Golang retry.Start函数代码示例

// WaitForFullReplication waits until all stores in the cluster
// have no ranges with replication pending.
func (tc *TestCluster) WaitForFullReplication() error {
	opts := retry.Options{
		InitialBackoff: time.Millisecond * 10,
		MaxBackoff:     time.Millisecond * 100,
		Multiplier:     2,
	}

	notReplicated := true
	for r := retry.Start(opts); r.Next() && notReplicated; {
		notReplicated = false
		for _, s := range tc.Servers {
			err := s.Stores().VisitStores(func(s *storage.Store) error {
				if err := s.ComputeMetrics(); err != nil {
					return err
				}
				if s.Registry().GetGauge("ranges.replication-pending").Value() > 0 {
					notReplicated = true
				}
				return nil
			})
			if err != nil {
				return err
			}
			if notReplicated {
				break
			}
		}
	}
	return nil
}

// request returns the result of performing a http get request.
func request(url string, httpClient *http.Client) ([]byte, bool) {
	for r := retry.Start(retryOptions); r.Next(); {
		req, err := http.NewRequest("GET", url, nil)
		if err != nil {
			log.Fatal(err)
			return nil, false
		}
		req.Header.Set(util.AcceptHeader, util.JSONContentType)
		resp, err := httpClient.Do(req)
		if err != nil {
			log.Infof("could not GET %s - %s", url, err)
			continue
		}
		defer resp.Body.Close()
		body, err := ioutil.ReadAll(resp.Body)
		if err != nil {
			log.Infof("could not ready body for %s - %s", url, err)
			continue
		}
		if resp.StatusCode != http.StatusOK {
			log.Infof("could not GET %s - statuscode: %d - body: %s", url, resp.StatusCode, body)
			continue
		}
		returnedContentType := resp.Header.Get(util.ContentTypeHeader)
		if returnedContentType != util.JSONContentType {
			log.Infof("unexpected content type: %v", returnedContentType)
			continue
		}
		log.Infof("OK response from %s", url)
		return body, true
	}
	log.Warningf("There was an error retrieving %s", url)
	return nil, false
}

// GetSnapshot wraps Snapshot() but does not require the replica lock
// to be held and it will block instead of returning
// ErrSnapshotTemporaryUnavailable.
func (r *Replica) GetSnapshot() (raftpb.Snapshot, error) {
	retryOptions := retry.Options{
		InitialBackoff: 1 * time.Millisecond,
		MaxBackoff:     50 * time.Millisecond,
		Multiplier:     2,
	}
	for retry := retry.Start(retryOptions); retry.Next(); {
		r.mu.Lock()
		snap, err := r.Snapshot()
		snapshotChan := r.mu.snapshotChan
		r.mu.Unlock()
		if err == raft.ErrSnapshotTemporarilyUnavailable {
			if snapshotChan == nil {
				// The call to Snapshot() didn't start an async process due to
				// rate limiting. Try again later.
				continue
			}
			var ok bool
			snap, ok = <-snapshotChan
			if ok {
				return snap, nil
			}
			// Each snapshot worker's output can only be consumed once.
			// We could be racing with raft itself, so if we get a closed
			// channel loop back and try again.
		} else {
			return snap, err
		}
	}
	panic("unreachable") // due to infinite retries
}

func (txn *Txn) exec(retryable func(txn *Txn) error) error {
	// Run retryable in a retry loop until we encounter a success or
	// error condition this loop isn't capable of handling.
	var err error
	for r := retry.Start(txn.db.txnRetryOptions); r.Next(); {
		err = retryable(txn)
		if err == nil && txn.Proto.Status == roachpb.PENDING {
			// retryable succeeded, but didn't commit.
			err = txn.commit(nil)
		}
		if restartErr, ok := err.(roachpb.TransactionRestartError); ok {
			if log.V(2) {
				log.Warning(err)
			}
			switch restartErr.CanRestartTransaction() {
			case roachpb.TransactionRestart_IMMEDIATE:
				r.Reset()
				continue
			case roachpb.TransactionRestart_BACKOFF:
				continue
			}
			// By default, fall through and break.
		}
		break
	}
	txn.Cleanup(err)
	return err
}

