multitsdb.go

// Copyright (c) The Thanos Authors.
// Licensed under the Apache License 2.0.

package receive

import (
	"context"
	"io/ioutil"
	"os"
	"path"
	"path/filepath"
	"sync"

	"github.com/go-kit/log"
	"github.com/go-kit/log/level"
	"github.com/pkg/errors"
	"github.com/prometheus/client_golang/prometheus"
	"github.com/prometheus/prometheus/model/labels"
	"github.com/prometheus/prometheus/storage"
	"github.com/prometheus/prometheus/tsdb"
	"go.uber.org/atomic"
	"golang.org/x/sync/errgroup"

	"github.com/thanos-io/thanos/pkg/block/metadata"
	"github.com/thanos-io/thanos/pkg/component"
	"github.com/thanos-io/thanos/pkg/errutil"
	"github.com/thanos-io/thanos/pkg/exemplars"
	"github.com/thanos-io/thanos/pkg/objstore"
	"github.com/thanos-io/thanos/pkg/shipper"
	"github.com/thanos-io/thanos/pkg/store"
	"github.com/thanos-io/thanos/pkg/store/labelpb"
)

type GetStatsFunc func(tenantID, statsByLabelName string) *tsdb.Stats

type MultiTSDB struct {
	dataDir         string
	logger          log.Logger
	reg             prometheus.Registerer
	tsdbOpts        *tsdb.Options
	tenantLabelName string
	labels          labels.Labels
	bucket          objstore.Bucket

	mtx                   *sync.RWMutex
	tenants               map[string]*tenant
	allowOutOfOrderUpload bool
	hashFunc              metadata.HashFunc
}

// NewMultiTSDB creates new MultiTSDB.
// NOTE: Passed labels has to be sorted by name.
func NewMultiTSDB(
	dataDir string,
	l log.Logger,
	reg prometheus.Registerer,
	tsdbOpts *tsdb.Options,
	labels labels.Labels,
	tenantLabelName string,
	bucket objstore.Bucket,
	allowOutOfOrderUpload bool,
	hashFunc metadata.HashFunc,
) *MultiTSDB {
	if l == nil {
		l = log.NewNopLogger()
	}

	return &MultiTSDB{
		dataDir:               dataDir,
		logger:                log.With(l, "component", "multi-tsdb"),
		reg:                   reg,
		tsdbOpts:              tsdbOpts,
		mtx:                   &sync.RWMutex{},
		tenants:               map[string]*tenant{},
		labels:                labels,
		tenantLabelName:       tenantLabelName,
		bucket:                bucket,
		allowOutOfOrderUpload: allowOutOfOrderUpload,
		hashFunc:              hashFunc,
	}
}

type tenant struct {
	readyS        *ReadyStorage
	storeTSDB     *store.TSDBStore
	exemplarsTSDB *exemplars.TSDB
	ship          *shipper.Shipper

	mtx *sync.RWMutex
}

func newTenant() *tenant {
	return &tenant{
		readyS: &ReadyStorage{},
		mtx:    &sync.RWMutex{},
	}
}

func (t *tenant) readyStorage() *ReadyStorage {
	return t.readyS
}

func (t *tenant) store() *store.TSDBStore {
	t.mtx.RLock()
	defer t.mtx.RUnlock()
	return t.storeTSDB
}

func (t *tenant) exemplars() *exemplars.TSDB {
	t.mtx.RLock()
	defer t.mtx.RUnlock()
	return t.exemplarsTSDB
}

func (t *tenant) shipper() *shipper.Shipper {
	t.mtx.RLock()
	defer t.mtx.RUnlock()
	return t.ship
}

func (t *tenant) set(storeTSDB *store.TSDBStore, tenantTSDB *tsdb.DB, ship *shipper.Shipper, exemplarsTSDB *exemplars.TSDB) {
	t.readyS.Set(tenantTSDB)
	t.mtx.Lock()
	t.storeTSDB = storeTSDB
	t.ship = ship
	t.exemplarsTSDB = exemplarsTSDB
	t.mtx.Unlock()
}

func (t *MultiTSDB) Open() error {
	if err := os.MkdirAll(t.dataDir, 0750); err != nil {
		return err
	}

	files, err := ioutil.ReadDir(t.dataDir)
	if err != nil {
		return err
	}

	var g errgroup.Group
	for _, f := range files {
		f := f
		if !f.IsDir() {
			continue
		}

		g.Go(func() error {
			_, err := t.getOrLoadTenant(f.Name(), true)
			return err
		})
	}

	return g.Wait()
}

func (t *MultiTSDB) Flush() error {
	t.mtx.RLock()
	defer t.mtx.RUnlock()

	errmtx := &sync.Mutex{}
	merr := errutil.MultiError{}
	wg := &sync.WaitGroup{}
	for id, tenant := range t.tenants {
		db := tenant.readyStorage().Get()
		if db == nil {
			level.Error(t.logger).Log("msg", "flushing TSDB failed; not ready", "tenant", id)
			continue
		}
		level.Info(t.logger).Log("msg", "flushing TSDB", "tenant", id)
		wg.Add(1)
		go func() {
			head := db.Head()
			if err := db.CompactHead(tsdb.NewRangeHead(head, head.MinTime(), head.MaxTime()-1)); err != nil {
				errmtx.Lock()
				merr.Add(err)
				errmtx.Unlock()
			}
			wg.Done()
		}()
	}

