VictoriaMetrics/app/vmalert/alerting.go

518 lines
15 KiB
Go
Raw Normal View History

package main
import (
"context"
"fmt"
"hash/fnv"
"sort"
"strconv"
"sync"
"time"
"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/config"
"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/datasource"
"github.com/VictoriaMetrics/VictoriaMetrics/app/vmalert/notifier"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/prompbmarshal"
app/vmalert: extend metrics set exported by `vmalert` #573 (#654) * app/vmalert: extend metrics set exported by `vmalert` #573 New metrics were added to improve observability: + vmalert_alerts_pending{alertname, group} - number of pending alerts per group per alert; + vmalert_alerts_acitve{alertname, group} - number of active alerts per group per alert; + vmalert_alerts_error{alertname, group} - is 1 if alertname ended up with error during prev execution, is 0 if no errors happened; + vmalert_recording_rules_error{recording, group} - is 1 if recording rule ended up with error during prev execution, is 0 if no errors happened; * vmalert_iteration_total{group, file} - now contains group and file name labels. This should improve control over specific groups; * vmalert_iteration_duration_seconds{group, file} - now contains group and file name labels. This should improve control over specific groups; Some collisions for alerts and recording rules are possible, because neither group name nor alert/recording rule name are unique for compatibility reasons. Commit contains list of TODOs for Unregistering metrics since groups and rules are ephemeral and could be removed without application restart. In order to unlock Unregistering feature corresponding PR was filed - https://github.com/VictoriaMetrics/metrics/pull/13 * app/vmalert: extend metrics set exported by `vmalert` #573 The changes are following: * add an ID label to rules metrics, since `name` collisions within one group is a common case - see the k8s example alerts; * supports metrics unregistering on rule updates. Consider the case when one rule was added or removed from the group, or the whole group was added or removed. The change depends on https://github.com/VictoriaMetrics/metrics/pull/16 where race condition for Unregister method was fixed.
2020-08-09 08:41:29 +02:00
"github.com/VictoriaMetrics/metrics"
)
// AlertingRule is basic alert entity
type AlertingRule struct {
Type datasource.Type
RuleID uint64
Name string
Expr string
For time.Duration
Labels map[string]string
Annotations map[string]string
GroupID uint64
GroupName string
EvalInterval time.Duration
q datasource.Querier
// guard status fields
mu sync.RWMutex
// stores list of active alerts
alerts map[uint64]*notifier.Alert
// stores last moment of time Exec was called
lastExecTime time.Time
// stores last error that happened in Exec func
// resets on every successful Exec
// may be used as Health state
lastExecError error
app/vmalert: extend metrics set exported by `vmalert` #573 (#654) * app/vmalert: extend metrics set exported by `vmalert` #573 New metrics were added to improve observability: + vmalert_alerts_pending{alertname, group} - number of pending alerts per group per alert; + vmalert_alerts_acitve{alertname, group} - number of active alerts per group per alert; + vmalert_alerts_error{alertname, group} - is 1 if alertname ended up with error during prev execution, is 0 if no errors happened; + vmalert_recording_rules_error{recording, group} - is 1 if recording rule ended up with error during prev execution, is 0 if no errors happened; * vmalert_iteration_total{group, file} - now contains group and file name labels. This should improve control over specific groups; * vmalert_iteration_duration_seconds{group, file} - now contains group and file name labels. This should improve control over specific groups; Some collisions for alerts and recording rules are possible, because neither group name nor alert/recording rule name are unique for compatibility reasons. Commit contains list of TODOs for Unregistering metrics since groups and rules are ephemeral and could be removed without application restart. In order to unlock Unregistering feature corresponding PR was filed - https://github.com/VictoriaMetrics/metrics/pull/13 * app/vmalert: extend metrics set exported by `vmalert` #573 The changes are following: * add an ID label to rules metrics, since `name` collisions within one group is a common case - see the k8s example alerts; * supports metrics unregistering on rule updates. Consider the case when one rule was added or removed from the group, or the whole group was added or removed. The change depends on https://github.com/VictoriaMetrics/metrics/pull/16 where race condition for Unregister method was fixed.
