feat: wait for checks before shutdown

canary-checker.properties

@@ -6,5 +6,6 @@
 
 # topology.runNow=true
 log.level.db=warn
+# check.concurrency=100
 
 # jobs.ComponentRelationshipSync.runNow=true

cmd/operator.go

@@ -8,6 +8,7 @@ import (
 	apicontext "github.com/flanksource/canary-checker/api/context"
 	"github.com/flanksource/canary-checker/pkg/cache"
 	"github.com/flanksource/canary-checker/pkg/jobs"
+	"github.com/flanksource/canary-checker/pkg/jobs/canary"
 	canaryJobs "github.com/flanksource/canary-checker/pkg/jobs/canary"
 	"github.com/flanksource/canary-checker/pkg/runner"
 	"github.com/flanksource/canary-checker/pkg/utils"
@@ -96,6 +97,12 @@ func run() error {
 		// so we use a goroutine to unblock server start
 		// to prevent health check from failing
 		go jobs.Start()
+
+		// TODO: stop the cron scheduler so that no more checks are scheduled
+
+		shutdown.AddHookWithPriority("check jobs", shutdown.PriorityJobs, func() {
+			canary.AcquireAllCheckLocks(ctx)
+		})
 	}
 
 	go serve()

cmd/serve.go

@@ -16,6 +16,7 @@ import (
 	"github.com/flanksource/canary-checker/pkg/db"
 	"github.com/flanksource/canary-checker/pkg/echo"
 	"github.com/flanksource/canary-checker/pkg/jobs"
+	"github.com/flanksource/canary-checker/pkg/jobs/canary"
 	canaryJobs "github.com/flanksource/canary-checker/pkg/jobs/canary"
 	echov4 "github.com/labstack/echo/v4"
 
@@ -49,6 +50,12 @@ var Serve = &cobra.Command{
 		canaryJobs.StartScanCanaryConfigs(apicontext.DefaultContext, dataFile, configFiles)
 		if executor {
 			jobs.Start()
+
+			// TODO: stop the cron scheduler so that no more checks are scheduled
+
+			shutdown.AddHookWithPriority("check jobs", shutdown.PriorityJobs, func() {
+				canary.AcquireAllCheckLocks(apicontext.DefaultContext)
+			})
 		}
 
 		serve()

pkg/jobs/canary/sync.go

@@ -19,8 +19,29 @@ import (
 	"github.com/flanksource/duty/job"
 	"github.com/flanksource/duty/models"
 	"github.com/robfig/cron/v3"
+	"golang.org/x/sync/semaphore"
 )
 
+const propertyCheckConcurrency = "check.concurrency"
+
+var (
+	// The maximum number of checks that can run concurrently
+	defaultCheckConcurrency = 50
+
+	// Holds in the lock for every running check.
+	// Can be overwritten by 'check.concurrency' property.
+	globalCheckSemaphore *semaphore.Weighted
+)
+
+// AcquireAllCheckLocks blocks until the global check sempahore is fully acquired.
+//
+// This helps to ensure that no checks are currently running.
+func AcquireAllCheckLocks(ctx context.Context) {
+	ctx.Logger.V(6).Infof("acquiring all check locks")
+	globalCheckSemaphore.Acquire(ctx, int64(ctx.Properties().Int(propertyCheckConcurrency, defaultCheckConcurrency)))
+	ctx.Logger.V(6).Infof("acquired all check locks")
+}
+
 var canaryJobs sync.Map
 
 const DefaultCanarySchedule = "@every 5m"
@@ -140,6 +161,7 @@ func newCanaryJob(c CanaryJob) {
 		IgnoreSuccessHistory: true,
 		Retention:            job.RetentionBalanced,
 		ResourceID:           c.DBCanary.ID.String(),
+		Semaphores:           []*semaphore.Weighted{globalCheckSemaphore},
 		ResourceType:         "canary",
 		ID:                   fmt.Sprintf("%s/%s", c.Canary.Namespace, c.Canary.Name),
 		Fn:                   c.Run,
@@ -159,6 +181,10 @@ var SyncCanaryJobs = &job.Job{
 	Schedule:   "@every 5m",
 	Retention:  job.RetentionFew,
 	Fn: func(ctx job.JobRuntime) error {
+		if globalCheckSemaphore == nil {
+			globalCheckSemaphore = semaphore.NewWeighted(int64(ctx.Properties().Int(propertyCheckConcurrency, defaultCheckConcurrency)))
+		}
+
 		canaries, err := db.GetAllCanariesForSync(ctx.Context, runner.WatchNamespace)
 		if err != nil {
 			return err