prometheus/db.go

499 lines
10 KiB
Go
Raw Normal View History

2016-11-15 09:34:25 +00:00
// Package tsdb implements a time series storage for float64 sample data.
package tsdb
import (
"bytes"
2016-12-04 12:16:11 +00:00
"fmt"
"os"
"path/filepath"
"reflect"
2016-12-15 07:31:26 +00:00
"strconv"
"sync"
2016-11-15 09:34:25 +00:00
"time"
"unsafe"
2016-11-15 09:34:25 +00:00
2016-12-15 07:31:26 +00:00
"golang.org/x/sync/errgroup"
"github.com/fabxc/tsdb/chunks"
2016-12-21 08:39:01 +00:00
"github.com/fabxc/tsdb/labels"
"github.com/go-kit/kit/log"
2016-12-31 08:48:49 +00:00
"github.com/prometheus/client_golang/prometheus"
2016-11-15 09:34:25 +00:00
)
2016-12-09 09:00:14 +00:00
// DefaultOptions used for the DB. They are sane for setups using
// millisecond precision timestamps.
2016-11-15 09:34:25 +00:00
var DefaultOptions = &Options{
2016-12-26 15:55:44 +00:00
Retention: 15 * 24 * 3600 * 1000, // 15 days
DisableWAL: false,
2016-11-15 09:34:25 +00:00
}
// Options of the DB storage.
type Options struct {
2016-12-26 15:55:44 +00:00
Retention int64
DisableWAL bool
2016-11-15 09:34:25 +00:00
}
// DB is a time series storage.
type DB struct {
logger log.Logger
opts *Options
path string
2016-11-15 09:34:25 +00:00
2016-12-15 07:36:09 +00:00
shards []*Shard
2016-11-15 09:34:25 +00:00
}
2016-12-04 12:16:11 +00:00
// TODO(fabxc): make configurable
const (
shardShift = 0
2016-12-15 07:36:09 +00:00
numShards = 1 << shardShift
maxChunkSize = 1024
2016-12-04 12:16:11 +00:00
)
2016-11-15 09:34:25 +00:00
// Open or create a new DB.
func Open(path string, l log.Logger, opts *Options) (*DB, error) {
if opts == nil {
opts = DefaultOptions
}
2016-12-04 12:16:11 +00:00
if err := os.MkdirAll(path, 0777); err != nil {
return nil, err
}
if l == nil {
l = log.NewLogfmtLogger(os.Stdout)
l = log.NewContext(l).With("ts", log.DefaultTimestampUTC, "caller", log.DefaultCaller)
}
2016-11-15 09:34:25 +00:00
c := &DB{
2016-12-04 12:16:11 +00:00
logger: l,
opts: opts,
path: path,
2016-11-15 09:34:25 +00:00
}
2016-12-04 12:16:11 +00:00
// Initialize vertical shards.
// TODO(fabxc): validate shard number to be power of 2, which is required
// for the bitshift-modulo when finding the right shard.
2016-12-15 07:36:09 +00:00
for i := 0; i < numShards; i++ {
l := log.NewContext(l).With("shard", i)
2016-12-15 07:31:26 +00:00
d := shardDir(path, i)
2016-12-31 08:48:49 +00:00
s, err := OpenShard(d, i, l)
2016-12-15 07:31:26 +00:00
if err != nil {
return nil, fmt.Errorf("initializing shard %q failed: %s", d, err)
}
c.shards = append(c.shards, s)
2016-12-04 12:16:11 +00:00
}
2016-11-15 09:34:25 +00:00
return c, nil
}
2016-12-15 07:31:26 +00:00
func shardDir(base string, i int) string {
return filepath.Join(base, strconv.Itoa(i))
}
2016-12-04 12:16:11 +00:00
// Close the database.
func (db *DB) Close() error {
2016-12-15 07:31:26 +00:00
var g errgroup.Group
2016-12-09 09:00:14 +00:00
2016-12-15 07:31:26 +00:00
for _, shard := range db.shards {
g.Go(shard.Close)
}
2016-12-09 09:00:14 +00:00
2016-12-15 07:31:26 +00:00
return g.Wait()
2016-12-04 12:16:11 +00:00
}
// Appender allows committing batches of samples to a database.
// The data held by the appender is reset after Commit returns.
type Appender interface {
// AddSeries registers a new known series label set with the appender
// and returns a reference number used to add samples to it over the
// life time of the Appender.
// AddSeries(Labels) uint64
// Add adds a sample pair for the referenced series.
