Merge pull request #13 from SberMarket-Tech/retry-on-error

Add retries for committing completed jobs

dead_pool_reaper.go

@@ -15,7 +15,7 @@ import (
 
 const (
 	deadTime          = 10 * time.Second // 2 x heartbeat
-	reapPeriod        = 10 * time.Minute
+	defaultReapPeriod = 5 * time.Minute
 	reapJitterSecs    = 30
 	requeueKeysPerJob = 4
 )
@@ -31,7 +31,11 @@ type deadPoolReaper struct {
 	doneStoppingChan chan struct{}
 }
 
-func newDeadPoolReaper(namespace string, pool Pool, curJobTypes []string) *deadPoolReaper {
+func newDeadPoolReaper(namespace string, pool Pool, curJobTypes []string, reapPeriod time.Duration) *deadPoolReaper {
+	if reapPeriod == 0 {
+		reapPeriod = defaultReapPeriod
+	}
+
 	return &deadPoolReaper{
 		namespace:        namespace,
 		pool:             pool,
@@ -53,6 +57,8 @@ func (r *deadPoolReaper) stop() {
 }
 
 func (r *deadPoolReaper) loop() {
+	Logger.Printf("Reaper: started with a period of %v", r.reapPeriod)
+
 	// Reap immediately after we provide some time for initialization
 	timer := time.NewTimer(r.deadTime)
 	defer timer.Stop()
@@ -66,7 +72,6 @@ func (r *deadPoolReaper) loop() {
 			// Schedule next occurrence periodically with jitter
 			timer.Reset(r.reapPeriod + time.Duration(rand.Intn(reapJitterSecs))*time.Second)
 
-			// Reap
 			if err := r.reap(); err != nil {
 				logError("dead_pool_reaper.reap", err)
 			}
@@ -80,24 +85,35 @@ func (r *deadPoolReaper) reap() (err error) {
 		return err
 	}
 
+	Logger.Printf("Reaper: trying to acquire lock...")
+
 	acquired, err := r.acquireLock(lockValue)
 	if err != nil {
+		Logger.Printf("Reaper: acquiring lock: %v", err)
 		return err
 	}
 
 	// Another reaper is already running
 	if !acquired {
+		Logger.Printf("Reaper: locked by another process")
 		return nil
 	}
 
+	Logger.Printf("Reaper: lock is acquired")
+
 	defer func() {
 		err = r.releaseLock(lockValue)
 	}()
 
 	rErr := r.reapDeadPools()
 	cErr := r.clearUnknownPools()
 
-	return multierr.Combine(rErr, cErr)
+	// TODO: consider refactoring requeueInProgressJobs and cleanStaleLockInfo
+	// and removing removeDanglingLocks. There was a block where lock is 1 and
+	// lock_info is 0.
+	dErr := r.removeDanglingLocks()
+
+	return multierr.Combine(err, rErr, cErr, dErr)
 }
 
 // reapDeadPools collects the IDs of expired heartbeat pools and releases the
@@ -108,6 +124,8 @@ func (r *deadPoolReaper) reapDeadPools() error {
 		return err
 	}
 
+	Logger.Printf("Reaper: dead pools: %v", deadPoolIDs)
+
 	conn := r.pool.Get()
 	defer conn.Close()
 
@@ -150,6 +168,8 @@ func (r *deadPoolReaper) clearUnknownPools() error {
 		return err
 	}
 
+	Logger.Printf("Reaper: unknown pools: %v", unknownPools)
+
 	for poolID, jobTypes := range unknownPools {
 		if err = r.requeueInProgressJobs(poolID, jobTypes); err != nil {
 			return err
@@ -293,6 +313,31 @@ func (r *deadPoolReaper) getUnknownPools() (map[string][]string, error) {
 	return pools, nil
 }
 
+// removeDanglingLocks adjusts the lock keys according to the lock_info numbers.
+// TODO: it's better to find where the inconsistency comes from.
+func (r *deadPoolReaper) removeDanglingLocks() error {
+	keysCount := len(r.curJobTypes) * 2               // lock and lock_info keys
+	scriptArgs := make([]interface{}, 0, keysCount+1) // +1 for keys count arg
+	scriptArgs = append(scriptArgs, keysCount)
+
+	for _, j := range r.curJobTypes {
+		scriptArgs = append(scriptArgs, redisKeyJobsLock(r.namespace, j))
+		scriptArgs = append(scriptArgs, redisKeyJobsLockInfo(r.namespace, j))
+	}
+
+	conn := r.pool.Get()
+	defer conn.Close()
+
+	keys, err := redis.Strings(redisRemoveDanglingLocksScript.Do(conn, scriptArgs...))
+	if err != nil {
+		return err
+	}
+
+	Logger.Printf("Reaper: dangling locks: %v", keys)
+
+	return nil
+}
+
 // acquireLock acquires lock with a value and an expiration time for reap period.
 func (r *deadPoolReaper) acquireLock(value string) (bool, error) {
 	conn := r.pool.Get()

dead_pool_reaper_test.go

@@ -48,7 +48,7 @@ func TestDeadPoolReaper(t *testing.T) {
 	assert.NoError(t, err)
 
