]
-    {
+    if recently_sent(receipt_id) {
-        let mut locks = ALREADY_QUEUED_RECEIPTS.lock().unwrap();
-        if let Some(time) = locks.get(receipt_id) {
-            if time.elapsed() < std::time::Duration::from_secs(120) {
         tracing::info!(
             "Blocking queued receipt, as it was already sent within the last 120 seconds"
         );
         return Ok(());
     }
-        }
-        locks.insert(receipt_id.to_owned(), std::time::Instant::now());
-        locks.retain(|_, time| time.elapsed() < std::time::Duration::from_secs(120));
-    }
 
     let app_db = ctx.app_db.read().await.clone();
     let Some(row) = load_queued_receipt_row(&app_db.pool, receipt_id).await? else {
@@ -599,6 +625,9 @@ pub(crate) async fn send_queued_receipt(ctx: &Arc<Context>, receipt_id: &str) ->
         }
     }
 
+    #[cfg(not(debug_assertions))]
+    note_sent(receipt_id);
+
     let mut t = app_db.pool.begin().await?;
     if let Some(message_id) = receipt.message_id {
         sqlx::query!(

rust/src/api/runtime/client.rs

@@ -160,6 +160,27 @@ impl ApiClient {
                 if let Err(error) = crate::api::Server::request_pending_messages(&context).await {
                     tracing::warn!("mailbox catch-up request failed: {error}");
                 }
+
+                // The outbox needs the same safety net as the mailbox. Every
+                // other flush is tied to an event -- authenticating, a message
+                // arriving, a background job -- so on a connection that stays
+                // up while nothing comes in, a send that failed once had
+                // nothing left to retry it. A one-way conversation is exactly
+                // the case with no inbound message to trigger the sweep.
+                //
+                // Not in a notification worker: its whole budget belongs to the
+                // alert, and it must not spend the one app-database connection
+                // on the outbox while a drain is waiting for it.
+                if !context.is_notification_runtime() {
+                    if let Err(error) =
+                        crate::api::messages::incoming::messages::retransmit_queued_receipts(
+                            &context,
+                        )
+                        .await
+                    {
+                        tracing::warn!("outbox catch-up flush failed: {error}");
+                    }
+                }
             }
         });
     }

rust/src/api/runtime/request.rs

@@ -277,17 +277,20 @@ impl ApiClient {
             if let Some(contact_id) = contact_id {
                 let context = self.context.upgrade().ok_or(TwonlyError::Initialization)?;
                 let database = context.app_db.read().await.clone();
-                let mut transaction = database.pool.begin().await?;
+                // Only the mark, never the queue. `account_deleted` is a latch
+                // any contact-scoped request can set, and this one may have
+                // been about something else entirely -- or the peer may since
+                // have re-registered. `send_queued_receipt` asks the server
+                // again before it believes the latch, and drops only what the
+                // server positively denies, so emptying the outbox here threw
+                // away the very messages that path exists to save: unsent,
+                // unacknowledged and unlogged.
                 sqlx::query!(
                     "UPDATE contacts SET account_deleted = 1 WHERE user_id = ?",
                     contact_id
                 )
-                .execute(&mut *transaction)
+                .execute(&database.pool)
                 .await?;
-                sqlx::query!("DELETE FROM receipts WHERE contact_id = ?", contact_id)
-                    .execute(&mut *transaction)
-                    .await?;
-                transaction.commit().await?;
             }
         }
         Ok(())