fix: export prometheus metrics for cache GC triggers (#9815)

Bonus change to use channel to serialize triggers,
instead of using atomic variables. More efficient
mechanism for synchronization.

Co-authored-by: Nitish Tiwari <nitish@minio.io>
master
Harshavardhana 5 years ago committed by GitHub
parent 2073b79633
commit f9aa239973
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
  1. 87
      cmd/disk-cache-backend.go
  2. 23
      cmd/disk-cache-stats.go
  3. 8
      cmd/disk-cache-utils.go
  4. 30
      cmd/disk-cache.go
  5. 21
      cmd/metrics.go
  6. 23
      docs/metrics/prometheus/README.md

@ -125,12 +125,14 @@ func (m *cacheMeta) ToObjectInfo(bucket, object string) (o ObjectInfo) {
// represents disk cache struct
type diskCache struct {
gcCounter uint64 // ref: https://golang.org/pkg/sync/atomic/#pkg-note-BUG
// is set to 0 if drive is offline
online uint32
online uint32 // ref: https://golang.org/pkg/sync/atomic/#pkg-note-BUG
purgeRunning int32
dir string // caching directory
quotaPct int // max usage in %
triggerGC chan struct{}
dir string // caching directory
stats CacheDiskStats // disk cache stats for prometheus
quotaPct int // max usage in %
pool sync.Pool
after int // minimum accesses before an object is cached.
lowWatermark int
@ -142,12 +144,14 @@ type diskCache struct {
}
// Inits the disk cache dir if it is not initialized already.
func newDiskCache(dir string, quotaPct, after, lowWatermark, highWatermark int) (*diskCache, error) {
func newDiskCache(ctx context.Context, dir string, quotaPct, after, lowWatermark, highWatermark int) (*diskCache, error) {
if err := os.MkdirAll(dir, 0777); err != nil {
return nil, fmt.Errorf("Unable to initialize '%s' dir, %w", dir, err)
}
cache := diskCache{
dir: dir,
triggerGC: make(chan struct{}),
stats: CacheDiskStats{Dir: dir},
quotaPct: quotaPct,
after: after,
lowWatermark: lowWatermark,
@ -161,6 +165,8 @@ func newDiskCache(dir string, quotaPct, after, lowWatermark, highWatermark int)
},
nsMutex: newNSLock(false),
}
go cache.purgeWait(ctx)
cache.diskUsageHigh() // update if cache usage is already high.
cache.NewNSLockFn = func(ctx context.Context, cachePath string) RWLocker {
return cache.nsMutex.NewNSLock(ctx, nil, cachePath, "")
}
@ -181,7 +187,12 @@ func (c *diskCache) diskUsageLow() bool {
return false
}
usedPercent := (di.Total - di.Free) * 100 / di.Total
return int(usedPercent) < gcStopPct
low := int(usedPercent) < gcStopPct
atomic.StoreUint64(&c.stats.UsagePercent, usedPercent)
if low {
atomic.StoreInt32(&c.stats.UsageState, 0)
}
return low
}
// Returns if the disk usage reaches high water mark w.r.t the configured cache quota.
@ -196,7 +207,12 @@ func (c *diskCache) diskUsageHigh() bool {
return false
}
usedPercent := (di.Total - di.Free) * 100 / di.Total
return int(usedPercent) >= gcTriggerPct
high := int(usedPercent) >= gcTriggerPct
atomic.StoreUint64(&c.stats.UsagePercent, usedPercent)
if high {
atomic.StoreInt32(&c.stats.UsageState, 1)
}
return high
}
// Returns if size space can be allocated without exceeding
@ -230,24 +246,36 @@ var (
errDoneForNow = errors.New("done for now")
)
func (c *diskCache) purgeWait(ctx context.Context) {
for {
select {
case <-ctx.Done():
case <-c.triggerGC: // wait here until someone triggers.
c.purge(ctx)
}
}
}
// Purge cache entries that were not accessed.
func (c *diskCache) purge(ctx context.Context) {
if c.diskUsageLow() {
if atomic.LoadInt32(&c.purgeRunning) == 1 || c.diskUsageLow() {
return
}
toFree := c.toClear()
if toFree == 0 {
return
}
atomic.StoreInt32(&c.purgeRunning, 1) // do not run concurrent purge()
defer atomic.StoreInt32(&c.purgeRunning, 0)
// expiry for cleaning up old cache.json files that
// need to be cleaned up.
expiry := UTCNow().Add(-cacheExpiryDays)
// defaulting max hits count to 100
scorer, err := newFileScorer(toFree, time.Now().Unix(), 100)
if err != nil {
logger.LogIf(ctx, err)
return
}
// ignore error we know what value we are passing.
scorer, _ := newFileScorer(toFree, time.Now().Unix(), 100)
// this function returns FileInfo for cached range files and cache data file.
fiStatFn := func(ranges map[string]string, dataFile, pathPrefix string) map[string]os.FileInfo {
@ -326,27 +354,20 @@ func (c *diskCache) purge(ctx context.Context) {
return
}
for _, path := range scorer.fileNames() {
removeAll(path)
slashIdx := strings.LastIndex(path, SlashSeparator)
pathPrefix := path[0:slashIdx]
fname := path[slashIdx+1:]
if fname == cacheDataFile {
removeAll(pathPrefix)
scorer.purgeFunc(func(qfile queuedFile) {
fileName := qfile.name
removeAll(fileName)
slashIdx := strings.LastIndex(fileName, SlashSeparator)
if slashIdx >= 0 {
fileNamePrefix := fileName[0:slashIdx]
fname := fileName[slashIdx+1:]
if fname == cacheDataFile {
removeAll(fileNamePrefix)
}
}
}
}
func (c *diskCache) incGCCounter() {
atomic.AddUint64(&c.gcCounter, 1)
}
func (c *diskCache) resetGCCounter() {
atomic.StoreUint64(&c.gcCounter, 0)
}
})
func (c *diskCache) gcCount() uint64 {
return atomic.LoadUint64(&c.gcCounter)
scorer.reset()
}
// sets cache drive status
@ -630,7 +651,7 @@ func newCacheEncryptMetadata(bucket, object string, metadata map[string]string)
// Caches the object to disk
func (c *diskCache) Put(ctx context.Context, bucket, object string, data io.Reader, size int64, rs *HTTPRangeSpec, opts ObjectOptions, incHitsOnly bool) error {
if c.diskUsageHigh() {
c.incGCCounter()
c.triggerGC <- struct{}{}
io.Copy(ioutil.Discard, data)
return errDiskFull
}

