// stop background processing of the target manager. If removeTargets is true,
// existing targets will be stopped and removed.
func (tm *TargetManager) stop(removeTargets bool) {
log.Info("Stopping target manager...")
defer log.Info("Target manager stopped.")
tm.m.Lock()
provs := []TargetProvider{}
for _, ps := range tm.providers {
provs = append(provs, ps...)
}
tm.m.Unlock()
var wg sync.WaitGroup
wg.Add(len(provs))
for _, prov := range provs {
go func(p TargetProvider) {
p.Stop()
wg.Done()
}(prov)
}
wg.Wait()
tm.m.Lock()
defer tm.m.Unlock()
if removeTargets {
tm.removeTargets(nil)
}
tm.running = false
}
func (p *persistence) rebuildLabelIndexes(
fpToSeries map[clientmodel.Fingerprint]*memorySeries,
) error {
count := 0
log.Info("Rebuilding label indexes.")
log.Info("Indexing metrics in memory.")
for fp, s := range fpToSeries {
p.indexMetric(fp, s.metric)
count++
if count%10000 == 0 {
log.Infof("%d metrics queued for indexing.", count)
}
}
log.Info("Indexing archived metrics.")
var fp codable.Fingerprint
var m codable.Metric
if err := p.archivedFingerprintToMetrics.ForEach(func(kv index.KeyValueAccessor) error {
if err := kv.Key(&fp); err != nil {
return err
}
if err := kv.Value(&m); err != nil {
return err
}
p.indexMetric(clientmodel.Fingerprint(fp), clientmodel.Metric(m))
count++
if count%10000 == 0 {
log.Infof("%d metrics queued for indexing.", count)
}
return nil
}); err != nil {
return err
}
log.Info("All requests for rebuilding the label indexes queued. (Actual processing may lag behind.)")
return nil
}
// stop background processing of the target manager. If removeTargets is true,
// existing targets will be stopped and removed.
func (tm *TargetManager) stop(removeTargets bool) {
log.Info("Stopping target manager...")
defer log.Info("Target manager stopped.")
close(tm.done)
tm.mtx.Lock()
defer tm.mtx.Unlock()
if removeTargets {
tm.removeTargets(nil)
}
tm.running = false
}
func (w *fileWatcher) Watch(cb ReloadCallback) {
watcher, err := fsnotify.NewWatcher()
if err != nil {
log.Fatal(err)
}
err = watcher.WatchFlags(w.fileName, fsnotify.FSN_MODIFY)
if err != nil {
log.Fatal(err)
}
for {
select {
case ev := <-watcher.Event:
log.Infof("Config file changed (%s), attempting reload", ev)
conf, err := LoadFromFile(w.fileName)
if err != nil {
log.Error("Error loading new config: ", err)
failedConfigReloads.Inc()
} else {
cb(&conf)
log.Info("Config reloaded successfully")
configReloads.Inc()
}
// Re-add the file watcher since it can get lost on some changes. E.g.
// saving a file with vim results in a RENAME-MODIFY-DELETE event
// sequence, after which the newly written file is no longer watched.
err = watcher.WatchFlags(w.fileName, fsnotify.FSN_MODIFY)
case err := <-watcher.Error:
log.Error("Error watching config: ", err)
}
}
}
func newMesosExporter(opts *exporterOpts) *periodicExporter {
e := &periodicExporter{
errors: prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "mesos_exporter",
Name: "slave_scrape_errors_total",
Help: "Current total scrape errors",
},
[]string{"slave"},
),
opts: opts,
}
e.slaves.urls = []string{e.opts.localURL}
if e.opts.autoDiscover {
log.Info("auto discovery enabled from command line flag.")
// Update nr. of mesos slaves every 10 minutes
e.updateSlaves()
go runEvery(e.updateSlaves, 10*time.Minute)
}
// Fetch slave metrics every interval
go runEvery(e.scrapeSlaves, e.opts.interval)
return e
}
// GetJson return json from server
func GetJson(url string, accessKey string, secretKey string, target interface{}) error {
start := time.Now()
// Counter for internal exporter metrics
measure.FunctionCountTotal.With(prometheus.Labels{"pkg": "utils", "fnc": "GetJson"}).Inc()
log.Info("Scraping: ", url)
client := &http.Client{}
req, err := http.NewRequest("GET", url, nil)
req.SetBasicAuth(accessKey, secretKey)
resp, err := client.Do(req)
if err != nil {
log.Error("Error Collecting JSON from API: ", err)
panic(err)
}
// Timings recorded as part of internal metrics
elapsed := float64((time.Since(start)) / time.Microsecond)
measure.FunctionDurations.WithLabelValues("hosts", "getJSON").Observe(elapsed)
return json.NewDecoder(resp.Body).Decode(target)
}
// Run starts background processing to handle target updates.
func (tm *TargetManager) Run() {
log.Info("Starting target manager...")
sources := map[string]struct{}{}
for scfg, provs := range tm.providers {
for _, prov := range provs {
ch := make(chan *config.TargetGroup)
go tm.handleTargetUpdates(scfg, ch)
for _, src := range prov.Sources() {
src = fullSource(scfg, src)
sources[src] = struct{}{}
}
// Run the target provider after cleanup of the stale targets is done.