2016-12-29 10:03:39 +00:00
Add(lset labels.Labels, t int64, v float64) error
// Commit submits the collected samples and purges the batch.
Commit() error
}
// Appender returns a new appender against the database.
func (db *DB) Appender() Appender {
return &bucketAppender{
db: db,
buckets: make([][]hashedSample, numShards),
}
}
type bucketAppender struct {
db *DB
buckets [][]hashedSample
}
2016-12-29 10:03:39 +00:00
func (ba *bucketAppender) Add(lset labels.Labels, t int64, v float64) error {
h := lset.Hash()
s := h >> (64 - shardShift)
ba.buckets[s] = append(ba.buckets[s], hashedSample{
hash: h,
labels: lset,
t: t,
v: v,
})
2016-12-29 10:03:39 +00:00
return nil
}
func (ba *bucketAppender) reset() {
for i := range ba.buckets {
ba.buckets[i] = ba.buckets[i][:0]
2016-12-09 12:41:38 +00:00
}
}
func (ba *bucketAppender) Commit() error {
defer ba.reset()
2016-12-09 12:41:38 +00:00
var merr MultiError
// Spill buckets into shards.
for s, b := range ba.buckets {
merr.Add(ba.db.shards[s].appendBatch(b))
}
return merr.Err()
2016-12-09 12:41:38 +00:00
}
type hashedSample struct {
hash uint64
2016-12-21 08:39:01 +00:00
labels labels.Labels
2016-12-22 11:05:24 +00:00
ref uint32
t int64
v float64
2016-12-09 15:54:38 +00:00
}
2016-12-09 09:00:14 +00:00
const sep = '\xff'
2016-12-15 07:36:09 +00:00
// Shard handles reads and writes of time series falling into
2016-12-09 09:00:14 +00:00
// a hashed shard of a series.
2016-12-15 07:36:09 +00:00
type Shard struct {
2017-01-02 21:24:35 +00:00
path string
logger log.Logger
metrics *shardMetrics
2016-12-09 09:00:14 +00:00
2016-12-15 07:31:26 +00:00
mtx sync.RWMutex
persisted persistedBlocks
head *HeadBlock
compactor *compactor
2017-01-02 21:24:35 +00:00
donec chan struct{}
persistc chan struct{}
2016-12-09 09:00:14 +00:00
}
2016-12-31 08:48:49 +00:00
type shardMetrics struct {
persistences prometheus.Counter
persistenceDuration prometheus.Histogram
samplesAppended prometheus.Counter
}
func newShardMetrics(r prometheus.Registerer, i int) *shardMetrics {
shardLabel := prometheus.Labels{
"shard": fmt.Sprintf("%d", i),
}
m := &shardMetrics{
persistences: prometheus.NewCounter(prometheus.CounterOpts{
Name: "tsdb_shard_persistences_total",
Help: "Total number of head persistances that ran so far.",
ConstLabels: shardLabel,
}),
persistenceDuration: prometheus.NewHistogram(prometheus.HistogramOpts{
Name: "tsdb_shard_persistence_duration_seconds",
Help: "Duration of persistences in seconds.",
ConstLabels: shardLabel,
Buckets: prometheus.ExponentialBuckets(0.25, 2, 5),
}),
samplesAppended: prometheus.NewCounter(prometheus.CounterOpts{
Name: "tsdb_shard_samples_appended_total",
Help: "Total number of appended samples for the shard.",
ConstLabels: shardLabel,
}),
}
if r != nil {
r.MustRegister(
m.persistences,
m.persistenceDuration,
m.samplesAppended,
)
}
return m
}
2016-12-15 07:36:09 +00:00
// OpenShard returns a new Shard.
2016-12-31 08:48:49 +00:00
func OpenShard(path string, i int, logger log.Logger) (*Shard, error) {
2016-12-15 07:31:26 +00:00
// Create directory if shard is new.
if _, err := os.Stat(path); os.IsNotExist(err) {
if err := os.MkdirAll(path, 0777); err != nil {
return nil, err
}
}
// Initialize previously persisted blocks.