2020-08-09 08:41:29 +02:00
metrics *alertingRuleMetrics
}
type alertingRuleMetrics struct {
errors *gauge
pending *gauge
active *gauge
}
func newAlertingRule(qb datasource.QuerierBuilder, group *Group, cfg config.Rule) *AlertingRule {
app/vmalert: extend metrics set exported by `vmalert` #573 (#654) * app/vmalert: extend metrics set exported by `vmalert` #573 New metrics were added to improve observability: + vmalert_alerts_pending{alertname, group} - number of pending alerts per group per alert; + vmalert_alerts_acitve{alertname, group} - number of active alerts per group per alert; + vmalert_alerts_error{alertname, group} - is 1 if alertname ended up with error during prev execution, is 0 if no errors happened; + vmalert_recording_rules_error{recording, group} - is 1 if recording rule ended up with error during prev execution, is 0 if no errors happened; * vmalert_iteration_total{group, file} - now contains group and file name labels. This should improve control over specific groups; * vmalert_iteration_duration_seconds{group, file} - now contains group and file name labels. This should improve control over specific groups; Some collisions for alerts and recording rules are possible, because neither group name nor alert/recording rule name are unique for compatibility reasons. Commit contains list of TODOs for Unregistering metrics since groups and rules are ephemeral and could be removed without application restart. In order to unlock Unregistering feature corresponding PR was filed - https://github.com/VictoriaMetrics/metrics/pull/13 * app/vmalert: extend metrics set exported by `vmalert` #573 The changes are following: * add an ID label to rules metrics, since `name` collisions within one group is a common case - see the k8s example alerts; * supports metrics unregistering on rule updates. Consider the case when one rule was added or removed from the group, or the whole group was added or removed. The change depends on https://github.com/VictoriaMetrics/metrics/pull/16 where race condition for Unregister method was fixed.
2020-08-09 08:41:29 +02:00
ar := &AlertingRule{
Type: cfg.Type,
RuleID: cfg.ID,
Name: cfg.Alert,
Expr: cfg.Expr,
For: cfg.For.Duration(),
Labels: cfg.Labels,
Annotations: cfg.Annotations,
GroupID: group.ID(),
GroupName: group.Name,
EvalInterval: group.Interval,
q: qb.BuildWithParams(datasource.QuerierParams{
DataSourceType: &cfg.Type,
EvaluationInterval: group.Interval,
ExtraLabels: group.ExtraFilterLabels,
}),
alerts: make(map[uint64]*notifier.Alert),
metrics: &alertingRuleMetrics{},
}
app/vmalert: extend metrics set exported by `vmalert` #573 (#654) * app/vmalert: extend metrics set exported by `vmalert` #573 New metrics were added to improve observability: + vmalert_alerts_pending{alertname, group} - number of pending alerts per group per alert; + vmalert_alerts_acitve{alertname, group} - number of active alerts per group per alert; + vmalert_alerts_error{alertname, group} - is 1 if alertname ended up with error during prev execution, is 0 if no errors happened; + vmalert_recording_rules_error{recording, group} - is 1 if recording rule ended up with error during prev execution, is 0 if no errors happened; * vmalert_iteration_total{group, file} - now contains group and file name labels. This should improve control over specific groups; * vmalert_iteration_duration_seconds{group, file} - now contains group and file name labels. This should improve control over specific groups; Some collisions for alerts and recording rules are possible, because neither group name nor alert/recording rule name are unique for compatibility reasons. Commit contains list of TODOs for Unregistering metrics since groups and rules are ephemeral and could be removed without application restart. In order to unlock Unregistering feature corresponding PR was filed - https://github.com/VictoriaMetrics/metrics/pull/13 * app/vmalert: extend metrics set exported by `vmalert` #573 The changes are following: * add an ID label to rules metrics, since `name` collisions within one group is a common case - see the k8s example alerts; * supports metrics unregistering on rule updates. Consider the case when one rule was added or removed from the group, or the whole group was added or removed. The change depends on https://github.com/VictoriaMetrics/metrics/pull/16 where race condition for Unregister method was fixed.
2020-08-09 08:41:29 +02:00
labels := fmt.Sprintf(`alertname=%q, group=%q, id="%d"`, ar.Name, group.Name, ar.ID())
ar.metrics.pending = getOrCreateGauge(fmt.Sprintf(`vmalert_alerts_pending{%s}`, labels),
func() float64 {
ar.mu.Lock()
defer ar.mu.Unlock()
var num int
for _, a := range ar.alerts {
if a.State == notifier.StatePending {
num++
}
}
return float64(num)
})
ar.metrics.active = getOrCreateGauge(fmt.Sprintf(`vmalert_alerts_firing{%s}`, labels),
func() float64 {
ar.mu.Lock()
defer ar.mu.Unlock()
var num int
for _, a := range ar.alerts {
if a.State == notifier.StateFiring {
num++
}
}
return float64(num)
})
ar.metrics.errors = getOrCreateGauge(fmt.Sprintf(`vmalert_alerts_error{%s}`, labels),
func() float64 {
ar.mu.Lock()
defer ar.mu.Unlock()
if ar.lastExecError == nil {
return 0
}
return 1
})
return ar
}
// Close unregisters rule metrics
func (ar *AlertingRule) Close() {
metrics.UnregisterMetric(ar.metrics.active.name)
metrics.UnregisterMetric(ar.metrics.pending.name)
metrics.UnregisterMetric(ar.metrics.errors.name)
}
// String implements Stringer interface
func (ar *AlertingRule) String() string {
return ar.Name
}
// ID returns unique Rule ID
// within the parent Group.
