From 90277ceed52b6e79c48b6872ecadcb007afe61aa Mon Sep 17 00:00:00 2001
From: Anton <ustyugov@gmail.com>
Date: Sat, 21 Mar 2026 23:18:42 +0200
Subject: [PATCH] admin/plugin: migrate inline job details asynchronously to
 avoid slow startup (#8721)

loadPersistedMonitorState performed a backward-compatibility migration that
wrote every job with inline rich detail fields to a dedicated per-job detail
file synchronously during startup. On deployments with many historical jobs
(e.g. 1000+) stored on distributed block storage (e.g. Longhorn), each
individual file write requires an fsync round-trip, making startup
disproportionately slow and causing readiness/liveness probe failures.

The in-memory state is populated correctly before the goroutine is started
because stripTrackedJobDetailFields is still called in-place; only the disk
writes are deferred. A completion log message at V(1) is emitted once the
background migration finishes.

Co-authored-by: Anton Ustyugov <anton@devops>
---
 weed/admin/plugin/plugin_monitor.go | 32 ++++++++++++++++++++++++++---
 1 file changed, 29 insertions(+), 3 deletions(-)

diff --git a/weed/admin/plugin/plugin_monitor.go b/weed/admin/plugin/plugin_monitor.go
index 332738a0d..73202d0a9 100644
--- a/weed/admin/plugin/plugin_monitor.go
+++ b/weed/admin/plugin/plugin_monitor.go
@@ -63,6 +63,12 @@ func (r *Plugin) loadPersistedMonitorState() error {
 	}
 
 	if len(trackedJobs) > 0 {
+		// Collect jobs that still carry inline rich details (old format) so we
+		// can migrate them to dedicated per-job detail files asynchronously.
+		// Writing many files synchronously during startup can be extremely slow
+		// on distributed storage (e.g. Longhorn), so we defer those writes.
+		var jobsToMigrate []TrackedJob
+
 		r.jobsMu.Lock()
 		for i := range trackedJobs {
 			job := trackedJobs[i]
@@ -78,10 +84,9 @@ func (r *Plugin) loadPersistedMonitorState() error {
 			}
 			// Backward compatibility: migrate older inline detail payloads
 			// out of tracked_jobs.json into dedicated per-job detail files.
+			// Collect for async migration instead of writing synchronously.
 			if hasTrackedJobRichDetails(job) {
-				if err := r.store.SaveJobDetail(job); err != nil {
-					glog.Warningf("Plugin failed to migrate detail snapshot for job %s: %v", job.JobID, err)
-				}
+				jobsToMigrate = append(jobsToMigrate, job)
 			}
 			stripTrackedJobDetailFields(&job)
 			jobCopy := job
@@ -89,6 +94,10 @@ func (r *Plugin) loadPersistedMonitorState() error {
 		}
 		r.pruneTrackedJobsLocked()
 		r.jobsMu.Unlock()
+
+		if len(jobsToMigrate) > 0 {
+			go r.migrateInlineJobDetails(jobsToMigrate)
+		}
 	}
 
 	if len(activities) > maxActivityRecords {
@@ -103,6 +112,23 @@ func (r *Plugin) loadPersistedMonitorState() error {
 	return nil
 }
 
+// migrateInlineJobDetails writes inline detail payloads from old-format
+// tracked_jobs.json entries to dedicated per-job detail files. It is called
+// asynchronously during startup to avoid blocking plugin initialisation.
+func (r *Plugin) migrateInlineJobDetails(jobs []TrackedJob) {
+	migrated := 0
+	for _, job := range jobs {
+		if err := r.store.SaveJobDetail(job); err != nil {
+			glog.Warningf("Plugin failed to migrate detail snapshot for job %s: %v", job.JobID, err)
+		} else {
+			migrated++
+		}
+	}
+	if migrated > 0 {
+		glog.V(1).Infof("Plugin migrated %d inline job details to dedicated files", migrated)
+	}
+}
+
 // ExpireJob marks an active job as failed so it no longer blocks scheduling.
 func (r *Plugin) ExpireJob(jobID, reason string) (*TrackedJob, bool, error) {
 	normalizedJobID := strings.TrimSpace(jobID)