// recordJoinEvent begins an asynchronous task which attempts to log a "node
// join" or "node restart" event. This query will retry until it succeeds or the
// server stops.
func (n *Node) recordJoinEvent() {
	if !n.ctx.LogRangeEvents {
		return
	}

	logEventType := sql.EventLogNodeRestart
	if n.initialBoot {
		logEventType = sql.EventLogNodeJoin
	}

	n.stopper.RunWorker(func() {
		for r := retry.Start(retry.Options{Closer: n.stopper.ShouldStop()}); r.Next(); {
			if err := n.ctx.DB.Txn(func(txn *client.Txn) error {
				return n.eventLogger.InsertEventRecord(txn,
					logEventType,
					int32(n.Descriptor.NodeID),
					int32(n.Descriptor.NodeID),
					struct {
						Descriptor roachpb.NodeDescriptor
						ClusterID  uuid.UUID
						StartedAt  int64
					}{n.Descriptor, n.ClusterID, n.startedAt},
				)
			}); err != nil {
				log.Warningc(n.context(context.TODO()), "unable to log %s event for node %d: %s", logEventType, n.Descriptor.NodeID, err)
			} else {
				return
			}
		}
	})
}

// recordJoinEvent begins an asynchronous task which attempts to log a "node
// join" or "node restart" event. This query will retry until it succeeds or the
// server stops.
func (n *Node) recordJoinEvent() {
	if !n.ctx.LogRangeEvents {
		return
	}

	logEventType := sql.EventLogNodeRestart
	if n.initialBoot {
		logEventType = sql.EventLogNodeJoin
	}

	n.stopper.RunWorker(func() {
		retryOpts := base.DefaultRetryOptions()
		retryOpts.Closer = n.stopper.ShouldStop()
		for r := retry.Start(retryOpts); r.Next(); {
			if err := n.ctx.DB.Txn(n.Ctx(), func(txn *client.Txn) error {
				return n.eventLogger.InsertEventRecord(txn,
					logEventType,
					int32(n.Descriptor.NodeID),
					int32(n.Descriptor.NodeID),
					struct {
						Descriptor roachpb.NodeDescriptor
						ClusterID  uuid.UUID
						StartedAt  int64
					}{n.Descriptor, n.ClusterID, n.startedAt},
				)
			}); err != nil {
				log.Warningf(n.Ctx(), "%s: unable to log %s event: %s", n, logEventType, err)
			} else {
				return
			}
		}
	})
}

// maybeWarnAboutInit looks for signs indicating a cluster which
// hasn't been initialized and warns. There's no absolutely sure way
// to determine whether the current node is simply waiting to be
// bootstrapped to an existing cluster vs. the operator having failed
// to initialize the cluster via the "cockroach init" command, so
// we can only warn.
//
// This method checks whether all gossip bootstrap hosts are
// connected, and whether the node itself is a bootstrap host, but
// there is still no sentinel gossip.
func (g *Gossip) maybeWarnAboutInit(stopper *stop.Stopper) {
	stopper.RunWorker(func() {
		// Wait 5s before first check.
		select {
		case <-stopper.ShouldStop():
			return
		case <-time.After(5 * time.Second):
		}
		retryOptions := retry.Options{
			InitialBackoff: 5 * time.Second,  // first backoff at 5s
			MaxBackoff:     60 * time.Second, // max backoff is 60s
			Multiplier:     2,                // doubles
			Stopper:        stopper,          // stop no matter what on stopper
		}
		// This will never error because of infinite retries.
		for r := retry.Start(retryOptions); r.Next(); {
			g.mu.Lock()
			hasSentinel := g.is.getInfo(KeySentinel) != nil
			triedAll := g.triedAll
			g.mu.Unlock()
			// If we have the sentinel, exit the retry loop.
			if hasSentinel {
				break
			}
			// Otherwise, if all bootstrap hosts are connected, warn.
			if triedAll {
				log.Warningf("connected to gossip but missing sentinel. Has the cluster been initialized? " +
					"Use \"cockroach init\" to initialize.")
			}
		}
	})
}