	wg.Wait()
	return merr.Err()
}

func (t *MultiTSDB) Close() error {
	t.mtx.Lock()
	defer t.mtx.Unlock()

	merr := errutil.MultiError{}
	for id, tenant := range t.tenants {
		db := tenant.readyStorage().Get()
		if db == nil {
			level.Error(t.logger).Log("msg", "closing TSDB failed; not ready", "tenant", id)
			continue
		}
		level.Info(t.logger).Log("msg", "closing TSDB", "tenant", id)
		merr.Add(db.Close())
	}
	return merr.Err()
}

func (t *MultiTSDB) Sync(ctx context.Context) (int, error) {
	if t.bucket == nil {
		return 0, errors.New("bucket is not specified, Sync should not be invoked")
	}

	t.mtx.RLock()
	defer t.mtx.RUnlock()

	var (
		errmtx   = &sync.Mutex{}
		merr     = errutil.MultiError{}
		wg       = &sync.WaitGroup{}
		uploaded atomic.Int64
	)

	for tenantID, tenant := range t.tenants {
		level.Debug(t.logger).Log("msg", "uploading block for tenant", "tenant", tenantID)
		s := tenant.shipper()
		if s == nil {
			continue
		}
		wg.Add(1)
		go func() {
			up, err := s.Sync(ctx)
			if err != nil {
				errmtx.Lock()
				merr.Add(errors.Wrap(err, "upload"))
				errmtx.Unlock()
			}
			uploaded.Add(int64(up))
			wg.Done()
		}()
	}
	wg.Wait()
	return int(uploaded.Load()), merr.Err()
}

func (t *MultiTSDB) RemoveLockFilesIfAny() error {
	fis, err := ioutil.ReadDir(t.dataDir)
	if err != nil {
		if os.IsNotExist(err) {
			return nil
		}
		return err
	}

	merr := errutil.MultiError{}
	for _, fi := range fis {
		if !fi.IsDir() {
			continue
		}
		if err := os.Remove(filepath.Join(t.defaultTenantDataDir(fi.Name()), "lock")); err != nil {
			if os.IsNotExist(err) {
				continue
			}
			merr.Add(err)
			continue
		}
		level.Info(t.logger).Log("msg", "a leftover lockfile found and removed", "tenant", fi.Name())
	}
	return merr.Err()
}

func (t *MultiTSDB) TSDBStores() map[string]store.InfoStoreServer {
	t.mtx.RLock()
	defer t.mtx.RUnlock()

	res := make(map[string]store.InfoStoreServer, len(t.tenants))
	for k, tenant := range t.tenants {
		s := tenant.store()
		if s != nil {
			res[k] = s
		}
	}
	return res
}

func (t *MultiTSDB) TSDBExemplars() map[string]*exemplars.TSDB {
	t.mtx.RLock()
	defer t.mtx.RUnlock()

	res := make(map[string]*exemplars.TSDB, len(t.tenants))
	for k, tenant := range t.tenants {
		e := tenant.exemplars()
		if e != nil {
			res[k] = e
		}
	}
	return res
}

func (t *MultiTSDB) Stats(tenantID, statsByLabelName string) *tsdb.Stats {
	t.mtx.RLock()
	defer t.mtx.RUnlock()

	tenant, ok := t.tenants[tenantID]
	if !ok {
		return nil
	}

	return tenant.readyS.get().db.Head().Stats(statsByLabelName)
}

func (t *MultiTSDB) startTSDB(logger log.Logger, tenantID string, tenant *tenant) error {
	reg := prometheus.WrapRegistererWith(prometheus.Labels{"tenant": tenantID}, t.reg)
	lset := labelpb.ExtendSortedLabels(t.labels, labels.FromStrings(t.tenantLabelName, tenantID))
	dataDir := t.defaultTenantDataDir(tenantID)

	level.Info(logger).Log("msg", "opening TSDB")
	opts := *t.tsdbOpts
	s, err := tsdb.Open(
		dataDir,
		logger,
		&UnRegisterer{Registerer: reg},
		&opts,
		nil,
	)
	if err != nil {
		t.mtx.Lock()
		delete(t.tenants, tenantID)
		t.mtx.Unlock()
		return err
	}
	var ship *shipper.Shipper
	if t.bucket != nil {
		ship = shipper.New(
			logger,
			reg,
			dataDir,
			t.bucket,
			func() labels.Labels { return lset },
			metadata.ReceiveSource,
			false,
			t.allowOutOfOrderUpload,
			t.hashFunc,
		)
	}
	tenant.set(store.NewTSDBStore(logger, s, component.Receive, lset), s, ship, exemplars.NewTSDB(s, lset))
	level.Info(logger).Log("msg", "TSDB is now ready")
	return nil
}

func (t *MultiTSDB) defaultTenantDataDir(tenantID string) string {
	return path.Join(t.dataDir, tenantID)
}

func (t *MultiTSDB) getOrLoadTenant(tenantID string, blockingStart bool) (*tenant, error) {
	// Fast path, as creating tenants is a very rare operation.
	t.mtx.RLock()
	tenant, exist := t.tenants[tenantID]
	t.mtx.RUnlock()
	if exist {
		return tenant, nil
	}

