diff --git a/backend/src/db_connect.rs b/backend/src/db_connect.rs
index d4dee19f82..b18f675e49 100644
--- a/backend/src/db_connect.rs
+++ b/backend/src/db_connect.rs
@@ -153,6 +153,15 @@ pub async fn connect(
         pool_options = pool_options.idle_timeout(Duration::from_secs(60));
     }
     pool_options
+        // Clears transaction state sqlx is not tracking, so a connection whose session was
+        // left inside a transaction is not handed to the next borrower. See
+        // `windmill_common::db::connection_reset` for how such a session comes about and why
+        // this only runs once one has been observed.
+        .after_release(|conn, _| {
+            Box::pin(windmill_common::db::connection_reset::reset_on_release(
+                conn,
+            ))
+        })
         .after_connect(move |conn, _| {
             if worker_mode {
                 Box::pin(async move {
diff --git a/backend/windmill-common/src/db.rs b/backend/windmill-common/src/db.rs
index 0048abf438..6f62038c1c 100644
--- a/backend/windmill-common/src/db.rs
+++ b/backend/windmill-common/src/db.rs
@@ -290,3 +290,76 @@ impl<'b> DbExecutor<'b> for &'b mut PgConnection {
         &mut **self
     }
 }
+
+/// Guards the pool against a Postgres session left inside a transaction sqlx is not
+/// tracking.
+///
+/// sqlx only queues its rollback-on-drop once `begin()` has returned, so a future
+/// cancelled while `BEGIN` is still in flight — a disconnecting API client, a
+/// `tokio::time::timeout`, an aborted task — hands the connection back with the
+/// transaction still open server-side. sqlx's on-release `ping` is a bare
+/// `wait_until_ready` that reports such a connection as healthy, so it is reused: the
+/// next borrower's statements silently run inside that transaction and hold their locks
+/// for as long as it lives, and the first error turns the session into
+/// `idle in transaction (aborted)`, after which every unrelated query on it fails with
+/// `25P02` until `max_lifetime` recycles it half an hour later.
+///
+/// Rolling back on every release would add a round trip to every query, which measured at
+/// about a third of the throughput on small ones, so the reset is armed only once Postgres
+/// has reported a state that proves a connection is carrying such leftover state.
+pub mod connection_reset {
+    use std::sync::atomic::{AtomicU64, Ordering};
+    use std::sync::LazyLock;
+    use std::time::{Duration, Instant};
+
+    static PROCESS_START: LazyLock<Instant> = LazyLock::new(Instant::now);
+    /// Milliseconds since `PROCESS_START` until which releases roll back; 0 = never armed.
+    static RESET_UNTIL_MS: AtomicU64 = AtomicU64::new(0);
+
+    /// Long enough to outlast the connections that were already checked out when the
+    /// first poisoned one surfaced; one release cycle clears the pool.
+    const RESET_WINDOW: Duration = Duration::from_secs(60);
+
+    fn now_ms() -> u64 {
+        PROCESS_START.elapsed().as_millis() as u64
+    }
+
+    fn arm() {
+        RESET_UNTIL_MS.fetch_max(
+            now_ms() + RESET_WINDOW.as_millis() as u64,
+            Ordering::Relaxed,
+        );
+    }
+
+    /// Reading the clock is skipped entirely while the pool has never been armed, which
+    /// is the steady state.
+    fn armed() -> bool {
+        let until = RESET_UNTIL_MS.load(Ordering::Relaxed);
+        until != 0 && until > now_ms()
+    }
+
+    /// Body of the pool's `after_release` hook. `ROLLBACK` ends both a leaked-open and an
+    /// aborted transaction, and is a no-op warning on a session that has neither. Reporting
+    /// the failure makes sqlx discard the connection, which is also what a session we could
+    /// not clean deserves.