// waitForOneVersion returns once there are no unexpired leases on the
// previous version of the table descriptor. It returns the current version.
// After returning there can only be versions of the descriptor >= to the
// returned version. Lease acquisition (see acquire()) maintains the
// invariant that no new leases for desc.Version-1 will be granted once
// desc.Version exists.
func (s LeaseStore) waitForOneVersion(tableID sqlbase.ID, retryOpts retry.Options) (
	sqlbase.DescriptorVersion, error,
) {
	desc := &sqlbase.Descriptor{}
	descKey := sqlbase.MakeDescMetadataKey(tableID)
	var tableDesc *sqlbase.TableDescriptor
	for r := retry.Start(retryOpts); r.Next(); {
		// Get the current version of the table descriptor non-transactionally.
		//
		// TODO(pmattis): Do an inconsistent read here?
		if err := s.db.GetProto(descKey, desc); err != nil {
			return 0, err
		}
		tableDesc = desc.GetTable()
		if tableDesc == nil {
			return 0, errors.Errorf("ID %d is not a table", tableID)
		}
		// Check to see if there are any leases that still exist on the previous
		// version of the descriptor.
		now := s.clock.Now()
		count, err := s.countLeases(tableDesc.ID, tableDesc.Version-1, now.GoTime())
		if err != nil {
			return 0, err
		}
		if count == 0 {
			break
		}
		log.Infof(context.TODO(), "publish (count leases): descID=%d name=%s version=%d count=%d",
			tableDesc.ID, tableDesc.Name, tableDesc.Version-1, count)
	}
	return tableDesc.Version, nil
}

// Send sends call to Cockroach via an RPC request. Errors which are retryable
// are retried with backoff in a loop using the default retry options. Other
// errors sending the request are retried indefinitely using the same client
// command ID to avoid reporting failure when in fact the command may have gone
// through and been executed successfully. We retry here to eventually get
// through with the same client command ID and be given the cached response.
func (s *rpcSender) Send(_ context.Context, call proto.Call) {
	method := fmt.Sprintf("Server.%s", call.Args.Method())

	var err error
	for r := retry.Start(s.retryOpts); r.Next(); {
		select {
		case <-s.client.Healthy():
		default:
			err = fmt.Errorf("failed to send RPC request %s: client is unhealthy", method)
			log.Warning(err)
			continue
		}

		if err = s.client.Call(method, call.Args, call.Reply); err != nil {
			// Assume all errors sending request are retryable. The actual
			// number of things that could go wrong is vast, but we don't
			// want to miss any which should in theory be retried with the
			// same client command ID. We log the error here as a warning so
			// there's visiblity that this is happening. Some of the errors
			// we'll sweep up in this net shouldn't be retried, but we can't
			// really know for sure which.
			log.Warningf("failed to send RPC request %s: %s", method, err)
			continue
		}

		// On successful post, we're done with retry loop.
		break
	}
	if err != nil {
		call.Reply.Header().SetGoError(err)
	}
}

// Tests a batch of queries very similar to those that that PGBench runs
// in its TPC-B(ish) mode.
func runPgbenchQueryParallel(b *testing.B, db *sql.DB) {
	if err := pgbench.SetupBenchDB(db, 20000, true /*quiet*/); err != nil {
		b.Fatal(err)
	}

	retryOpts := retry.Options{
		InitialBackoff: 1 * time.Millisecond,
		MaxBackoff:     200 * time.Millisecond,
		Multiplier:     2,
	}

	b.ResetTimer()
	b.RunParallel(func(pb *testing.PB) {
		src := rand.New(rand.NewSource(5432))
		r := retry.Start(retryOpts)
		var err error
		for pb.Next() {
			r.Reset()
			for r.Next() {
				err = pgbench.RunOne(db, src, 20000)
				if err == nil {
					break
				}
			}
			if err != nil {
				b.Fatal(err)
			}
		}
	})
	b.StopTimer()
}

// get performs an HTTPS GET to the specified path for a specific node.
func get(t *testing.T, client *http.Client, node *localcluster.Container, path string) []byte {
	url := fmt.Sprintf("https://%s%s", node.Addr(""), path)
	// TODO(bram) #2059: Remove retry logic.
	for r := retry.Start(retryOptions); r.Next(); {
		resp, err := client.Get(url)
		if err != nil {
			t.Logf("could not GET %s - %s", url, err)
			continue
		}
		defer resp.Body.Close()
		body, err := ioutil.ReadAll(resp.Body)
		if err != nil {
			t.Logf("could not read body for %s - %s", url, err)
			continue
		}
		if resp.StatusCode != http.StatusOK {
			t.Logf("could not GET %s - statuscode: %d - body: %s", url, resp.StatusCode, body)
			continue
		}
		t.Logf("OK response from %s", url)
		return body
	}
	t.Fatalf("There was an error retrieving %s", url)
	return []byte("")
}