defer func(p TargetProvider, c chan *config.TargetGroup) {
go p.Run(c)
}(prov, ch)
}
}
tm.m.Lock()
defer tm.m.Unlock()
tm.removeTargets(func(src string) bool {
if _, ok := sources[src]; ok {
return false
}
return true
})
tm.running = true
}
// Run starts background processing to handle target updates.
func (tm *TargetManager) Run() {
log.Info("Starting target manager...")
tm.done = make(chan struct{})
sources := map[string]struct{}{}
updates := []<-chan targetGroupUpdate{}
for scfg, provs := range tm.providers {
for _, prov := range provs {
// Get an initial set of available sources so we don't remove
// target groups from the last run that are still available.
for _, src := range prov.Sources() {
sources[src] = struct{}{}
}
tgc := make(chan *config.TargetGroup)
// Run the target provider after cleanup of the stale targets is done.
defer func(prov TargetProvider, tgc chan *config.TargetGroup) {
go prov.Run(tgc, tm.done)
}(prov, tgc)
tgupc := make(chan targetGroupUpdate)
updates = append(updates, tgupc)
go func(scfg *config.ScrapeConfig) {
defer close(tgupc)
for {
select {
case tg := <-tgc:
if tg == nil {
break
}
tgupc <- targetGroupUpdate{tg: tg, scfg: scfg}
case <-tm.done:
return
}
}
}(scfg)
}
}
// Merge all channels of incoming target group updates into a single
// one and keep applying the updates.
go tm.handleUpdates(merge(tm.done, updates...), tm.done)
tm.mtx.Lock()
defer tm.mtx.Unlock()
// Remove old target groups that are no longer in the set of sources.
tm.removeTargets(func(src string) bool {
if _, ok := sources[src]; ok {
return false
}
return true
})
tm.running = true
}
func startPolling() {
t := time.NewTicker(time.Duration(*pollRate) * time.Minute)
log.Info("Starting polling of tanks.")
for {
readTanks()
<-t.C
}
}
// checkpointFPMappings persists the fingerprint mappings. This method is not
// goroutine-safe.
//
// Description of the file format, v1:
//
// (1) Magic string (const mappingsMagicString).
//
// (2) Uvarint-encoded format version (const mappingsFormatVersion).
//
// (3) Uvarint-encoded number of mappings in fpMappings.
//
// (4) Repeated once per mapping:
//
// (4.1) The raw fingerprint as big-endian uint64.
//
// (4.2) The uvarint-encoded number of sub-mappings for the raw fingerprint.
//
// (4.3) Repeated once per sub-mapping:
//
// (4.3.1) The uvarint-encoded length of the unique metric string.
// (4.3.2) The unique metric string.
// (4.3.3) The mapped fingerprint as big-endian uint64.
func (p *persistence) checkpointFPMappings(fpm fpMappings) (err error) {
log.Info("Checkpointing fingerprint mappings...")
begin := time.Now()
f, err := os.OpenFile(p.mappingsTempFileName(), os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0640)
if err != nil {
return
}
defer func() {
f.Sync()
closeErr := f.Close()
if err != nil {
return
}
err = closeErr
if err != nil {
return
}
err = os.Rename(p.mappingsTempFileName(), p.mappingsFileName())
duration := time.Since(begin)
log.Infof("Done checkpointing fingerprint mappings in %v.", duration)
}()
w := bufio.NewWriterSize(f, fileBufSize)
if _, err = w.WriteString(mappingsMagicString); err != nil {
return
}
if _, err = codable.EncodeUvarint(w, mappingsFormatVersion); err != nil {
return
}
if _, err = codable.EncodeUvarint(w, uint64(len(fpm))); err != nil {
return
}
for fp, mappings := range fpm {
if err = codable.EncodeUint64(w, uint64(fp)); err != nil {
return
}
if _, err = codable.EncodeUvarint(w, uint64(len(mappings))); err != nil {
return
}
for ms, mappedFP := range mappings {
if _, err = codable.EncodeUvarint(w, uint64(len(ms))); err != nil {
return
}
if _, err = w.WriteString(ms); err != nil {
return
}
if err = codable.EncodeUint64(w, uint64(mappedFP)); err != nil {
return
}
}
}
err = w.Flush()
return
}
// Stop stops sending samples to the remote storage and waits for pending
// sends to complete.
func (t *StorageQueueManager) Stop() {
log.Infof("Stopping remote storage...")
close(t.queue)
<-t.drained
for i := 0; i < maxConcurrentSends; i++ {
t.sendSemaphore <- true
}
log.Info("Remote storage stopped.")
