From ebdc8c19da54756a0969e2b2c624f2eae3ea8aff Mon Sep 17 00:00:00 2001
From: Stefan Boberg <stefan@boberg.org>
Date: Fri, 27 Feb 2026 12:05:41 +0100
Subject: made local action submits more async by introducing "submitting"
 state

---
 src/zencompute/localrunner.cpp | 57 +++++++++++++++++++++++++++++-------------
 1 file changed, 40 insertions(+), 17 deletions(-)

(limited to 'src/zencompute/localrunner.cpp')

diff --git a/src/zencompute/localrunner.cpp b/src/zencompute/localrunner.cpp
index c167b6055..ceb1c19d4 100644
--- a/src/zencompute/localrunner.cpp
+++ b/src/zencompute/localrunner.cpp
@@ -181,14 +181,16 @@ LocalProcessRunner::QueryCapacity()
 		return 0;
 	}
 
-	size_t RunningCount = m_RunningMap.size();
+	const size_t InFlightCount = m_RunningMap.size() + m_SubmittingCount.load(std::memory_order_relaxed);
 
-	if (RunningCount >= size_t(m_MaxRunningActions))
+	if (const size_t MaxRunningActions = m_MaxRunningActions; InFlightCount >= MaxRunningActions)
 	{
 		return 0;
 	}
-
-	return m_MaxRunningActions - RunningCount;
+	else
+	{
+		return MaxRunningActions - InFlightCount;
+	}
 }
 
 std::vector<SubmitResult>
@@ -206,26 +208,47 @@ LocalProcessRunner::SubmitActions(const std::vector<Ref<RunnerAction>>& Actions)
 		return Results;
 	}
 
-	// For nontrivial batches, submit actions in parallel via the worker pool.
-	// PrepareActionSubmission (sandbox creation, worker manifesting, input
-	// writing) is heavily I/O-bound and benefits from overlapping.
+	// For nontrivial batches, check capacity upfront and accept what fits.
+	// Accepted actions are transitioned to Submitting and dispatched to the
+	// worker pool as fire-and-forget, so SubmitActions returns immediately
+	// and the scheduler thread is free to handle completions and updates.
 
-	std::vector<std::future<SubmitResult>> Futures;
-	Futures.reserve(Actions.size());
+	size_t Available = QueryCapacity();
 
-	for (const Ref<RunnerAction>& Action : Actions)
+	std::vector<SubmitResult> Results(Actions.size());
+
+	size_t AcceptCount = std::min(Available, Actions.size());
+
+	for (size_t i = 0; i < AcceptCount; ++i)
 	{
-		std::packaged_task<SubmitResult()> Task([this, Action]() { return SubmitAction(Action); });
+		const Ref<RunnerAction>& Action = Actions[i];
 
-		Futures.push_back(m_WorkerPool.EnqueueTask(std::move(Task), WorkerThreadPool::EMode::EnableBacklog));
-	}
+		Action->SetActionState(RunnerAction::State::Submitting);
+		m_SubmittingCount.fetch_add(1, std::memory_order_relaxed);
 
-	std::vector<SubmitResult> Results;
-	Results.reserve(Futures.size());
+		Results[i] = SubmitResult{.IsAccepted = true};
+
+		m_WorkerPool.ScheduleWork(
+			[this, Action]() {
+				auto CountGuard = MakeGuard([this] { m_SubmittingCount.fetch_sub(1, std::memory_order_relaxed); });
+
+				SubmitResult Result = SubmitAction(Action);
+
+				if (!Result.IsAccepted)
+				{
+					// This might require another state? We should
+					// distinguish between outright rejections (e.g. invalid action)
+					// and transient failures (e.g. failed to launch process) which might
+					// be retried by the scheduler, but for now just fail the action
+					Action->SetActionState(RunnerAction::State::Failed);
+				}
+			},
+			WorkerThreadPool::EMode::EnableBacklog);
+	}
 
-	for (auto& Future : Futures)
+	for (size_t i = AcceptCount; i < Actions.size(); ++i)
 	{
-		Results.push_back(Future.get());
+		Results[i] = SubmitResult{.IsAccepted = false};
 	}
 
 	return Results;
-- 
cgit v1.2.3