func (txn *Txn) exec(retryable func(txn *Txn) error) (err error) {
	// Run retryable in a retry loop until we encounter a success or
	// error condition this loop isn't capable of handling.
	for r := retry.Start(txn.db.txnRetryOptions); r.Next(); {
		txn.haveTxnWrite, txn.haveEndTxn = false, false // always reset before [re]starting txn
		if err = retryable(txn); err == nil {
			if !txn.haveEndTxn && txn.haveTxnWrite {
				// If there were no errors running retryable, commit the txn. This
				// may block waiting for outstanding writes to complete in case
				// retryable didn't -- we need the most recent of all response
				// timestamps in order to commit.
				err = txn.Commit()
			}
		}
		if restartErr, ok := err.(proto.TransactionRestartError); ok {
			if log.V(2) {
				log.Warning(err)
			}
			if restartErr.CanRestartTransaction() == proto.TransactionRestart_IMMEDIATE {
				r.Reset()
				continue
			} else if restartErr.CanRestartTransaction() == proto.TransactionRestart_BACKOFF {
				continue
			}
			// By default, fall through and break.
		}
		break
	}
	if err != nil && txn.haveTxnWrite {
		if replyErr := txn.Rollback(); replyErr != nil {
			log.Errorf("failure aborting transaction: %s; abort caused by: %s", replyErr, err)
		}
	}
	return
}

// GetSnapshot wraps Snapshot() but does not require the replica lock
// to be held and it will block instead of returning
// ErrSnapshotTemporaryUnavailable.
func (r *Replica) GetSnapshot(ctx context.Context) (raftpb.Snapshot, error) {
	retryOptions := retry.Options{
		InitialBackoff: 1 * time.Millisecond,
		MaxBackoff:     50 * time.Millisecond,
		Multiplier:     2,
		Closer:         r.store.Stopper().ShouldQuiesce(),
	}
	for retry := retry.Start(retryOptions); retry.Next(); {
		log.Tracef(ctx, "snapshot retry loop pass %d", retry.CurrentAttempt())
		r.mu.Lock()
		snap, err := r.SnapshotWithContext(ctx)
		snapshotChan := r.mu.snapshotChan
		r.mu.Unlock()
		if err == raft.ErrSnapshotTemporarilyUnavailable {
			if snapshotChan == nil {
				// The call to Snapshot() didn't start an async process due to
				// rate limiting. Try again later.
				continue
			}
			var ok bool
			snap, ok = <-snapshotChan
			if ok {
				return snap, nil
			}
			// Each snapshot worker's output can only be consumed once.
			// We could be racing with raft itself, so if we get a closed
			// channel loop back and try again.
		} else {
			return snap, err
		}
	}
	return raftpb.Snapshot{}, &roachpb.NodeUnavailableError{}
}

func (txn *Txn) exec(retryable func(txn *Txn) *roachpb.Error) *roachpb.Error {
	// Run retryable in a retry loop until we encounter a success or
	// error condition this loop isn't capable of handling.
	var pErr *roachpb.Error
	for r := retry.Start(txn.db.txnRetryOptions); r.Next(); {
		pErr = retryable(txn)
		if pErr == nil && txn.Proto.Status == roachpb.PENDING {
			// retryable succeeded, but didn't commit.
			pErr = txn.commit(nil)
		}

		if pErr != nil {
			switch pErr.TransactionRestart {
			case roachpb.TransactionRestart_IMMEDIATE:
				if log.V(2) {
					log.Warning(pErr)
				}
				r.Reset()
				continue
			case roachpb.TransactionRestart_BACKOFF:
				if log.V(2) {
					log.Warning(pErr)
				}
				continue
			}
			// By default, fall through and break.
		}
		break
	}
	txn.Cleanup(pErr)
	return pErr
}

// get performs an HTTPS GET to the specified path for a specific node.
func get(t *testing.T, base, rel string) []byte {
	// TODO(bram) #2059: Remove retry logic.
	url := fmt.Sprintf("%s/%s", base, rel)
	for r := retry.Start(retryOptions); r.Next(); {
		resp, err := cluster.HTTPClient.Get(url)
		if err != nil {
			log.Infof("could not GET %s - %s", url, err)
			continue
		}
		defer resp.Body.Close()
		body, err := ioutil.ReadAll(resp.Body)
		if err != nil {
			log.Infof("could not read body for %s - %s", url, err)
			continue
		}
		if resp.StatusCode != http.StatusOK {
			log.Infof("could not GET %s - statuscode: %d - body: %s", url, resp.StatusCode, body)
			continue
		}
		if log.V(1) {
			log.Infof("OK response from %s", url)
		}
		return body
	}
	t.Fatalf("There was an error retrieving %s", url)
	return []byte("")
}