func (ar *AlertingRule) ID() uint64 {
return ar.RuleID
}
// ExecRange executes alerting rule on the given time range similarly to Exec.
// It doesn't update internal states of the Rule and meant to be used just
// to get time series for backfilling.
// It returns ALERT and ALERT_FOR_STATE time series as result.
func (ar *AlertingRule) ExecRange(ctx context.Context, start, end time.Time) ([]prompbmarshal.TimeSeries, error) {
series, err := ar.q.QueryRange(ctx, ar.Expr, start, end)
if err != nil {
return nil, err
}
var result []prompbmarshal.TimeSeries
qFn := func(query string) ([]datasource.Metric, error) {
return nil, fmt.Errorf("`query` template isn't supported in replay mode")
}
for _, s := range series {
// extra labels could contain templates, so we expand them first
labels, err := expandLabels(s, qFn, ar)
if err != nil {
return nil, fmt.Errorf("failed to expand labels: %s", err)
}
for k, v := range labels {
// apply extra labels to datasource
// so the hash key will be consistent on restore
s.SetLabel(k, v)
}
a, err := ar.newAlert(s, time.Time{}, qFn) // initial alert
if err != nil {
return nil, fmt.Errorf("failed to create alert: %s", err)
}
if ar.For == 0 { // if alert is instant
a.State = notifier.StateFiring
for i := range s.Values {
result = append(result, ar.alertToTimeSeries(a, s.Timestamps[i])...)
}
continue
}
// if alert with For > 0
prevT := time.Time{}
//activeAt := time.Time{}
for i := range s.Values {
at := time.Unix(s.Timestamps[i], 0)
if at.Sub(prevT) > ar.EvalInterval {
// reset to Pending if there are gaps > EvalInterval between DPs
a.State = notifier.StatePending
//activeAt = at
a.Start = at
} else if at.Sub(a.Start) >= ar.For {
a.State = notifier.StateFiring
}
prevT = at
result = append(result, ar.alertToTimeSeries(a, s.Timestamps[i])...)
}
}
return result, nil
}
// Exec executes AlertingRule expression via the given Querier.
// Based on the Querier results AlertingRule maintains notifier.Alerts
func (ar *AlertingRule) Exec(ctx context.Context) ([]prompbmarshal.TimeSeries, error) {
qMetrics, err := ar.q.Query(ctx, ar.Expr)
ar.mu.Lock()
defer ar.mu.Unlock()
ar.lastExecError = err
ar.lastExecTime = time.Now()
if err != nil {
return nil, fmt.Errorf("failed to execute query %q: %w", ar.Expr, err)
}
for h, a := range ar.alerts {
// cleanup inactive alerts from previous Exec
if a.State == notifier.StateInactive {
delete(ar.alerts, h)
}
}
qFn := func(query string) ([]datasource.Metric, error) { return ar.q.Query(ctx, query) }
updated := make(map[uint64]struct{})
// update list of active alerts
for _, m := range qMetrics {
// extra labels could contain templates, so we expand them first
labels, err := expandLabels(m, qFn, ar)
if err != nil {
return nil, fmt.Errorf("failed to expand labels: %s", err)
}
for k, v := range labels {
// apply extra labels to datasource
// so the hash key will be consistent on restore
m.SetLabel(k, v)
}
h := hash(m)
if _, ok := updated[h]; ok {
// duplicate may be caused by extra labels
// conflicting with the metric labels
return nil, fmt.Errorf("labels %v: %w", m.Labels, errDuplicate)
}
updated[h] = struct{}{}
if a, ok := ar.alerts[h]; ok {
if a.Value != m.Values[0] {
// update Value field with latest value
a.Value = m.Values[0]
// and re-exec template since Value can be used
// in annotations
a.Annotations, err = a.ExecTemplate(qFn, ar.Annotations)
if err != nil {
return nil, err
}
}
continue
}
a, err := ar.newAlert(m, ar.lastExecTime, qFn)
if err != nil {
ar.lastExecError = err
return nil, fmt.Errorf("failed to create alert: %w", err)
}
a.ID = h
a.State = notifier.StatePending
ar.alerts[h] = a
}
for h, a := range ar.alerts {
// if alert wasn't updated in this iteration
// means it is resolved already
if _, ok := updated[h]; !ok {
if a.State == notifier.StatePending {
// alert was in Pending state - it is not
// active anymore