}
// StacksURLCheck - Checks the API version for Rancher to determine the correct URL
func StacksURLCheck(rancherURL string) string {
var stacksEndpoint string
if strings.Contains(rancherURL, "v1") {
log.Info("Version 1 API detected, using legacy API fields")
stacksEndpoint = (rancherURL + "/environments/")
} else if strings.Contains(rancherURL, "v2") {
log.Info("Version 2 API detected, using updated API fields")
stacksEndpoint = (rancherURL + "/stacks/")
} else {
log.Info("No known API version detected, defaulting to /stacks/")
stacksEndpoint = (rancherURL + "/stacks/")
}
return stacksEndpoint
}
func newMesosExporter(opts *exporterOpts) *periodicExporter {
e := &periodicExporter{
errors: prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "mesos_exporter",
Name: "slave_scrape_errors_total",
Help: "Current total scrape errors",
},
[]string{"slave"},
),
opts: opts,
}
if opts.queryURL == "" {
log.Fatal("Flag '-exporter.url' not set")
}
switch opts.mode {
case "discover":
log.Info("starting mesos_exporter in scrape mode 'discover'")
e.queryURL = parseMasterURL(opts.queryURL)
// Update nr. of mesos slaves.
e.updateSlaves()
go runEvery(e.updateSlaves, e.opts.autoDiscoverInterval)
// Fetch slave metrics every interval.
go runEvery(e.scrapeSlaves, e.opts.interval)
case "master":
log.Info("starting mesos_exporter in scrape mode 'master'")
e.queryURL = parseMasterURL(opts.queryURL)
case "slave":
log.Info("starting mesos_exporter in scrape mode 'slave'")
e.slaves.urls = []string{opts.queryURL}
default:
log.Fatalf("Invalid value '%s' of flag '-exporter.mode' - must be one of 'discover', 'master' or 'slave'", opts.mode)
}
return e
}
// Stop implements Storage.
func (s *memorySeriesStorage) Stop() error {
log.Info("Stopping local storage...")
log.Info("Stopping maintenance loop...")
close(s.loopStopping)
<-s.loopStopped
log.Info("Stopping chunk eviction...")
close(s.evictStopping)
<-s.evictStopped
// One final checkpoint of the series map and the head chunks.
if err := s.persistence.checkpointSeriesMapAndHeads(s.fpToSeries, s.fpLocker); err != nil {
return err
}
if err := s.persistence.close(); err != nil {
return err
}
log.Info("Local storage stopped.")
return nil
}
func (s *memorySeriesStorage) loop() {
checkpointTimer := time.NewTimer(s.checkpointInterval)
dirtySeriesCount := 0
defer func() {
checkpointTimer.Stop()
log.Info("Maintenance loop stopped.")
close(s.loopStopped)
}()
memoryFingerprints := s.cycleThroughMemoryFingerprints()
archivedFingerprints := s.cycleThroughArchivedFingerprints()
loop:
for {
select {
case <-s.loopStopping:
break loop
case <-checkpointTimer.C:
err := s.persistence.checkpointSeriesMapAndHeads(s.fpToSeries, s.fpLocker)
if err != nil {
log.Errorln("Error while checkpointing:", err)
} else {
dirtySeriesCount = 0
}
checkpointTimer.Reset(s.checkpointInterval)
case fp := <-memoryFingerprints:
if s.maintainMemorySeries(fp, model.Now().Add(-s.dropAfter)) {
dirtySeriesCount++
// Check if we have enough "dirty" series so that we need an early checkpoint.
// However, if we are already behind persisting chunks, creating a checkpoint
// would be counterproductive, as it would slow down chunk persisting even more,
// while in a situation like that, where we are clearly lacking speed of disk
// maintenance, the best we can do for crash recovery is to persist chunks as
// quickly as possible. So only checkpoint if the storage is not in "graceful
// degradation mode".
if dirtySeriesCount >= s.checkpointDirtySeriesLimit && !s.isDegraded() {
checkpointTimer.Reset(0)
}
}
case fp := <-archivedFingerprints:
s.maintainArchivedSeries(fp, model.Now().Add(-s.dropAfter))
}
}
// Wait until both channels are closed.
for range memoryFingerprints {
}
for range archivedFingerprints {
}
}
func main() {
flag.Parse()
exporter := newZooKeeperExporter(flag.Args(), *useExhibitor)
prometheus.MustRegister(exporter)
http.Handle(*metricPath, prometheus.Handler())
http.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
http.Redirect(w, r, *metricPath, http.StatusMovedPermanently)
})
log.Info("starting mesos_exporter on ", *addr)
log.Fatal(http.ListenAndServe(*addr, nil))
}
func main() {
flag.Parse()
if rancherURL == "" {
log.Fatal("CATTLE_URL must be set and non-empty")
}
log.Info("Starting Prometheus Exporter for Rancher. Listen Address: ", listenAddress, " metricsPath: ", metricsPath, " rancherURL: ", rancherURL, " AccessKey: ", accessKey)
log.Info("System Services Reported on:", hideSys)
// Register internal metrics
measure.Init()
// Pass URL & Credentials out to the Exporters
servicesExporter := services.NewExporter(rancherURL, accessKey, secretKey, hideSys)
stacksExporter := stacks.NewExporter(rancherURL, accessKey, secretKey, hideSys)
hostsExporter := hosts.NewExporter(rancherURL, accessKey, secretKey)
// Register Metrics from each of the endpoints
prometheus.MustRegister(servicesExporter)
prometheus.MustRegister(stacksExporter)
prometheus.MustRegister(hostsExporter)
http.Handle(metricsPath, prometheus.Handler())
http.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
w.Write([]byte(`<html>
<head><title>Rancher exporter</title></head>
<body>
<h1>rancher exporter</h1>
<p><a href='` + metricsPath + `'>Metrics</a></p>
</body>
</html>
`))
})
log.Infof("Starting Server: %s", listenAddress)
log.Fatal(http.ListenAndServe(listenAddress, nil))
}
// Start implements Storage.
