// TODO: also check the build slaves to see if there are alerts for currently running builds that
// haven't shown up in CBE yet.
func (a *Analyzer) builderAlerts(masterName string, builderName string, b *messages.Builder) ([]messages.Alert, []error) {
if len(b.CachedBuilds) == 0 {
// TODO: Make an alert for this?
return nil, []error{errNoRecentBuilds}
}
recentBuildIDs := b.CachedBuilds
// Should be a *reverse* sort.
sort.Sort(buildNums(recentBuildIDs))
if len(recentBuildIDs) > a.MaxRecentBuilds {
recentBuildIDs = recentBuildIDs[:a.MaxRecentBuilds]
}
alerts, errs := []messages.Alert{}, []error{}
lastBuild, lastCompletedBuild, err := a.lastBuilds(masterName, builderName, recentBuildIDs)
if err != nil {
errs = append(errs, err)
return nil, errs
}
// Examining only the latest build is probably suboptimal since if it's still in progress it might
// not have hit a step that is going to fail and has failed repeatedly for the last few builds.
// AKA "Reliable failures". TODO: Identify "Reliable failures"
lastStep, lastUpdated, err := a.latestBuildStep(lastBuild)
if err != nil {
errs = append(errs, fmt.Errorf("Couldn't get latest build step for %s.%s: %v", masterName, builderName, err))
return alerts, errs
}
elapsed := a.Now().Sub(lastUpdated.Time())
links := []messages.Link{
{"Builder", client.BuilderURL(masterName, builderName)},
{"Last build", client.BuildURL(masterName, builderName, lastBuild.Number)},
{"Last build step", client.StepURL(masterName, builderName, lastStep, lastBuild.Number)},
}
switch b.State {
case messages.StateBuilding:
if elapsed > a.HungBuilderThresh && lastStep != StepCompletedRun {
alerts = append(alerts, messages.Alert{
Key: fmt.Sprintf("%s.%s.hung", masterName, builderName),
Title: fmt.Sprintf("%s.%s is hung in step %s.", masterName, builderName, lastStep),
Body: fmt.Sprintf("%s.%s has been building for %v (last step update %s), past the alerting threshold of %v", masterName, builderName, elapsed, lastUpdated.Time(), a.HungBuilderThresh),
Severity: hungBuilderSev,
Time: messages.TimeToEpochTime(a.Now()),
Links: links,
})
// Note, just because it's building doesn't mean it's in a good state. If the last N builds
// all failed (for some large N) then this might still be alertable.
}
case messages.StateOffline:
if elapsed > a.OfflineBuilderThresh {
alerts = append(alerts, messages.Alert{
Key: fmt.Sprintf("%s.%s.offline", masterName, builderName),
Title: fmt.Sprintf("%s.%s is offline.", masterName, builderName),
Body: fmt.Sprintf("%s.%s has been offline for %v (last step update %s %v), past the alerting threshold of %v", masterName, builderName, elapsed, lastUpdated.Time(), float64(lastUpdated), a.OfflineBuilderThresh),
Severity: offlineBuilderSev,
Time: messages.TimeToEpochTime(a.Now()),
Links: links,
})
}
case messages.StateIdle:
if b.PendingBuilds > a.IdleBuilderCountThresh {
alerts = append(alerts, messages.Alert{
Key: fmt.Sprintf("%s.%s.idle", masterName, builderName),
Title: fmt.Sprintf("%s.%s is idle with too many pending builds.", masterName, builderName),
Body: fmt.Sprintf("%s.%s is idle with %d pending builds, past the alerting threshold of %d", masterName, builderName, b.PendingBuilds, a.IdleBuilderCountThresh),
Severity: idleBuilderSev,
Time: messages.TimeToEpochTime(a.Now()),
Links: links,
})
}
default:
log.Errorf("Unknown %s.%s builder state: %s", masterName, builderName, b.State)
}
// Check for alerts on the most recent complete build
log.Infof("Checking %d most recent builds for alertable step failures: %s/%s", len(recentBuildIDs), masterName, builderName)
as, es := a.builderStepAlerts(masterName, builderName, []int64{lastCompletedBuild.Number})
if len(as) > 0 {
mostRecentComplete := 0
for i, id := range recentBuildIDs {
if id == lastCompletedBuild.Number {
mostRecentComplete = i
}
}
as, es = a.builderStepAlerts(masterName, builderName, recentBuildIDs[mostRecentComplete:])
alerts = append(alerts, as...)
errs = append(errs, es...)