2016-12-22 14:54:39 +00:00
pbs, head, err := findBlocks(path)
2016-12-15 07:31:26 +00:00
if err != nil {
return nil, err
}
2016-12-22 11:05:24 +00:00
// TODO(fabxc): get time from client-defined `now` function.
baset := time.Unix(0, 0).UnixNano() / int64(time.Millisecond)
2016-12-22 11:05:24 +00:00
if len(pbs) > 0 {
2016-12-22 14:54:39 +00:00
baset = pbs[len(pbs)-1].stats.MaxTime
2016-12-22 11:05:24 +00:00
}
2016-12-22 14:54:39 +00:00
if head == nil {
2016-12-22 19:00:24 +00:00
head, err = OpenHeadBlock(filepath.Join(path, fmt.Sprintf("%d", baset)), baset)
2016-12-22 14:54:39 +00:00
if err != nil {
return nil, err
}
2016-12-22 11:05:24 +00:00
}
2016-12-15 07:36:09 +00:00
s := &Shard{
2016-12-09 12:41:38 +00:00
path: path,
logger: logger,
2016-12-31 08:48:49 +00:00
metrics: newShardMetrics(prometheus.DefaultRegisterer, i),
2016-12-22 11:05:24 +00:00
head: head,
2016-12-15 07:31:26 +00:00
persisted: pbs,
2017-01-02 21:24:35 +00:00
persistc: make(chan struct{}, 1),
donec: make(chan struct{}),
2016-12-09 09:00:14 +00:00
}
2017-01-02 21:24:35 +00:00
if s.compactor, err = newCompactor(s, logger); err != nil {
return nil, err
}
2017-01-02 21:24:35 +00:00
go s.run()
2016-12-15 07:31:26 +00:00
return s, nil
2016-12-09 12:41:38 +00:00
}
2017-01-02 21:24:35 +00:00
func (s *Shard) run() {
for range s.persistc {
start := time.Now()
if err := s.persist(); err != nil {
s.logger.Log("msg", "persistence error", "err", err)
}
s.metrics.persistenceDuration.Observe(time.Since(start).Seconds())
s.metrics.persistences.Inc()
}
close(s.donec)
}
2016-12-15 07:36:09 +00:00
// Close the shard.
func (s *Shard) Close() error {
2017-01-02 21:24:35 +00:00
close(s.persistc)
<-s.donec
2017-01-02 21:24:35 +00:00
var merr MultiError
merr.Add(s.compactor.Close())
2016-12-15 07:31:26 +00:00
2017-01-02 21:24:35 +00:00
s.mtx.Lock()
defer s.mtx.Unlock()
2016-12-15 07:31:26 +00:00
for _, pb := range s.persisted {
2017-01-02 21:24:35 +00:00
merr.Add(pb.Close())
2016-12-15 07:31:26 +00:00
}
2017-01-02 21:24:35 +00:00
merr.Add(s.head.Close())
2016-12-15 07:31:26 +00:00
2017-01-02 21:24:35 +00:00
return merr.Err()
2016-12-09 09:00:14 +00:00
}
func (s *Shard) appendBatch(samples []hashedSample) error {
if len(samples) == 0 {
return nil
}
2016-12-09 12:41:38 +00:00
s.mtx.Lock()
defer s.mtx.Unlock()
2016-12-22 00:12:28 +00:00
// TODO(fabxc): distinguish samples between concurrent heads for
// different time blocks. Those may occurr during transition to still
// allow late samples to arrive for a previous block.
err := s.head.appendBatch(samples)
if err == nil {
s.metrics.samplesAppended.Add(float64(len(samples)))
}
2016-12-09 12:41:38 +00:00
2016-12-22 00:12:28 +00:00
// TODO(fabxc): randomize over time and use better scoring function.
2016-12-31 08:48:49 +00:00
if s.head.stats.SampleCount/(uint64(s.head.stats.ChunkCount)+1) > 400 {
2016-12-09 12:41:38 +00:00
select {
2017-01-02 21:24:35 +00:00
case s.persistc <- struct{}{}:
2016-12-09 12:41:38 +00:00
default:
}
}
2016-12-22 00:12:28 +00:00
return err
2016-12-09 12:41:38 +00:00
}
func intervalOverlap(amin, amax, bmin, bmax int64) bool {
if bmin >= amin && bmin <= amax {
return true
}
if amin >= bmin && amin <= bmax {
return true
}
return false
}
func intervalContains(min, max, t int64) bool {
return t >= min && t <= max
}
2016-12-13 14:26:58 +00:00
// blocksForRange returns all blocks within the shard that may contain
// data for the given time range.
func (s *Shard) blocksForInterval(mint, maxt int64) []block {
var bs []block
for _, b := range s.persisted {
bmin, bmax := b.interval()
if intervalOverlap(mint, maxt, bmin, bmax) {
bs = append(bs, b)
}
}
hmin, hmax := s.head.interval()
if intervalOverlap(mint, maxt, hmin, hmax) {
bs = append(bs, s.head)
}
return bs
2016-12-13 14:26:58 +00:00
}
2016-12-09 12:41:38 +00:00
// TODO(fabxc): make configurable.