 	// Test getting dead pool
-	reaper := newDeadPoolReaper(ns, pool, []string{})
+	reaper := newDeadPoolReaper(ns, pool, []string{}, 0)
 	deadPools, err := reaper.findDeadPools()
 	assert.NoError(t, err)
 	assert.Equal(t, map[string][]string{"2": {"type1", "type2"}, "3": {"type1", "type2"}}, deadPools)
@@ -127,7 +127,7 @@ func TestDeadPoolReaperNoHeartbeat(t *testing.T) {
 	assert.EqualValues(t, 3, numPools)
 
 	// Test getting dead pool ids
-	reaper := newDeadPoolReaper(ns, pool, []string{"type1"})
+	reaper := newDeadPoolReaper(ns, pool, []string{"type1"}, 0)
 	deadPools, err := reaper.findDeadPools()
 	assert.NoError(t, err)
 	assert.Equal(t, map[string][]string{"1": nil, "2": nil, "3": nil}, deadPools)
@@ -210,7 +210,7 @@ func TestDeadPoolReaperNoJobTypes(t *testing.T) {
 	assert.NoError(t, err)
 
 	// Test getting dead pool
-	reaper := newDeadPoolReaper(ns, pool, []string{})
+	reaper := newDeadPoolReaper(ns, pool, []string{}, 0)
 	deadPools, err := reaper.findDeadPools()
 	assert.NoError(t, err)
 	assert.Equal(t, map[string][]string{"2": {"type1", "type2"}}, deadPools)
@@ -283,7 +283,7 @@ func TestDeadPoolReaperWithWorkerPools(t *testing.T) {
 
 	// setup a worker pool and start the reaper, which should restart the stale job above
 	wp := setupTestWorkerPool(pool, ns, job1, 1, JobOptions{Priority: 1})
-	wp.deadPoolReaper = newDeadPoolReaper(wp.namespace, wp.pool, []string{"job1"})
+	wp.deadPoolReaper = newDeadPoolReaper(wp.namespace, wp.pool, []string{"job1"}, 0)
 	wp.deadPoolReaper.deadTime = expectedDeadTime
 	wp.deadPoolReaper.start()
 
@@ -327,7 +327,7 @@ func TestDeadPoolReaperCleanStaleLocks(t *testing.T) {
 	err = conn.Flush()
 	assert.NoError(t, err)
 
-	reaper := newDeadPoolReaper(ns, pool, jobNames)
+	reaper := newDeadPoolReaper(ns, pool, jobNames, 0)
 	// clean lock info for workerPoolID1
 	err = reaper.cleanStaleLockInfo(workerPoolID1, jobNames)
 	assert.NoError(t, err)
@@ -384,7 +384,7 @@ func TestDeadPoolReaperTakeDeadPools(t *testing.T) {
 	assert.NoError(t, err)
 
 	// Test getting dead pools
-	reaper := newDeadPoolReaper(ns, pool, []string{})
+	reaper := newDeadPoolReaper(ns, pool, []string{}, 0)
 	deadPools, err := reaper.findDeadPools()
 	assert.NoError(t, err)
 	assert.Equal(t, map[string][]string{"2": {"type1", "type2"}, "3": nil}, deadPools)
@@ -395,7 +395,7 @@ func TestReaperLock(t *testing.T) {
 	ns := "work"
 	cleanKeyspace(ns, pool)
 
-	reaper := newDeadPoolReaper(ns, pool, []string{})
+	reaper := newDeadPoolReaper(ns, pool, []string{}, 0)
 
 	value, err := genValue()
 	assert.NoError(t, err)
@@ -472,7 +472,7 @@ func TestDeadPoolReaperGetUnknownPools(t *testing.T) {
 	assert.NoError(t, conn.Flush())
 
