fix: resolve executor timeout due to lock contention in scheduler event loop

The metrics instrumentation PR introduced a hot-path call to total_pending_tasks()
after every scheduler event. This method acquires read locks on all execution graphs,
which combined with concurrent poll_work calls (acquiring write locks), caused
deadlock-like behavior due to tokio RwLock fairness semantics.

Changes:
- Remove total_pending_tasks() from event loop hot path
- Add periodic background task (5s interval) for pending_tasks metric
- Remove diagnostic logging and unnecessary timeouts
- Clean up unused imports

Author: phillipleblanc

ballista/scheduler/src/cluster/mod.rs

@@ -447,6 +447,7 @@ pub(crate) async fn bind_task_bias(
             continue;
         }
         let mut graph = job_info.execution_graph.write().await;
+
         let session_id = graph.session_id().to_string();
         let mut black_list = vec![];
         while let Some((running_stage, task_id_gen)) =
@@ -548,6 +549,7 @@ pub(crate) async fn bind_task_round_robin(
             continue;
         }
         let mut graph = job_info.execution_graph.write().await;
+
         let session_id = graph.session_id().to_string();
         let mut black_list = vec![];
         while let Some((running_stage, task_id_gen)) =

ballista/scheduler/src/scheduler_server/grpc.rs

@@ -118,6 +118,7 @@ impl<T: 'static + AsLogicalPlan, U: 'static + AsExecutionPlan> SchedulerGrpc
             }];
             let available_slots = available_slots.iter_mut().collect();
             let running_jobs = self.state.task_manager.get_running_job_cache();
+
             let binding_result = match self.state.config.task_distribution {
                 TaskDistributionPolicy::Bias => {
                     bind_task_bias(available_slots, running_jobs, |_| false).await

ballista/scheduler/src/scheduler_server/mod.rs

@@ -169,6 +169,7 @@ impl<T: 'static + AsLogicalPlan, U: 'static + AsExecutionPlan> SchedulerServer<T
         self.state.init().await?;
         self.query_stage_event_loop.start()?;
         self.expire_dead_executors()?;
+        self.start_pending_tasks_metrics_loop();
 
         Ok(())
     }
@@ -339,6 +340,30 @@ impl<T: 'static + AsLogicalPlan, U: 'static + AsExecutionPlan> SchedulerServer<T
         Ok(())
     }
 
+    /// Spawns a background task that periodically updates the pending tasks metric.
+    ///
+    /// This metric requires iterating over all active jobs and acquiring read locks,
+    /// which can cause lock contention if done in the main event loop. Running it
+    /// periodically in a background task provides observability without impacting
+    /// scheduler performance.
+    fn start_pending_tasks_metrics_loop(&self) {
+        let state = self.state.clone();
+        tokio::task::spawn(async move {
+            // Update every 5 seconds - frequent enough for observability,
+            // infrequent enough to avoid lock contention
+            const UPDATE_INTERVAL: Duration = Duration::from_secs(5);
+
+            loop {
+                let pending_tasks = state.task_manager.total_pending_tasks().await;
+                state
+                    .metrics_collector
+                    .set_pending_tasks_queue_size(pending_tasks as u64);
+
+                tokio::time::sleep(UPDATE_INTERVAL).await;
+            }
+        });
+    }
+
     pub(crate) fn remove_executor(
         executor_manager: ExecutorManager,
         event_sender: EventSender<QueryStageSchedulerEvent>,

ballista/scheduler/src/scheduler_server/query_stage_scheduler.rs

@@ -324,14 +324,20 @@ impl<T: 'static + AsLogicalPlan, U: 'static + AsExecutionPlan>
             }
         }
 
-        // Update queue size metrics after processing each event
+        // Update pending jobs queue size metric (this is a cheap O(1) operation)
         let pending_jobs = self.state.task_manager.pending_job_number();
         self.metrics_collector
             .set_pending_jobs_queue_size(pending_jobs as u64);
 
-        let pending_tasks = self.state.task_manager.total_pending_tasks().await;
-        self.metrics_collector
-            .set_pending_tasks_queue_size(pending_tasks as u64);
+        // NOTE: We intentionally do NOT call total_pending_tasks() here.
+        // That method iterates over all active jobs and acquires read locks on each
+        // execution graph, which causes lock contention with concurrent poll_work
+        // calls that acquire write locks during task binding. This was causing
+        // deadlocks and executor timeouts under load.
+        //
+        // The pending_jobs metric above provides sufficient visibility into queue depth.
+        // If detailed pending task counts are needed, they should be collected
+        // periodically in a background task, not on every event.
 
         Ok(())
     }

ballista/scheduler/src/state/task_manager.rs

@@ -266,6 +266,11 @@ impl<T: 'static + AsLogicalPlan, U: 'static + AsExecutionPlan> TaskManager<T, U>
     ///
     /// A pending task is a task that is available to schedule on an executor
     /// but cannot be scheduled because no resources are available.
+    ///
+    /// NOTE: This method iterates over all active jobs and acquires read locks
+    /// on each execution graph. It should NOT be called frequently (e.g., in a
+    /// hot loop or after every event) as it can cause lock contention with
+    /// concurrent task binding operations.
     pub async fn total_pending_tasks(&self) -> usize {
         let mut total = 0;
         for entry in self.active_job_cache.iter() {