	// Slow path needs to lock fully and attempt to read again to prevent race
	// conditions, where since the fast path was tried, there may have actually
	// been the same tenant inserted in the map.
	t.mtx.Lock()
	tenant, exist = t.tenants[tenantID]
	if exist {
		t.mtx.Unlock()
		return tenant, nil
	}

	tenant = newTenant()
	t.tenants[tenantID] = tenant
	t.mtx.Unlock()

	logger := log.With(t.logger, "tenant", tenantID)
	if !blockingStart {
		go func() {
			if err := t.startTSDB(logger, tenantID, tenant); err != nil {
				level.Error(logger).Log("msg", "failed to start tsdb asynchronously", "err", err)
			}
		}()
		return tenant, nil
	}
	return tenant, t.startTSDB(logger, tenantID, tenant)
}

func (t *MultiTSDB) TenantAppendable(tenantID string) (Appendable, error) {
	tenant, err := t.getOrLoadTenant(tenantID, false)
	if err != nil {
		return nil, err
	}
	return tenant.readyStorage(), nil
}

// ErrNotReady is returned if the underlying storage is not ready yet.
var ErrNotReady = errors.New("TSDB not ready")

// ReadyStorage implements the Storage interface while allowing to set the actual
// storage at a later point in time.
// TODO: Replace this with upstream Prometheus implementation when it is exposed.
type ReadyStorage struct {
	mtx sync.RWMutex
	a   *adapter
}

// Set the storage.
func (s *ReadyStorage) Set(db *tsdb.DB) {
	s.mtx.Lock()
	defer s.mtx.Unlock()

	s.a = &adapter{db: db}
}

// Get the storage.
func (s *ReadyStorage) Get() *tsdb.DB {
	if x := s.get(); x != nil {
		return x.db
	}
	return nil
}

func (s *ReadyStorage) get() *adapter {
	s.mtx.RLock()
	x := s.a
	s.mtx.RUnlock()
	return x
}

// StartTime implements the Storage interface.
func (s *ReadyStorage) StartTime() (int64, error) {
	return 0, errors.New("not implemented")
}

// Querier implements the Storage interface.
func (s *ReadyStorage) Querier(ctx context.Context, mint, maxt int64) (storage.Querier, error) {
	if x := s.get(); x != nil {
		return x.Querier(ctx, mint, maxt)
	}
	return nil, ErrNotReady
}

// ExemplarQuerier implements the Storage interface.
func (s *ReadyStorage) ExemplarQuerier(ctx context.Context) (storage.ExemplarQuerier, error) {
	if x := s.get(); x != nil {
		return x.ExemplarQuerier(ctx)
	}
	return nil, ErrNotReady
}

// Appender implements the Storage interface.
func (s *ReadyStorage) Appender(ctx context.Context) (storage.Appender, error) {
	if x := s.get(); x != nil {
		return x.Appender(ctx)
	}
	return nil, ErrNotReady
}

// Close implements the Storage interface.
func (s *ReadyStorage) Close() error {
	if x := s.Get(); x != nil {
		return x.Close()
	}
	return nil
}

// adapter implements a storage.Storage around TSDB.
type adapter struct {
	db *tsdb.DB
}

// StartTime implements the Storage interface.
func (a adapter) StartTime() (int64, error) {
	return 0, errors.New("not implemented")
}

func (a adapter) Querier(ctx context.Context, mint, maxt int64) (storage.Querier, error) {
	return a.db.Querier(ctx, mint, maxt)
}

func (a adapter) ExemplarQuerier(ctx context.Context) (storage.ExemplarQuerier, error) {
	return a.db.ExemplarQuerier(ctx)
}

// Appender returns a new appender against the storage.
func (a adapter) Appender(ctx context.Context) (storage.Appender, error) {
	return a.db.Appender(ctx), nil
}

// Close closes the storage and all its underlying resources.
func (a adapter) Close() error {
	return a.db.Close()
}

// UnRegisterer is a Prometheus registerer that
// ensures that collectors can be registered
// by unregistering already-registered collectors.
// FlushableStorage uses this registerer in order
// to not lose metric values between DB flushes.
type UnRegisterer struct {
	prometheus.Registerer
}

func (u *UnRegisterer) MustRegister(cs ...prometheus.Collector) {
	for _, c := range cs {
		if err := u.Register(c); err != nil {
			if _, ok := err.(prometheus.AlreadyRegisteredError); ok {
				if ok = u.Unregister(c); !ok {
					panic("unable to unregister existing collector")
				}
				u.Registerer.MustRegister(c)
				continue
			}
			panic(err)
		}
	}
}