func (s *memorySeriesStorage) Start() (err error) {
var syncStrategy syncStrategy
switch s.options.SyncStrategy {
case Never:
syncStrategy = func() bool { return false }
case Always:
syncStrategy = func() bool { return true }
case Adaptive:
syncStrategy = func() bool { return !s.isDegraded() }
default:
panic("unknown sync strategy")
}
var p *persistence
p, err = newPersistence(s.options.PersistenceStoragePath, s.options.Dirty, s.options.PedanticChecks, syncStrategy)
if err != nil {
return err
}
s.persistence = p
// Persistence must start running before loadSeriesMapAndHeads() is called.
go s.persistence.run()
defer func() {
if err != nil {
if e := p.close(); e != nil {
log.Errorln("Error closing persistence:", e)
}
}
}()
log.Info("Loading series map and head chunks...")
s.fpToSeries, s.numChunksToPersist, err = p.loadSeriesMapAndHeads()
if err != nil {
return err
}
log.Infof("%d series loaded.", s.fpToSeries.length())
s.numSeries.Set(float64(s.fpToSeries.length()))
s.mapper, err = newFPMapper(s.fpToSeries, p)
if err != nil {
return err
}
go s.handleEvictList()
go s.loop()
return nil
}
func (s *memorySeriesStorage) handleEvictList() {
ticker := time.NewTicker(maxEvictInterval)
count := 0
for {
// To batch up evictions a bit, this tries evictions at least
// once per evict interval, but earlier if the number of evict
// requests with evict==true that have happened since the last
// evict run is more than maxMemoryChunks/1000.
select {
case req := <-s.evictRequests:
if req.evict {
req.cd.evictListElement = s.evictList.PushBack(req.cd)
count++
if count > s.maxMemoryChunks/1000 {
s.maybeEvict()
count = 0
}
} else {
if req.cd.evictListElement != nil {
s.evictList.Remove(req.cd.evictListElement)
req.cd.evictListElement = nil
}
}
case <-ticker.C:
if s.evictList.Len() > 0 {
s.maybeEvict()
}
case <-s.evictStopping:
// Drain evictRequests forever in a goroutine to not let
// requesters hang.
go func() {
for {
<-s.evictRequests
}
}()
ticker.Stop()
log.Info("Chunk eviction stopped.")
close(s.evictStopped)
return
}
}
}
func testChunk(t *testing.T, encoding chunkEncoding) {
samples := make(model.Samples, 500000)
for i := range samples {
samples[i] = &model.Sample{
Timestamp: model.Time(i),
Value: model.SampleValue(float64(i) * 0.2),
}
}
s, closer := NewTestStorage(t, encoding)
defer closer.Close()
for _, sample := range samples {
s.Append(sample)
}
s.WaitForIndexing()
for m := range s.fpToSeries.iter() {
s.fpLocker.Lock(m.fp)
var values []model.SamplePair
for _, cd := range m.series.chunkDescs {
if cd.isEvicted() {
continue
}
for sample := range cd.c.newIterator().values() {
values = append(values, *sample)
}
}
for i, v := range values {
if samples[i].Timestamp != v.Timestamp {
t.Errorf("%d. Got %v; want %v", i, v.Timestamp, samples[i].Timestamp)
}
if samples[i].Value != v.Value {
t.Errorf("%d. Got %v; want %v", i, v.Value, samples[i].Value)
}
}
s.fpLocker.Unlock(m.fp)
}
log.Info("test done, closing")
}
func (e *Exporter) gatherMetrics(rancherURL string, accessKey string, secretKey string, ch chan<- prometheus.Metric) error {
// Reset guageVecs back to 0
for _, m := range e.gaugeVecs {
m.Reset()
}
// Set the correct API endpoint for hosts
endpoint := (rancherURL + "/hosts/")
// Scrape EndPoint for JSON Data
data := new(Data)
err := utils.GetJson(endpoint, accessKey, secretKey, &data)
if err != nil {
log.Error("Error getting JSON from URL ", endpoint)
return err
}
log.Info("JSON Fetched for hosts: ", data)
// Host Metrics
for _, x := range data.Data {
// Pre-defines the known states from the Rancher API
states := []string{"activating", "active", "deactivating", "error", "erroring", "inactive", "provisioned", "purged", "purging", "registering", "removed", "removing", "requested", "restoring", "updating_active", "updating_inactive"}
// Set the state of the service to 1 when it matches one of the known states
for _, y := range states {
if x.State == y {
e.gaugeVecs["HostState"].With(prometheus.Labels{"rancherURL": rancherURL, "name": x.Hostname, "state": y}).Set(1)
} else {
e.gaugeVecs["HostState"].With(prometheus.Labels{"rancherURL": rancherURL, "name": x.Hostname, "state": y}).Set(0)
}
}
}
return nil
}
func (w WebService) ServeForever(pathPrefix string) error {
http.Handle(pathPrefix+"favicon.ico", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
http.Error(w, "", 404)
}))
http.HandleFunc("/", prometheus.InstrumentHandlerFunc("index", func(rw http.ResponseWriter, req *http.Request) {
// The "/" pattern matches everything, so we need to check
// that we're at the root here.