const shardGracePeriod = 60 * 1000 // 60 seconds for millisecond scale
2016-12-15 07:36:09 +00:00
func (s *Shard) persist() error {
2016-12-09 12:41:38 +00:00
s.mtx.Lock()
// Set new head block.
head := s.head
2016-12-22 19:00:24 +00:00
newHead, err := OpenHeadBlock(filepath.Join(s.path, fmt.Sprintf("%d", head.stats.MaxTime)), head.stats.MaxTime)
2016-12-22 11:05:24 +00:00
if err != nil {
s.mtx.Unlock()
return err
}
s.head = newHead
2016-12-09 12:41:38 +00:00
s.mtx.Unlock()
// TODO(fabxc): add grace period where we can still append to old head shard
// before actually persisting it.
2017-01-02 13:41:13 +00:00
dir := filepath.Join(s.path, fmt.Sprintf("%d", head.stats.MinTime))
2016-12-09 12:41:38 +00:00
2017-01-02 15:58:47 +00:00
if err := persist(dir, head.persist); err != nil {
2016-12-09 20:23:34 +00:00
return err
}
2017-01-02 15:58:47 +00:00
s.logger.Log("samples", head.stats.SampleCount, "chunks", head.stats.ChunkCount, "msg", "persisted head")
2016-12-09 12:41:38 +00:00
// Reopen block as persisted block for querying.
2017-01-02 13:41:13 +00:00
pb, err := newPersistedBlock(dir)
if err != nil {
return err
}
s.mtx.Lock()
s.persisted = append(s.persisted, pb)
s.mtx.Unlock()
s.compactor.trigger()
2016-12-09 12:41:38 +00:00
return nil
}
2016-12-09 09:00:14 +00:00
// chunkDesc wraps a plain data chunk and provides cached meta data about it.
type chunkDesc struct {
ref uint32
2016-12-21 08:39:01 +00:00
lset labels.Labels
2016-12-09 09:00:14 +00:00
chunk chunks.Chunk
// Caching fields.
firsTimestamp int64
2016-12-09 09:00:14 +00:00
lastTimestamp int64
lastValue float64
2016-12-22 19:57:00 +00:00
numSamples int
2016-12-09 09:00:14 +00:00
app chunks.Appender // Current appender for the chunks.
}
2016-12-31 09:10:27 +00:00
func (cd *chunkDesc) append(ts int64, v float64) {
if cd.numSamples == 0 {
cd.firsTimestamp = ts
2016-12-09 09:00:14 +00:00
}
2016-12-31 09:10:27 +00:00
cd.app.Append(ts, v)
2016-12-09 09:00:14 +00:00
cd.lastTimestamp = ts
cd.lastValue = v
2016-12-22 19:57:00 +00:00
cd.numSamples++
2016-12-09 09:00:14 +00:00
}
// The MultiError type implements the error interface, and contains the
// Errors used to construct it.
type MultiError []error
// Returns a concatenated string of the contained errors
func (es MultiError) Error() string {
var buf bytes.Buffer
if len(es) > 0 {
fmt.Fprintf(&buf, "%d errors: ", len(es))
2016-12-08 09:04:24 +00:00
}
for i, err := range es {
if i != 0 {
buf.WriteString("; ")
}
buf.WriteString(err.Error())
}
return buf.String()
2016-11-15 09:34:25 +00:00
}
2016-12-15 07:31:26 +00:00
// Add adds the error to the error list if it is not nil.
func (es *MultiError) Add(err error) {
if err == nil {
return
}
if merr, ok := err.(MultiError); ok {
*es = append(*es, merr...)
} else {
*es = append(*es, err)
2016-12-15 07:31:26 +00:00
}
}
// Err returns the error list as an error or nil if it is empty.
2016-12-15 07:31:26 +00:00
func (es MultiError) Err() error {
if len(es) == 0 {
return nil
}
return es
}
func yoloString(b []byte) string {
sh := (*reflect.SliceHeader)(unsafe.Pointer(&b))
h := reflect.StringHeader{
Data: sh.Data,
Len: sh.Len,
}
return *((*string)(unsafe.Pointer(&h)))
}
func yoloBytes(s string) []byte {
sh := (*reflect.StringHeader)(unsafe.Pointer(&s))
h := reflect.SliceHeader{
Cap: sh.Len,
Len: sh.Len,
Data: sh.Data,
}
return *((*[]byte)(unsafe.Pointer(&h)))
}