delete(ar.alerts, h)
continue
}
a.State = notifier.StateInactive
continue
}
if a.State == notifier.StatePending && time.Since(a.Start) >= ar.For {
a.State = notifier.StateFiring
alertsFired.Inc()
}
}
return ar.toTimeSeries(ar.lastExecTime.Unix()), nil
}
func expandLabels(m datasource.Metric, q notifier.QueryFn, ar *AlertingRule) (map[string]string, error) {
metricLabels := make(map[string]string)
for _, l := range m.Labels {
metricLabels[l.Name] = l.Value
}
tpl := notifier.AlertTplData{
Labels: metricLabels,
Value: m.Values[0],
Expr: ar.Expr,
}
return notifier.ExecTemplate(q, ar.Labels, tpl)
}
func (ar *AlertingRule) toTimeSeries(timestamp int64) []prompbmarshal.TimeSeries {
var tss []prompbmarshal.TimeSeries
for _, a := range ar.alerts {
if a.State == notifier.StateInactive {
continue
}
ts := ar.alertToTimeSeries(a, timestamp)
tss = append(tss, ts...)
}
return tss
}
// UpdateWith copies all significant fields.
// alerts state isn't copied since
// it should be updated in next 2 Execs
func (ar *AlertingRule) UpdateWith(r Rule) error {
nr, ok := r.(*AlertingRule)
if !ok {
return fmt.Errorf("BUG: attempt to update alerting rule with wrong type %#v", r)
}
ar.Expr = nr.Expr
ar.For = nr.For
ar.Labels = nr.Labels
ar.Annotations = nr.Annotations
ar.EvalInterval = nr.EvalInterval
ar.q = nr.q
return nil
}
// TODO: consider hashing algorithm in VM
func hash(m datasource.Metric) uint64 {
hash := fnv.New64a()
labels := m.Labels
sort.Slice(labels, func(i, j int) bool {
return labels[i].Name < labels[j].Name
})
for _, l := range labels {
// drop __name__ to be consistent with Prometheus alerting
if l.Name == "__name__" {
continue
}
hash.Write([]byte(l.Name))
hash.Write([]byte(l.Value))
hash.Write([]byte("\xff"))
}
return hash.Sum64()
}
func (ar *AlertingRule) newAlert(m datasource.Metric, start time.Time, qFn notifier.QueryFn) (*notifier.Alert, error) {
a := &notifier.Alert{
GroupID: ar.GroupID,
Name: ar.Name,
Labels: map[string]string{},
Value: m.Values[0],
Start: start,
Expr: ar.Expr,
}
// label defined here to make override possible by
// time series labels.
if ar.GroupName != "" {
a.Labels[alertGroupNameLabel] = ar.GroupName
}
for _, l := range m.Labels {
// drop __name__ to be consistent with Prometheus alerting
if l.Name == "__name__" {
continue
}
a.Labels[l.Name] = l.Value
}
var err error
a.Annotations, err = a.ExecTemplate(qFn, ar.Annotations)
return a, err
}
// AlertAPI generates APIAlert object from alert by its id(hash)
func (ar *AlertingRule) AlertAPI(id uint64) *APIAlert {
ar.mu.RLock()
defer ar.mu.RUnlock()
a, ok := ar.alerts[id]
if !ok {
return nil
}
return ar.newAlertAPI(*a)
}
// RuleAPI returns Rule representation in form
// of APIAlertingRule
func (ar *AlertingRule) RuleAPI() APIAlertingRule {
var lastErr string
if ar.lastExecError != nil {
lastErr = ar.lastExecError.Error()
}
return APIAlertingRule{
// encode as strings to avoid rounding
ID: fmt.Sprintf("%d", ar.ID()),
GroupID: fmt.Sprintf("%d", ar.GroupID),
Type: ar.Type.String(),
Name: ar.Name,
Expression: ar.Expr,
For: ar.For.String(),
LastError: lastErr,
LastExec: ar.lastExecTime,
Labels: ar.Labels,
Annotations: ar.Annotations,
}
}
// AlertsAPI generates list of APIAlert objects from existing alerts
func (ar *AlertingRule) AlertsAPI() []*APIAlert {
var alerts []*APIAlert
ar.mu.RLock()
for _, a := range ar.alerts {
alerts = append(alerts, ar.newAlertAPI(*a))
}
ar.mu.RUnlock()
return alerts
}
func (ar *AlertingRule) newAlertAPI(a notifier.Alert) *APIAlert {
return &APIAlert{
// encode as strings to avoid rounding
ID: fmt.Sprintf("%d", a.ID),
GroupID: fmt.Sprintf("%d", a.GroupID),
Name: a.Name,
Expression: ar.Expr,
Labels: a.Labels,
Annotations: a.Annotations,
State: a.State.String(),
ActiveAt: a.Start,
Value: strconv.FormatFloat(a.Value, 'e', -1, 64),
}
}
const (
// alertMetricName is the metric name for synthetic alert timeseries.