+    pub async fn reset_on_release(conn: &mut sqlx::PgConnection) -> Result<bool, sqlx::Error> {
+        if armed() {
+            use sqlx::Executor;
+            conn.execute("ROLLBACK").await?;
+        }
+        Ok(true)
+    }
+
+    /// Arms the reset when an error proves a session is stuck in an aborted transaction.
+    /// Called from `Error`'s `From<sqlx::Error>`, so it sees every query that reports one.
+    pub(crate) fn note_sqlx_error(err: &sqlx::Error) {
+        let sqlx::Error::Database(db_err) = err else {
+            return;
+        };
+        // 25P02 `in_failed_sql_transaction`: the connection this query ran on is sitting
+        // in a failed transaction block that nothing in sqlx will ever roll back.
+        if db_err.code().as_deref() == Some("25P02") {
+            arm();
+        }
+    }
+}
diff --git a/backend/windmill-common/src/error.rs b/backend/windmill-common/src/error.rs
index 0d5ebdc2c3..7ee8ee21cf 100644
--- a/backend/windmill-common/src/error.rs
+++ b/backend/windmill-common/src/error.rs
@@ -153,6 +153,7 @@ impl From<anyhow::Error> for Error {
 impl From<sqlx::Error> for Error {
     #[track_caller]
     fn from(e: sqlx::Error) -> Self {
+        crate::db::connection_reset::note_sqlx_error(&e);
         Self::SqlErr { error: e, location: prettify_location(std::panic::Location::caller()) }
     }
 }
diff --git a/backend/windmill-common/tests/connection_reset.rs b/backend/windmill-common/tests/connection_reset.rs
new file mode 100644
index 0000000000..0e5d8b6e95
--- /dev/null
+++ b/backend/windmill-common/tests/connection_reset.rs
@@ -0,0 +1,45 @@
+//! A pooled connection whose session is left inside a transaction sqlx is not tracking stays
+//! broken for every later borrower — `Rollback::drop` only fires while sqlx's own transaction
+//! depth is non-zero, and the on-release `ping` is a bare `wait_until_ready` that reports such
+//! a session as healthy. Before the reset hook this cost half an hour of unrelated `25P02`
+//! failures across the whole process, until `max_lifetime` recycled the connection.
+
+use sqlx::{Executor, Pool, Postgres};
+use windmill_common::db::connection_reset;
+use windmill_common::error::Error;
+
+#[sqlx::test(migrations = "../migrations")]
+async fn poisoned_connection_is_reset_before_being_handed_out_again(db: Pool<Postgres>) {
+    // One connection, so every query below lands on the same session.
+    let pool = sqlx::postgres::PgPoolOptions::new()
+        .max_connections(1)
+        .min_connections(0)
+        .after_release(|conn, _| Box::pin(connection_reset::reset_on_release(conn)))
+        .connect_with((*db.connect_options()).clone())
+        .await
+        .expect("failed to build pool");
+
+    // A batch that fails between BEGIN and COMMIT never reaches the COMMIT, leaving the
+    // session `idle in transaction (aborted)` while sqlx still believes it is not in a
+    // transaction — the same state a future cancelled during `begin()` leaves behind.
+    pool.execute("BEGIN; SELECT 1/0; COMMIT;")
+        .await
+        .expect_err("the batch must fail");
+
+    let err = sqlx::query_scalar::<_, i32>("SELECT 1")
+        .fetch_one(&pool)
+        .await
+        .expect_err("the next borrower inherits the aborted transaction");
+    assert_eq!(
+        err.as_database_error().and_then(|e| e.code()).as_deref(),
+        Some("25P02"),
+    );
+    // Converting the error is what arms the reset in production, where every `?` on a
+    // sqlx result goes through this same `From`.
+    let _ = Error::from(err);
+
+    sqlx::query_scalar::<_, i32>("SELECT 1")
+        .fetch_one(&pool)
+        .await
+        .expect("pool must hand out a usable connection again");
+}