 	// Run test
-	reaper := newDeadPoolReaper(ns, pool, jobNames)
+	reaper := newDeadPoolReaper(ns, pool, jobNames, 0)
 	unknownPools, err := reaper.getUnknownPools()
 	assert.NoError(t, err)
 	assert.Equal(t, map[string][]string{"2": {"type1", "type2"}, "3": {"type1", "type2"}}, unknownPools)
@@ -531,7 +531,7 @@ func TestDeadPoolReaperClearUnknownPool(t *testing.T) {
 	assert.NoError(t, conn.Flush())
 
 	// Run test
-	reaper := newDeadPoolReaper(ns, pool, jobNames)
+	reaper := newDeadPoolReaper(ns, pool, jobNames, 0)
 	err = reaper.clearUnknownPools()
 	assert.NoError(t, err)
 
@@ -551,3 +551,62 @@ func TestDeadPoolReaperClearUnknownPool(t *testing.T) {
 	assert.NoError(t, err)
 	assert.Equal(t, map[string]string{workerPoolID1: "0"}, nLockInfo2)
 }
+
+func TestDeadPoolReaperRemoveDanglingLocks(t *testing.T) {
+	pool := newTestPool(":6379")
+	ns := "work"
+	cleanKeyspace(ns, pool)
+
+	workerPoolID1, workerPoolID2 := "1", "2"
+
+	job1, job2, job3, job4 := "type1", "type2", "type3", "type4"
+	jobNames := []string{job1, job2, job3, job4}
+	lock1, lock2, lock3 := redisKeyJobsLock(ns, job1), redisKeyJobsLock(ns, job2), redisKeyJobsLock(ns, job3)
+	lockInfo1, lockInfo2 := redisKeyJobsLockInfo(ns, job1), redisKeyJobsLockInfo(ns, job2)
+
+	conn := pool.Get()
+	defer conn.Close()
+
+	// Create redis data
+	var err error
+
+	err = conn.Send("SET", lock1, 4) // One dangling lock
+	assert.NoError(t, err)
+
+	err = conn.Send("HMSET", lockInfo1,
+		workerPoolID1, 2,
+		workerPoolID2, 1,
+	)
+	assert.NoError(t, err)
+
+	err = conn.Send("SET", lock2, 1) // No dangling locks
+	assert.NoError(t, err)
+
+	err = conn.Send("HMSET", lockInfo2,
+		workerPoolID1, 0,
+		workerPoolID2, 1,
+	)
+	assert.NoError(t, err)
+
+	err = conn.Send("SET", lock3, 1) // One dangling lock
+	assert.NoError(t, err)
+
+	assert.NoError(t, conn.Flush())
+
+	reaper := newDeadPoolReaper(ns, pool, jobNames, 0)
+	err = reaper.removeDanglingLocks()
+	assert.NoError(t, err)
+
+	// Checks
+	nLock1, err := redis.Int(conn.Do("GET", lock1))
+	assert.NoError(t, err)
+	assert.Equal(t, 3, nLock1)
+
+	nLock2, err := redis.Int(conn.Do("GET", lock2))
+	assert.NoError(t, err)
+	assert.Equal(t, 1, nLock2)
+
+	nLock3, err := redis.Int(conn.Do("GET", lock3))
+	assert.NoError(t, err)
+	assert.Equal(t, 0, nLock3)
+}

redis.go

@@ -440,3 +440,37 @@ end
 
 return cjson.encode(unknownPools)
 `)
+
+// Used by the reaper to DECR dangling locks. Returns the dangling lock keys that
+// have been fixed.
+//
+// KEYS[1] = job's lock key
+// KEYS[2...] = job's lock info key
+// Returns: ["ns:jobs:job1:lock", "ns:jobs:job3:lock"]
+var redisRemoveDanglingLocksScript = redis.NewScript(-1, `
+local danglingLocks = {}
+
+for i=1,#KEYS,2 do
+    local lockKey = KEYS[i]
+    local lockInfoKey = KEYS[i+1]
+
+    local rlocks = redis.call('get', lockKey)
+    if rlocks ~= false then
+        local locks = tonumber(rlocks)
+        local lockInfo = redis.call('hvals', lockInfoKey)
+
+        local totalLocks = 0
+        for j=1,#lockInfo do
+            totalLocks = totalLocks + tonumber(lockInfo[j])
+        end
+
+        local diff = locks - totalLocks
+        if diff ~= 0 then
+            table.insert(danglingLocks, lockKey)
+            redis.call('decrby', lockKey, diff)
+        end
+    end
+end
+
+return danglingLocks
+`)