}
return alerts, errs
}
// stepFailureAlerts returns alerts generated from step failures. It applies filtering
// logic specified in the gatekeeper config to ignore some failures.
func (a *Analyzer) stepFailureAlerts(failures []stepFailure) ([]messages.Alert, error) {
ret := []messages.Alert{}
type res struct {
f stepFailure
a *messages.Alert
err error
}
// Might not need full capacity buffer, since some failures are ignored below.
rs := make(chan res, len(failures))
scannedFailures := []stepFailure{}
for _, failure := range failures {
// goroutine/channel because the reasonsForFailure call potentially
// blocks on IO.
if failure.step.Name == "steps" {
// check results to see if it's an array of [4]
// That's a purple failure, which should go to infra/trooper.
log.Infof("steps results: %+v", failure.step)
if len(failure.step.Results) > 0 {
if r, ok := failure.step.Results[0].(float64); ok && r == resInfraFailure {
// TODO: Create a trooper alert about this.
log.Errorf("INFRA FAILURE: %+v", failure)
}
}
continue
// The actual breaking step will appear later.
}
// Check the gatekeeper configs to see if this is ignorable.
if a.excludeFailure(failure.masterName, failure.builderName, failure.step.Name) {
continue
}
// Gets the named revision number from gnumbd metadata.
getCommitPos := func(b messages.Build, name string) (string, bool) {
for _, p := range b.Properties {
if p[0] == name {
s, ok := p[1].(string)
return s, ok
}
}
return "", false
}
scannedFailures = append(scannedFailures, failure)
go func(f stepFailure) {
alr := messages.Alert{
Title: fmt.Sprintf("Builder step failure: %s.%s", f.masterName, f.builderName),
Time: messages.EpochTime(a.Now().Unix()),
Type: "buildfailure",
}
regRanges := []messages.RegressionRange{}
revisionsByRepo := map[string][]string{}
// Get gnumbd sequence numbers for whatever this build pulled in.
chromiumPos, ok := getCommitPos(f.build, "got_revision_cp")
if ok {
regRanges = append(regRanges, messages.RegressionRange{
Repo: "chromium",
Positions: []string{chromiumPos},
})
}
blinkPos, ok := getCommitPos(f.build, "got_webkit_revision_cp")
if ok {
regRanges = append(regRanges, messages.RegressionRange{
Repo: "blink",
Positions: []string{blinkPos},
})
}
v8Pos, ok := getCommitPos(f.build, "got_v8_revision_cp")
if ok {
regRanges = append(regRanges, messages.RegressionRange{
Repo: "v8",
Positions: []string{v8Pos},
})
}
naclPos, ok := getCommitPos(f.build, "got_nacl_revision_cp")
if ok {
regRanges = append(regRanges, messages.RegressionRange{
Repo: "nacl",
Positions: []string{naclPos},
})
}
for _, change := range f.build.SourceStamp.Changes {
revisionsByRepo[change.Repository] = append(revisionsByRepo[change.Repository], change.Revision)
// change.Revision is *not* always a git hash. Sometimes it is a position from gnumbd.
// change.Revision is git hash or gnumbd depending on what exactly? Not obvious at this time.
// A potential problem here is when multiple repos have overlapping gnumbd ranges.
a.revisionSummaries[change.Revision] = messages.RevisionSummary{
GitHash: change.Revision,
Link: change.Revlink,
Description: trunc(change.Comments),
Author: change.Who,
When: change.When,
}
}
for repo, revisions := range revisionsByRepo {
regRanges = append(regRanges, messages.RegressionRange{
Repo: repo,
Revisions: revisions,
})
}
// If the builder has been failing on the same step for multiple builds in a row,
// we should have only one alert but indicate the range of builds affected.
// These are set in FirstFailure and LastFailure.
bf := messages.BuildFailure{
// FIXME: group builders?