if req.URL.Path == pathPrefix {
w.AlertsHandler.ServeHTTP(rw, req)
} else if req.URL.Path == strings.TrimRight(pathPrefix, "/") {
http.Redirect(rw, req, pathPrefix, http.StatusFound)
} else if !strings.HasPrefix(req.URL.Path, pathPrefix) {
// We're running under a prefix but the user requested something
// outside of it. Let's see if this page exists under the prefix.
http.Redirect(rw, req, pathPrefix+strings.TrimLeft(req.URL.Path, "/"), http.StatusFound)
} else {
http.NotFound(rw, req)
}
}))
http.Handle(pathPrefix+"alerts", prometheus.InstrumentHandler("alerts", w.AlertsHandler))
http.Handle(pathPrefix+"silences", prometheus.InstrumentHandler("silences", w.SilencesHandler))
http.Handle(pathPrefix+"status", prometheus.InstrumentHandler("status", w.StatusHandler))
http.Handle(pathPrefix+"metrics", prometheus.Handler())
if *useLocalAssets {
http.Handle(pathPrefix+"static/", http.StripPrefix(pathPrefix+"static/", http.FileServer(http.Dir("web/static"))))
} else {
http.Handle(pathPrefix+"static/", http.StripPrefix(pathPrefix+"static/", new(blob.Handler)))
}
http.Handle(pathPrefix+"api/", w.AlertManagerService.Handler())
log.Info("listening on ", *listenAddress)
return http.ListenAndServe(*listenAddress, nil)
}
// Run the rule manager's periodic rule evaluation.
func (m *Manager) Run() {
defer log.Info("Rule manager stopped.")
m.Lock()
lastInterval := m.interval
m.Unlock()
ticker := time.NewTicker(lastInterval)
defer ticker.Stop()
for {
// The outer select clause makes sure that m.done is looked at
// first. Otherwise, if m.runIteration takes longer than
// m.interval, there is only a 50% chance that m.done will be
// looked at before the next m.runIteration call happens.
select {
case <-m.done:
return
default:
select {
case <-ticker.C:
start := time.Now()
m.runIteration()
iterationDuration.Observe(float64(time.Since(start) / time.Millisecond))
m.Lock()
if lastInterval != m.interval {
ticker.Stop()
ticker = time.NewTicker(m.interval)
lastInterval = m.interval
}
m.Unlock()
case <-m.done:
return
}
}
}
}
func main() {
flag.Parse()
opts := &exporterOpts{
autoDiscoverInterval: *autoDiscoverInterval,
interval: *scrapeInterval,
mode: *scrapeMode,
queryURL: strings.TrimRight(*queryURL, "/"),
}
exporter := newMesosExporter(opts)
prometheus.MustRegister(exporter)
http.Handle(*metricsPath, prometheus.Handler())
http.HandleFunc("/status", func(w http.ResponseWriter, r *http.Request) {
fmt.Fprintf(w, "OK")
})
http.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
http.Redirect(w, r, *metricsPath, http.StatusMovedPermanently)
})
log.Info("starting mesos_exporter on ", *addr)
log.Fatal(http.ListenAndServe(*addr, nil))
}
// Stop the rule manager's rule evaluation cycles.
func (m *Manager) Stop() {
log.Info("Stopping rule manager...")
m.done <- true
}
// recoverFromCrash is called by loadSeriesMapAndHeads if the persistence
// appears to be dirty after the loading (either because the loading resulted in
// an error or because the persistence was dirty from the start). Not goroutine
// safe. Only call before anything else is running (except index processing
// queue as started by newPersistence).
func (p *persistence) recoverFromCrash(fingerprintToSeries map[clientmodel.Fingerprint]*memorySeries) error {
// TODO(beorn): We need proper tests for the crash recovery.
log.Warn("Starting crash recovery. Prometheus is inoperational until complete.")
log.Warn("To avoid crash recovery in future, shutdown Prometheus with SIGTERM or a HTTP POST to /-/quit.")
fpsSeen := map[clientmodel.Fingerprint]struct{}{}
count := 0
seriesDirNameFmt := fmt.Sprintf("%%0%dx", seriesDirNameLen)
// Delete the fingerprint mapping file as it might be stale or
// corrupt. We'll rebuild the mappings as we go.
os.Remove(p.mappingsFileName())
// The mappings to rebuild.
fpm := fpMappings{}
log.Info("Scanning files.")
for i := 0; i < 1<<(seriesDirNameLen*4); i++ {
dirname := path.Join(p.basePath, fmt.Sprintf(seriesDirNameFmt, i))
dir, err := os.Open(dirname)
if os.IsNotExist(err) {
continue
}
if err != nil {
return err
}
defer dir.Close()
for fis := []os.FileInfo{}; err != io.EOF; fis, err = dir.Readdir(1024) {
if err != nil {
return err
}
for _, fi := range fis {
fp, ok := p.sanitizeSeries(dirname, fi, fingerprintToSeries, fpm)
if ok {
fpsSeen[fp] = struct{}{}
}
count++
if count%10000 == 0 {
log.Infof("%d files scanned.", count)
}
}
}
}
log.Infof("File scan complete. %d series found.", len(fpsSeen))
log.Info("Checking for series without series file.")
for fp, s := range fingerprintToSeries {
if _, seen := fpsSeen[fp]; !seen {
// fp exists in fingerprintToSeries, but has no representation on disk.
if s.persistWatermark == len(s.chunkDescs) {
// Oops, everything including the head chunk was
// already persisted, but nothing on disk.
