cc-backend/metricdata/metricdata.go

245 lines
6.8 KiB
Go
Raw Normal View History

package metricdata
import (
"context"
"encoding/json"
2021-11-26 10:32:36 +01:00
"fmt"
"time"
2022-01-27 09:40:59 +01:00
"github.com/ClusterCockpit/cc-backend/config"
"github.com/ClusterCockpit/cc-backend/log"
2022-01-27 09:40:59 +01:00
"github.com/ClusterCockpit/cc-backend/schema"
"github.com/iamlouk/lrucache"
)
2021-12-08 10:14:45 +01:00
type MetricDataRepository interface {
2021-12-09 16:25:48 +01:00
// Initialize this MetricDataRepository. One instance of
// this interface will only ever be responsible for one cluster.
Init(rawConfig json.RawMessage) error
2021-12-09 16:25:48 +01:00
// Return the JobData for the given job, only with the requested metrics.
LoadData(job *schema.Job, metrics []string, scopes []schema.MetricScope, ctx context.Context) (schema.JobData, error)
2021-12-09 16:25:48 +01:00
2022-01-12 13:03:01 +01:00
// Return a map of metrics to a map of nodes to the metric statistics of the job. node scope assumed for now.
2021-12-17 15:49:22 +01:00
LoadStats(job *schema.Job, metrics []string, ctx context.Context) (map[string]map[string]schema.MetricStatistics, error)
2021-12-09 16:25:48 +01:00
// Return a map of hosts to a map of metrics at the requested scopes for that node.
LoadNodeData(cluster string, metrics, nodes []string, scopes []schema.MetricScope, from, to time.Time, ctx context.Context) (map[string]map[string][]*schema.JobMetric, error)
2021-12-08 10:14:45 +01:00
}
var metricDataRepos map[string]MetricDataRepository = map[string]MetricDataRepository{}
2021-11-26 10:32:36 +01:00
2021-12-08 10:14:45 +01:00
var JobArchivePath string
var useArchive bool
func Init(jobArchivePath string, disableArchive bool) error {
useArchive = !disableArchive
2021-12-08 10:14:45 +01:00
JobArchivePath = jobArchivePath
for _, cluster := range config.Clusters {
if cluster.MetricDataRepository != nil {
var kind struct {
Kind string `json:"kind"`
}
if err := json.Unmarshal(cluster.MetricDataRepository, &kind); err != nil {
return err
}
2022-01-24 10:06:25 +01:00
var mdr MetricDataRepository
switch kind.Kind {
2021-12-08 10:14:45 +01:00
case "cc-metric-store":
2022-01-24 10:06:25 +01:00
mdr = &CCMetricStore{}
case "influxdb":
mdr = &InfluxDBv2DataRepository{}
2022-01-24 10:06:25 +01:00
case "test":
mdr = &TestMetricDataRepository{}
2021-12-08 10:14:45 +01:00
default:
return fmt.Errorf("unkown metric data repository '%s' for cluster '%s'", kind.Kind, cluster.Name)
2021-12-08 10:14:45 +01:00
}
2022-01-24 10:06:25 +01:00
if err := mdr.Init(cluster.MetricDataRepository); err != nil {
2022-01-24 10:06:25 +01:00
return err
}
metricDataRepos[cluster.Name] = mdr
2021-12-08 10:14:45 +01:00
}
2021-11-26 10:32:36 +01:00
}
2021-12-08 10:14:45 +01:00
return nil
2021-11-26 10:32:36 +01:00
}
2022-01-20 10:08:50 +01:00
var cache *lrucache.Cache = lrucache.New(512 * 1024 * 1024)
// Fetches the metric data for a job.
func LoadData(job *schema.Job, metrics []string, scopes []schema.MetricScope, ctx context.Context) (schema.JobData, error) {
2022-03-21 13:30:19 +01:00
data := cache.Get(cacheKey(job, metrics, scopes), func() (_ interface{}, ttl time.Duration, size int) {
2022-01-20 10:08:50 +01:00
var jd schema.JobData
var err error
2022-02-16 09:06:23 +01:00
if job.State == schema.JobStateRunning ||
job.MonitoringStatus == schema.MonitoringStatusRunningOrArchiving ||
!useArchive {
2022-01-20 10:08:50 +01:00
repo, ok := metricDataRepos[job.Cluster]
if !ok {
return fmt.Errorf("no metric data repository configured for '%s'", job.Cluster), 0, 0
}
2022-01-20 10:08:50 +01:00
if scopes == nil {
scopes = append(scopes, schema.MetricScopeNode)
}
2021-12-08 10:14:45 +01:00
2022-01-20 10:08:50 +01:00
if metrics == nil {
cluster := config.GetCluster(job.Cluster)
2022-01-20 10:08:50 +01:00
for _, mc := range cluster.MetricConfig {
metrics = append(metrics, mc.Name)
}
}
2022-01-20 10:08:50 +01:00
jd, err = repo.LoadData(job, metrics, scopes, ctx)
if err != nil {
if len(jd) != 0 {
log.Errorf("partial error: %s", err.Error())
} else {
return err, 0, 0
}
2022-01-20 10:08:50 +01:00
}
2022-03-21 13:30:19 +01:00
size = jd.Size()
2022-01-20 10:08:50 +01:00
} else {
jd, err = loadFromArchive(job)
if err != nil {
return err, 0, 0
}
2022-03-21 13:30:19 +01:00
// Avoid sending unrequested data to the client:
2022-01-20 10:08:50 +01:00
if metrics != nil {
res := schema.JobData{}
for _, metric := range metrics {
2022-03-21 13:30:19 +01:00
if perscope, ok := jd[metric]; ok {
if len(scopes) > 1 {
subset := make(map[schema.MetricScope]*schema.JobMetric)
for _, scope := range scopes {
if jm, ok := perscope[scope]; ok {
subset[scope] = jm
}
}
perscope = subset
}
res[metric] = perscope
2022-01-20 10:08:50 +01:00
}
}
jd = res
}
2022-03-21 13:30:19 +01:00
size = 1 // loadFromArchive() caches in the same cache.
