X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/3f5b1be0f705c8f1491fdca5e3b57982d97378fd..7345838cb097f11e2ba8239020762ae867591510:/lib/dispatchcloud/scheduler/scheduler.go

diff --git a/lib/dispatchcloud/scheduler/scheduler.go b/lib/dispatchcloud/scheduler/scheduler.go
index c3e67dd11f..bc6574a21a 100644
--- a/lib/dispatchcloud/scheduler/scheduler.go
+++ b/lib/dispatchcloud/scheduler/scheduler.go
@@ -9,6 +9,7 @@ package scheduler
 import (
 	"context"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"git.arvados.org/arvados.git/sdk/go/arvados"
@@ -31,6 +32,7 @@ import (
 // shuts down idle workers, in case they are consuming quota.
 type Scheduler struct {
 	logger              logrus.FieldLogger
+	client              *arvados.Client
 	queue               ContainerQueue
 	pool                WorkerPool
 	reg                 *prometheus.Registry
@@ -45,18 +47,29 @@ type Scheduler struct {
 	stop    chan struct{}
 	stopped chan struct{}
 
+	last503time          time.Time // last time API responded 503
+	maxConcurrency       int       // dynamic container limit (0 = unlimited), see runQueue()
+	supervisorFraction   float64   // maximum fraction of "supervisor" containers (these are containers who's main job is to launch other containers, e.g. workflow runners)
+	maxInstances         int       // maximum number of instances the pool will bring up (0 = unlimited)
+	instancesWithinQuota int       // max concurrency achieved since last quota error (0 = no quota error yet)
+
 	mContainersAllocatedNotStarted   prometheus.Gauge
 	mContainersNotAllocatedOverQuota prometheus.Gauge
 	mLongestWaitTimeSinceQueue       prometheus.Gauge
+	mLast503Time                     prometheus.Gauge
+	mMaxContainerConcurrency         prometheus.Gauge
+
+	lastQueue atomic.Value // stores a []QueueEnt
 }
 
 // New returns a new unstarted Scheduler.
 //
 // Any given queue and pool should not be used by more than one
 // scheduler at a time.
-func New(ctx context.Context, queue ContainerQueue, pool WorkerPool, reg *prometheus.Registry, staleLockTimeout, queueUpdateInterval time.Duration) *Scheduler {
+func New(ctx context.Context, client *arvados.Client, queue ContainerQueue, pool WorkerPool, reg *prometheus.Registry, staleLockTimeout, queueUpdateInterval time.Duration, minQuota, maxInstances int, supervisorFraction float64) *Scheduler {
 	sch := &Scheduler{
 		logger:              ctxlog.FromContext(ctx),
+		client:              client,
 		queue:               queue,
 		pool:                pool,
 		reg:                 reg,
@@ -66,6 +79,13 @@ func New(ctx context.Context, queue ContainerQueue, pool WorkerPool, reg *promet
 		stop:                make(chan struct{}),
 		stopped:             make(chan struct{}),
 		uuidOp:              map[string]string{},
+		supervisorFraction:  supervisorFraction,
+		maxInstances:        maxInstances,
+	}
+	if minQuota > 0 {
+		sch.maxConcurrency = minQuota
+	} else {
+		sch.maxConcurrency = maxInstances
 	}
 	sch.registerMetrics(reg)
 	return sch
@@ -96,6 +116,32 @@ func (sch *Scheduler) registerMetrics(reg *prometheus.Registry) {
 		Help:      "Current longest wait time of any container since queuing, and before the start of crunch-run.",
 	})
 	reg.MustRegister(sch.mLongestWaitTimeSinceQueue)
+	sch.mLast503Time = prometheus.NewGauge(prometheus.GaugeOpts{
+		Namespace: "arvados",
+		Subsystem: "dispatchcloud",
+		Name:      "last_503_time",
+		Help:      "Time of most recent 503 error received from API.",
+	})
+	reg.MustRegister(sch.mLast503Time)
+	sch.mMaxContainerConcurrency = prometheus.NewGauge(prometheus.GaugeOpts{
+		Namespace: "arvados",
+		Subsystem: "dispatchcloud",
+		Name:      "max_concurrent_containers",
+		Help:      "Dynamically assigned limit on number of containers scheduled concurrency, set after receiving 503 errors from API.",
+	})
+	reg.MustRegister(sch.mMaxContainerConcurrency)
+	reg.MustRegister(prometheus.NewGaugeFunc(prometheus.GaugeOpts{
+		Namespace: "arvados",
+		Subsystem: "dispatchcloud",
+		Name:      "at_quota",
+		Help:      "Flag indicating the cloud driver is reporting an at-quota condition.",
+	}, func() float64 {
+		if sch.pool.AtQuota() {
+			return 1
+		} else {
+			return 0
+		}
+	}))
 }
 
 func (sch *Scheduler) updateMetrics() {
@@ -149,14 +195,23 @@ func (sch *Scheduler) run() {
 	}
 
 	// Keep the queue up to date.
-	poll := time.NewTicker(sch.queueUpdateInterval)
-	defer poll.Stop()
 	go func() {
-		for range poll.C {
+		for {
+			starttime := time.Now()
 			err := sch.queue.Update()
 			if err != nil {
 				sch.logger.Errorf("error updating queue: %s", err)
 			}
+			// If the previous update took a long time,
+			// that probably means the server is
+			// overloaded, so wait that long before doing
+			// another. Otherwise, wait for the configured
+			// poll interval.
+			delay := time.Since(starttime)
+			if delay < sch.queueUpdateInterval {
+				delay = sch.queueUpdateInterval
+			}
+			time.Sleep(delay)
 		}
 	}()