// Thus, we lost that series completely. Clean
// up the remnants.
delete(fingerprintToSeries, fp)
if err := p.purgeArchivedMetric(fp); err != nil {
// Purging the archived metric didn't work, so try
// to unindex it, just in case it's in the indexes.
p.unindexMetric(fp, s.metric)
}
log.Warnf("Lost series detected: fingerprint %v, metric %v.", fp, s.metric)
continue
}
// If we are here, the only chunks we have are the chunks in the checkpoint.
// Adjust things accordingly.
if s.persistWatermark > 0 || s.chunkDescsOffset != 0 {
minLostChunks := s.persistWatermark + s.chunkDescsOffset
if minLostChunks <= 0 {
log.Warnf(
"Possible loss of chunks for fingerprint %v, metric %v.",
fp, s.metric,
)
} else {
log.Warnf(
"Lost at least %d chunks for fingerprint %v, metric %v.",
minLostChunks, fp, s.metric,
)
}
s.chunkDescs = append(
make([]*chunkDesc, 0, len(s.chunkDescs)-s.persistWatermark),
s.chunkDescs[s.persistWatermark:]...,
)
numMemChunkDescs.Sub(float64(s.persistWatermark))
s.persistWatermark = 0
s.chunkDescsOffset = 0
}
maybeAddMapping(fp, s.metric, fpm)
fpsSeen[fp] = struct{}{} // Add so that fpsSeen is complete.
}
}
log.Info("Check for series without series file complete.")
if err := p.cleanUpArchiveIndexes(fingerprintToSeries, fpsSeen, fpm); err != nil {
return err
}
if err := p.rebuildLabelIndexes(fingerprintToSeries); err != nil {
return err
}
// Finally rewrite the mappings file if there are any mappings.
if len(fpm) > 0 {
if err := p.checkpointFPMappings(fpm); err != nil {
return err
}
}
p.setDirty(false)
log.Warn("Crash recovery complete.")
return nil
}
// Main manages the startup and shutdown lifecycle of the entire Prometheus server.
func Main() int {
if err := parse(os.Args[1:]); err != nil {
return 2
}
printVersion()
if cfg.printVersion {
return 0
}
var reloadables []Reloadable
var (
memStorage = local.NewMemorySeriesStorage(&cfg.storage)
remoteStorage = remote.New(&cfg.remote)
sampleAppender = storage.Fanout{memStorage}
)
if remoteStorage != nil {
sampleAppender = append(sampleAppender, remoteStorage)
reloadables = append(reloadables, remoteStorage)
}
var (
notificationHandler = notification.NewNotificationHandler(&cfg.notification)
targetManager = retrieval.NewTargetManager(sampleAppender)
queryEngine = promql.NewEngine(memStorage, &cfg.queryEngine)
)
ruleManager := rules.NewManager(&rules.ManagerOptions{
SampleAppender: sampleAppender,
NotificationHandler: notificationHandler,
QueryEngine: queryEngine,
ExternalURL: cfg.web.ExternalURL,
})
flags := map[string]string{}
cfg.fs.VisitAll(func(f *flag.Flag) {
flags[f.Name] = f.Value.String()
})
status := &web.PrometheusStatus{
TargetPools: targetManager.Pools,
Rules: ruleManager.Rules,
Flags: flags,
Birth: time.Now(),
}
webHandler := web.New(memStorage, queryEngine, ruleManager, status, &cfg.web)
reloadables = append(reloadables, status, targetManager, ruleManager, webHandler, notificationHandler)
if !reloadConfig(cfg.configFile, reloadables...) {
return 1
}
// Wait for reload or termination signals. Start the handler for SIGHUP as
// early as possible, but ignore it until we are ready to handle reloading
// our config.
hup := make(chan os.Signal)
hupReady := make(chan bool)
signal.Notify(hup, syscall.SIGHUP)
go func() {
<-hupReady
for {
select {
case <-hup:
case <-webHandler.Reload():
}
reloadConfig(cfg.configFile, reloadables...)
}
}()
// Start all components.
if err := memStorage.Start(); err != nil {
log.Errorln("Error opening memory series storage:", err)
return 1
}
defer func() {
if err := memStorage.Stop(); err != nil {
log.Errorln("Error stopping storage:", err)
}
}()
if remoteStorage != nil {
prometheus.MustRegister(remoteStorage)
go remoteStorage.Run()
defer remoteStorage.Stop()
}
// The storage has to be fully initialized before registering.
prometheus.MustRegister(memStorage)
prometheus.MustRegister(notificationHandler)
prometheus.MustRegister(configSuccess)
prometheus.MustRegister(configSuccessTime)
go ruleManager.Run()
defer ruleManager.Stop()
go notificationHandler.Run()
defer notificationHandler.Stop()
go targetManager.Run()
defer targetManager.Stop()
defer queryEngine.Stop()
go webHandler.Run()
// Wait for reload or termination signals.