Builders: []messages.AlertedBuilder{
{
Name: f.builderName,
URL: client.BuilderURL(f.masterName, f.builderName),
StartTime: f.build.CreatedTimestamp,
FirstFailure: f.build.Number,
LatestFailure: f.build.Number,
},
},
TreeCloser: a.wouldCloseTree(f.masterName, f.builderName, f.step.Name),
RegressionRanges: regRanges,
}
reasons := a.reasonsForFailure(f)
for _, r := range reasons {
bf.Reasons = append(bf.Reasons, messages.Reason{
TestName: r,
Step: f.step.Name,
URL: f.URL(),
})
}
alr.Key = alertKey(f.masterName, f.builderName, f.step.Name)
if len(bf.Reasons) == 0 {
log.Warningf("No reasons for step failure: %s", alr.Key)
bf.Reasons = append(bf.Reasons, messages.Reason{
Step: f.step.Name,
URL: f.URL(),
})
}
alr.Extension = bf
rs <- res{
f: f,
a: &alr,
err: nil,
}
}(failure)
}
for range scannedFailures {
r := <-rs
if r.a != nil {
ret = append(ret, *r.a)
}
}
return ret, nil
}
// stepFailureAlerts returns alerts generated from step failures. It applies filtering
// logic specified in the gatekeeper config to ignore some failures.
func (a *Analyzer) stepFailureAlerts(failures []stepFailure) ([]messages.Alert, error) {
ret := []messages.Alert{}
type res struct {
f stepFailure
a *messages.Alert
err error
}
// Might not need full capacity buffer, since some failures are ignored below.
rs := make(chan res, len(failures))
scannedFailures := []stepFailure{}
for _, failure := range failures {
// goroutine/channel because the reasonsForFailure call potentially
// blocks on IO.
if failure.step.Name == "steps" {
// check results to see if it's an array of [4]
// That's a purple failure, which should go to infra/trooper.
log.Infof("steps results: %+v", failure.step)
if len(failure.step.Results) > 0 {
if r, ok := failure.step.Results[0].(float64); ok && r == resInfraFailure {
// TODO: Create a trooper alert about this.
log.Errorf("INFRA FAILURE: %+v", failure)
}
}
continue
// The actual breaking step will appear later.
}
// Check the gatekeeper configs to see if this is ignorable.
if a.excludeFailure(failure.masterName, failure.builderName, failure.step.Name) {
continue
}
scannedFailures = append(scannedFailures, failure)
go func(f stepFailure) {
alr := messages.Alert{
Title: fmt.Sprintf("Builder step failure: %s.%s", f.masterName, f.builderName),
Time: messages.EpochTime(a.now().Unix()),
Type: "buildfailure",
}
regRanges := []messages.RegressionRange{}
revsByRepo := map[string][]string{}
for _, change := range f.build.SourceStamp.Changes {
// check change.Comments for text like
// "Cr-Commit-Position: refs/heads/master@{#330158}" to pick out revs from git commits.
revsByRepo[change.Repository] = append(revsByRepo[change.Repository], change.Revision)
}
for repo, revs := range revsByRepo {
regRanges = append(regRanges, messages.RegressionRange{
Repo: repo,
Revisions: revs,
})
}
// If the builder has been failing on the same step for multiple builds in a row,
// we should have only one alert but indicate the range of builds affected.
// These are set in FirstFailure and LastFailure.
bf := messages.BuildFailure{
// FIXME: group builders?
Builders: []messages.AlertedBuilder{
{
Name: f.builderName,
URL: client.BuilderURL(f.masterName, f.builderName),
StartTime: f.build.CreatedTimestamp,
FirstFailure: f.build.Number,
LatestFailure: f.build.Number,
},
},
TreeCloser: a.wouldCloseTree(f.masterName, f.builderName, f.step.Name),
RegressionRanges: regRanges,
}
reasons := a.reasonsForFailure(f)
for _, r := range reasons {
bf.Reasons = append(bf.Reasons, messages.Reason{
TestName: r,
Step: f.step.Name,
URL: f.URL(),
})
}
alr.Key = alertKey(f.masterName, f.builderName, f.step.Name)
if len(bf.Reasons) == 0 {
log.Warningf("No reasons for step failure: %s", alr.Key)
bf.Reasons = append(bf.Reasons, messages.Reason{
Step: f.step.Name,
URL: f.URL(),
})
}
alr.Extension = bf
rs <- res{
f: f,
a: &alr,
err: nil,
}
}(failure)
}
for range scannedFailures {
r := <-rs
if r.a != nil {
ret = append(ret, *r.a)
}
}
return ret, nil
}