@ -1,5 +1,5 @@
/*
* MinIO Cloud Storage, (C) 2019 MinIO, Inc.
* MinIO Cloud Storage, (C) 2019, 2020 MinIO, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
@ -16,14 +16,27 @@
package cmd
import "sync/atomic"
import (
"sync/atomic"
)
// CacheDiskStats represents cache disk statistics
// such as current disk usage and available.
type CacheDiskStats struct {
// indicates if usage is high or low, if high value is '1', if low its '0'
UsageState int32
// indicates the current usage percentage of this cache disk
UsagePercent uint64
Dir string
}
// CacheStats - represents bytes served from cache,
// cache hits and cache misses.
type CacheStats struct {
BytesServed uint64
Hits uint64
Misses uint64
BytesServed uint64
Hits uint64
Misses uint64
GetDiskStats func() []CacheDiskStats
}
// Increase total bytes served from cache

@ -439,6 +439,14 @@ func (f *fileScorer) fileObjInfos() []ObjectInfo {
return res
}
func (f *fileScorer) purgeFunc(p func(qfile queuedFile)) {
e := f.queue.Front()
for e != nil {
p(e.Value.(queuedFile))
e = e.Next()
}
}
// fileNames returns all queued file names.
func (f *fileScorer) fileNames() []string {
res := make([]string, 0, f.queue.Len())