close(hupReady) // Unblock SIGHUP handler.
term := make(chan os.Signal)
signal.Notify(term, os.Interrupt, syscall.SIGTERM)
select {
case <-term:
log.Warn("Received SIGTERM, exiting gracefully...")
case <-webHandler.Quit():
log.Warn("Received termination request via web service, exiting gracefully...")
case err := <-webHandler.ListenError():
log.Errorln("Error starting web server, exiting gracefully:", err)
}
log.Info("See you next time!")
return 0
}
func Main() int {
if err := parse(os.Args[1:]); err != nil {
return 2
}
versionInfoTmpl.Execute(os.Stdout, BuildInfo)
if cfg.printVersion {
return 0
}
memStorage := local.NewMemorySeriesStorage(&cfg.storage)
var (
sampleAppender storage.SampleAppender
remoteStorageQueues []*remote.StorageQueueManager
)
if cfg.opentsdbURL == "" && cfg.influxdbURL == "" {
log.Warnf("No remote storage URLs provided; not sending any samples to long-term storage")
sampleAppender = memStorage
} else {
fanout := storage.Fanout{memStorage}
addRemoteStorage := func(c remote.StorageClient) {
qm := remote.NewStorageQueueManager(c, 100*1024)
fanout = append(fanout, qm)
remoteStorageQueues = append(remoteStorageQueues, qm)
}
if cfg.opentsdbURL != "" {
addRemoteStorage(opentsdb.NewClient(cfg.opentsdbURL, cfg.remoteStorageTimeout))
}
if cfg.influxdbURL != "" {
addRemoteStorage(influxdb.NewClient(cfg.influxdbURL, cfg.remoteStorageTimeout, cfg.influxdbDatabase, cfg.influxdbRetentionPolicy))
}
sampleAppender = fanout
}
var (
notificationHandler = notification.NewNotificationHandler(&cfg.notification)
targetManager = retrieval.NewTargetManager(sampleAppender)
queryEngine = promql.NewEngine(memStorage, &cfg.queryEngine)
)
ruleManager := rules.NewManager(&rules.ManagerOptions{
SampleAppender: sampleAppender,
NotificationHandler: notificationHandler,
QueryEngine: queryEngine,
PrometheusURL: cfg.prometheusURL,
PathPrefix: cfg.web.PathPrefix,
})
flags := map[string]string{}
cfg.fs.VisitAll(func(f *flag.Flag) {
flags[f.Name] = f.Value.String()
})
status := &web.PrometheusStatus{
BuildInfo: BuildInfo,
TargetPools: targetManager.Pools,
Rules: ruleManager.Rules,
Flags: flags,
Birth: time.Now(),
}
webHandler := web.New(memStorage, queryEngine, ruleManager, status, &cfg.web)
if !reloadConfig(cfg.configFile, status, targetManager, ruleManager) {
os.Exit(1)
}
// Wait for reload or termination signals. Start the handler for SIGHUP as
// early as possible, but ignore it until we are ready to handle reloading
// our config.
hup := make(chan os.Signal)
hupReady := make(chan bool)
signal.Notify(hup, syscall.SIGHUP)
go func() {
<-hupReady
for range hup {
reloadConfig(cfg.configFile, status, targetManager, ruleManager)
}
}()
// Start all components.
if err := memStorage.Start(); err != nil {
log.Errorln("Error opening memory series storage:", err)
return 1
}
defer func() {
if err := memStorage.Stop(); err != nil {
log.Errorln("Error stopping storage:", err)
}
}()
// The storage has to be fully initialized before registering.
registry.MustRegister(memStorage)
registry.MustRegister(notificationHandler)
for _, q := range remoteStorageQueues {
registry.MustRegister(q)
go q.Run()
defer q.Stop()
}
go ruleManager.Run()
defer ruleManager.Stop()
go notificationHandler.Run()
defer notificationHandler.Stop()
go targetManager.Run()
defer targetManager.Stop()
defer queryEngine.Stop()
go webHandler.Run()
// Wait for reload or termination signals.
close(hupReady) // Unblock SIGHUP handler.
term := make(chan os.Signal)
signal.Notify(term, os.Interrupt, syscall.SIGTERM)
select {
case <-term:
log.Warn("Received SIGTERM, exiting gracefully...")
case <-webHandler.Quit():
log.Warn("Received termination request via web service, exiting gracefully...")
}
close(hup)
log.Info("See you next time!")
return 0
}
// Stop shuts down the notification handler.
func (n *NotificationHandler) Stop() {
log.Info("Stopping notification handler...")
close(n.pendingNotifications)
<-n.stopped
log.Info("Notification handler stopped.")
}
// checkpointSeriesMapAndHeads persists the fingerprint to memory-series mapping
// and all non persisted chunks. Do not call concurrently with
// loadSeriesMapAndHeads. This method will only write heads format v2, but
// loadSeriesMapAndHeads can also understand v1.
//
// Description of the file format (for both, v1 and v2):
//
// (1) Magic string (const headsMagicString).
//
// (2) Varint-encoded format version (const headsFormatVersion).