}
2022-03-21 13:30:19 +01:00
ttl = 5 * time.Hour
2022-01-20 10:08:50 +01:00
if job.State == schema.JobStateRunning {
ttl = 2 * time.Minute
}
2022-01-20 10:08:50 +01:00
prepareJobData(job, jd, scopes)
2022-03-21 13:30:19 +01:00
return jd, ttl, size
2022-01-20 10:08:50 +01:00
})
2021-11-26 10:32:36 +01:00
2022-01-20 10:08:50 +01:00
if err, ok := data.(error); ok {
return nil, err
}
2022-01-20 10:08:50 +01:00
return data.(schema.JobData), nil
}
// Used for the jobsFootprint GraphQL-Query. TODO: Rename/Generalize.
2021-12-17 15:49:22 +01:00
func LoadAverages(job *schema.Job, metrics []string, data [][]schema.Float, ctx context.Context) error {
if job.State != schema.JobStateRunning && useArchive {
return loadAveragesFromArchive(job, metrics, data)
}
repo, ok := metricDataRepos[job.Cluster]
if !ok {
return fmt.Errorf("no metric data repository configured for '%s'", job.Cluster)
}
stats, err := repo.LoadStats(job, metrics, ctx)
if err != nil {
return err
}
for i, m := range metrics {
nodes, ok := stats[m]
if !ok {
data[i] = append(data[i], schema.NaN)
continue
}
sum := 0.0
for _, node := range nodes {
sum += node.Avg
}
data[i] = append(data[i], schema.Float(sum))
}
return nil
}
// Used for the node/system view. Returns a map of nodes to a map of metrics.
func LoadNodeData(cluster string, metrics, nodes []string, scopes []schema.MetricScope, from, to time.Time, ctx context.Context) (map[string]map[string][]*schema.JobMetric, error) {
repo, ok := metricDataRepos[cluster]
if !ok {
return nil, fmt.Errorf("no metric data repository configured for '%s'", cluster)
}
if metrics == nil {
for _, m := range config.GetCluster(cluster).MetricConfig {
metrics = append(metrics, m.Name)
}
}
data, err := repo.LoadNodeData(cluster, metrics, nodes, scopes, from, to, ctx)
if err != nil {
if len(data) != 0 {
log.Errorf("partial error: %s", err.Error())
} else {
return nil, err
}
}
if data == nil {
return nil, fmt.Errorf("the metric data repository for '%s' does not support this query", cluster)
}
return data, nil
}
func cacheKey(job *schema.Job, metrics []string, scopes []schema.MetricScope) string {
// Duration and StartTime do not need to be in the cache key as StartTime is less unique than
// job.ID and the TTL of the cache entry makes sure it does not stay there forever.
2022-01-20 10:08:50 +01:00
return fmt.Sprintf("%d(%s):[%v],[%v]",
job.ID, job.State, metrics, scopes)
}
// For /monitoring/job/<job> and some other places, flops_any and mem_bw need to be available at the scope 'node'.
// If a job has a lot of nodes, statisticsSeries should be available so that a min/mean/max Graph can be used instead of
// a lot of single lines.
func prepareJobData(job *schema.Job, jobData schema.JobData, scopes []schema.MetricScope) {
const maxSeriesSize int = 15
for _, scopes := range jobData {
for _, jm := range scopes {
if jm.StatisticsSeries != nil || len(jm.Series) <= maxSeriesSize {
continue
}
jm.AddStatisticsSeries()
}
}
nodeScopeRequested := false
for _, scope := range scopes {
if scope == schema.MetricScopeNode {
nodeScopeRequested = true
}
}
if nodeScopeRequested {
jobData.AddNodeScope("flops_any")
jobData.AddNodeScope("mem_bw")
}
}