alertMetricName = "ALERTS"
// alertForStateMetricName is the metric name for 'for' state of alert.
alertForStateMetricName = "ALERTS_FOR_STATE"
// alertNameLabel is the label name indicating the name of an alert.
alertNameLabel = "alertname"
// alertStateLabel is the label name indicating the state of an alert.
alertStateLabel = "alertstate"
// alertGroupNameLabel defines the label name attached for generated time series.
alertGroupNameLabel = "alertgroup"
)
// alertToTimeSeries converts the given alert with the given timestamp to timeseries
func (ar *AlertingRule) alertToTimeSeries(a *notifier.Alert, timestamp int64) []prompbmarshal.TimeSeries {
var tss []prompbmarshal.TimeSeries
tss = append(tss, alertToTimeSeries(ar.Name, a, timestamp))
if ar.For > 0 {
tss = append(tss, alertForToTimeSeries(ar.Name, a, timestamp))
}
return tss
}
func alertToTimeSeries(name string, a *notifier.Alert, timestamp int64) prompbmarshal.TimeSeries {
labels := make(map[string]string)
for k, v := range a.Labels {
labels[k] = v
}
labels["__name__"] = alertMetricName
labels[alertNameLabel] = name
labels[alertStateLabel] = a.State.String()
return newTimeSeries([]float64{1}, []int64{timestamp}, labels)
}
// alertForToTimeSeries returns a timeseries that represents
// state of active alerts, where value is time when alert become active
func alertForToTimeSeries(name string, a *notifier.Alert, timestamp int64) prompbmarshal.TimeSeries {
labels := make(map[string]string)
for k, v := range a.Labels {
labels[k] = v
}
labels["__name__"] = alertForStateMetricName
labels[alertNameLabel] = name
return newTimeSeries([]float64{float64(a.Start.Unix())}, []int64{timestamp}, labels)
}
// Restore restores the state of active alerts basing on previously written timeseries.
// Restore restores only Start field. Field State will be always Pending and supposed
// to be updated on next Exec, as well as Value field.
// Only rules with For > 0 will be restored.
func (ar *AlertingRule) Restore(ctx context.Context, q datasource.Querier, lookback time.Duration, labels map[string]string) error {
if q == nil {
return fmt.Errorf("querier is nil")
}
qFn := func(query string) ([]datasource.Metric, error) { return ar.q.Query(ctx, query) }
// account for external labels in filter
var labelsFilter string
for k, v := range labels {
labelsFilter += fmt.Sprintf(",%s=%q", k, v)
}
// Get the last data point in range via MetricsQL `last_over_time`.
// We don't use plain PromQL since Prometheus doesn't support
// remote write protocol which is used for state persistence in vmalert.
expr := fmt.Sprintf("last_over_time(%s{alertname=%q%s}[%ds])",
alertForStateMetricName, ar.Name, labelsFilter, int(lookback.Seconds()))
qMetrics, err := q.Query(ctx, expr)
if err != nil {
return err
}
for _, m := range qMetrics {
labels := m.Labels
m.Labels = make([]datasource.Label, 0)
// drop all extra labels, so hash key will
// be identical to time series received in Exec
for _, l := range labels {
if l.Name == alertNameLabel || l.Name == alertGroupNameLabel {
continue
}
m.Labels = append(m.Labels, l)
}
a, err := ar.newAlert(m, time.Unix(int64(m.Values[0]), 0), qFn)
if err != nil {
return fmt.Errorf("failed to create alert: %w", err)
}
a.ID = hash(m)
a.State = notifier.StatePending
ar.alerts[a.ID] = a
logger.Infof("alert %q (%d) restored to state at %v", a.Name, a.ID, a.Start)
}
return nil
}