@ -24,6 +24,7 @@ import (
"net/http"
"strings"
"sync"
"sync/atomic"
"time"
"github.com/minio/minio/cmd/config/cache"
@ -282,10 +283,11 @@ func (c *cacheObjects) GetObjectNInfo(ctx context.Context, bucket, object string
// Reaching here implies cache miss
c.cacheStats.incMiss()
// Since we got here, we are serving the request from backend,
// and also adding the object to the cache.
if dcache.diskUsageHigh() {
dcache.incGCCounter()
dcache.triggerGC <- struct{}{} // this is non-blocking
}
bkReader, bkErr := c.GetObjectNInfoFn(ctx, bucket, object, rs, h, lockType, opts)
@ -544,7 +546,7 @@ func newCache(config cache.Config) ([]*diskCache, bool, error) {
if quota == 0 {
quota = config.Quota
}
cache, err := newDiskCache(dir, quota, config.After, config.WatermarkLow, config.WatermarkHigh)
cache, err := newDiskCache(ctx, dir, quota, config.After, config.WatermarkLow, config.WatermarkHigh)
if err != nil {
return nil, false, err
}
@ -677,7 +679,17 @@ func newServerCacheObjects(ctx context.Context, config cache.Config) (CacheObjec
return newObjectLayerFn().CopyObject(ctx, srcBucket, srcObject, destBucket, destObject, srcInfo, srcOpts, dstOpts)
},
}
c.cacheStats.GetDiskStats = func() []CacheDiskStats {
cacheDiskStats := make([]CacheDiskStats, len(c.cache))
for i := range c.cache {
cacheDiskStats[i] = CacheDiskStats{
Dir: c.cache[i].stats.Dir,
}
atomic.StoreInt32(&cacheDiskStats[i].UsageState, atomic.LoadInt32(&c.cache[i].stats.UsageState))
atomic.StoreUint64(&cacheDiskStats[i].UsagePercent, atomic.LoadUint64(&c.cache[i].stats.UsagePercent))
}
return cacheDiskStats
}
if migrateSw {
go c.migrateCacheFromV1toV2(ctx)
}
@ -697,19 +709,9 @@ func (c *cacheObjects) gc(ctx context.Context) {
if c.migrating {
continue
}
var wg sync.WaitGroup
for _, dcache := range c.cache {
if dcache.gcCount() == 0 {
continue
}
wg.Add(1)
go func(d *diskCache) {
defer wg.Done()
d.resetGCCounter()
d.purge(ctx)
}(dcache)
dcache.triggerGC <- struct{}{}
}
wg.Wait()
}
}
}

@ -260,6 +260,27 @@ func cacheMetricsPrometheus(ch chan<- prometheus.Metric) {
prometheus.CounterValue,
float64(cacheObjLayer.CacheStats().getBytesServed()),
)
for _, cdStats := range cacheObjLayer.CacheStats().GetDiskStats() {
// Cache disk usage percentage
ch <- prometheus.MustNewConstMetric(
prometheus.NewDesc(
prometheus.BuildFQName("cache", "usage", "percent"),
"Total percentage cache usage",
[]string{"disk"}, nil),
prometheus.GaugeValue,
float64(cdStats.UsagePercent),
cdStats.Dir,
)
ch <- prometheus.MustNewConstMetric(
prometheus.NewDesc(
prometheus.BuildFQName("cache", "usage", "high"),
"Indicates cache usage is high or low, relative to current cache 'quota' settings",
[]string{"disk"}, nil),
prometheus.GaugeValue,
float64(cdStats.UsageState),
cdStats.Dir,
)
}
}
// collects http metrics for MinIO server in Prometheus specific format

@ -110,7 +110,7 @@ These are the new set of metrics which will be in effect after `RELEASE.2019-10-
- Metrics that records the http statistics and latencies are labeled to their respective APIs (putobject,getobject etc).
- Disk usage metrics are distributed and labeled to the respective disk paths.
For more details, please check the `Migration guide for the new set of metrics`.
For more details, please check the `Migration guide for the new set of metrics`.
The list of metrics and its definition are as follows. (NOTE: instance here is one MinIO node)
@ -178,14 +178,23 @@ All metrics are labeled by `bucket`, each metric is displayed per bucket. `bucke
MinIO Gateway instances enabled with Disk-Caching expose caching related metrics.
| name | description |
|:---------------------|:----------------------------------------|
| `cache_data_served` | Total number of bytes served from cache |
| `cache_hits_total` | Total number of cache hits |
| `cache_misses_total` | Total number of cache misses |
#### Global cache metrics
| name | description |
|:---------------------|:--------------------------------------------------|
| `cache_hits_total` | Total number of cache hits |
| `cache_misses_total` | Total number of cache misses |
| `cache_data_served` | Total number of bytes served from cache |
### Gateway & Cache specific metrics
#### Per disk cache metrics
| `cache_usage_percent` | Total percentage cache usage |
| `cache_usage_state` | Indicates cache usage is high or low, relative to current cache 'quota' settings |
`cache_usage_state` holds only two states
- '1' indicates high disk usage
- '0' indicates low disk usage
### Gateway specific metrics
MinIO Gateway instance exposes metrics related to Gateway communication with the cloud backend (S3, Azure & GCS Gateway).
`<gateway_type>` changes based on the gateway in use can be 's3', 'gcs' or 'azure'. Other metrics are labeled with `method` that identifies HTTP GET, HEAD, PUT and POST requests to the backend.

Loading…
Cancel
Save