//
// (3) Number of series in checkpoint as big-endian uint64.
//
// (4) Repeated once per series:
//
// (4.1) A flag byte, see flag constants above. (Present but unused in v2.)
//
// (4.2) The fingerprint as big-endian uint64.
//
// (4.3) The metric as defined by codable.Metric.
//
// (4.4) The varint-encoded persistWatermark. (Missing in v1.)
//
// (4.5) The modification time of the series file as nanoseconds elapsed since
// January 1, 1970 UTC. -1 if the modification time is unknown or no series file
// exists yet. (Missing in v1.)
//
// (4.6) The varint-encoded chunkDescsOffset.
//
// (4.6) The varint-encoded savedFirstTime.
//
// (4.7) The varint-encoded number of chunk descriptors.
//
// (4.8) Repeated once per chunk descriptor, oldest to most recent, either
// variant 4.8.1 (if index < persistWatermark) or variant 4.8.2 (if index >=
// persistWatermark). In v1, everything is variant 4.8.1 except for a
// non-persisted head-chunk (determined by the flags).
//
// (4.8.1.1) The varint-encoded first time.
// (4.8.1.2) The varint-encoded last time.
//
// (4.8.2.1) A byte defining the chunk type.
// (4.8.2.2) The chunk itself, marshaled with the marshal() method.
//
func (p *persistence) checkpointSeriesMapAndHeads(fingerprintToSeries *seriesMap, fpLocker *fingerprintLocker) (err error) {
log.Info("Checkpointing in-memory metrics and chunks...")
begin := time.Now()
f, err := os.OpenFile(p.headsTempFileName(), os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0640)
if err != nil {
return
}
defer func() {
f.Sync()
closeErr := f.Close()
if err != nil {
return
}
err = closeErr
if err != nil {
return
}
err = os.Rename(p.headsTempFileName(), p.headsFileName())
duration := time.Since(begin)
p.checkpointDuration.Set(float64(duration) / float64(time.Millisecond))
log.Infof("Done checkpointing in-memory metrics and chunks in %v.", duration)
}()
w := bufio.NewWriterSize(f, fileBufSize)
if _, err = w.WriteString(headsMagicString); err != nil {
return
}
var numberOfSeriesOffset int
if numberOfSeriesOffset, err = codable.EncodeVarint(w, headsFormatVersion); err != nil {
return
}
numberOfSeriesOffset += len(headsMagicString)
numberOfSeriesInHeader := uint64(fingerprintToSeries.length())
// We have to write the number of series as uint64 because we might need
// to overwrite it later, and a varint might change byte width then.
if err = codable.EncodeUint64(w, numberOfSeriesInHeader); err != nil {
return
}
iter := fingerprintToSeries.iter()
defer func() {
// Consume the iterator in any case to not leak goroutines.
for range iter {
}
}()
var realNumberOfSeries uint64
for m := range iter {
func() { // Wrapped in function to use defer for unlocking the fp.
fpLocker.Lock(m.fp)
defer fpLocker.Unlock(m.fp)
if len(m.series.chunkDescs) == 0 {
// This series was completely purged or archived in the meantime. Ignore.
return
}
realNumberOfSeries++
// seriesFlags left empty in v2.
if err = w.WriteByte(0); err != nil {
return
}
if err = codable.EncodeUint64(w, uint64(m.fp)); err != nil {
return
}
var buf []byte
buf, err = codable.Metric(m.series.metric).MarshalBinary()
if err != nil {
return
}
w.Write(buf)
if _, err = codable.EncodeVarint(w, int64(m.series.persistWatermark)); err != nil {
return
}
if m.series.modTime.IsZero() {
if _, err = codable.EncodeVarint(w, -1); err != nil {
return
}
} else {
if _, err = codable.EncodeVarint(w, m.series.modTime.UnixNano()); err != nil {
return
}
}
if _, err = codable.EncodeVarint(w, int64(m.series.chunkDescsOffset)); err != nil {
return
}
if _, err = codable.EncodeVarint(w, int64(m.series.savedFirstTime)); err != nil {
return
}
if _, err = codable.EncodeVarint(w, int64(len(m.series.chunkDescs))); err != nil {
return
}
for i, chunkDesc := range m.series.chunkDescs {
if i < m.series.persistWatermark {
if _, err = codable.EncodeVarint(w, int64(chunkDesc.firstTime())); err != nil {
return
}
if _, err = codable.EncodeVarint(w, int64(chunkDesc.lastTime())); err != nil {
return
}
} else {
// This is the non-persisted head chunk. Fully marshal it.
if err = w.WriteByte(byte(chunkDesc.c.encoding())); err != nil {
return
}
if err = chunkDesc.c.marshal(w); err != nil {
return
}
}
}
// Series is checkpointed now, so declare it clean.
m.series.dirty = false
}()
if err != nil {
return
}
}
if err = w.Flush(); err != nil {
return
}
if realNumberOfSeries != numberOfSeriesInHeader {
// The number of series has changed in the meantime.
// Rewrite it in the header.
if _, err = f.Seek(int64(numberOfSeriesOffset), os.SEEK_SET); err != nil {
return
}
if err = codable.EncodeUint64(f, realNumberOfSeries); err != nil {
return
}
}
return
}
