go/cmd/vtbackup: wait for plugins to finish initializing

[maxenglander]

Description

VTB E2E tests sometimes fail due to:

1. Some VTB plugins (in particular plugin_opentsdb.go register with servenv.OnRun

   vitess/go/cmd/vtbackup/cli/plugin_opentsdb.go

   Line 24 in 3038580

   opentsdb.Init("vtbackup")

   vitess/go/stats/opentsdb/init.go

   Lines 36 to 38 in 3038580

   	func Init(prefix string) {
   	// Needs to happen in servenv.OnRun() instead of init because it requires flag parsing and logging
   	servenv.OnRun(func() {

2. VTB starts servenv in a goroutine

   vitess/go/cmd/vtbackup/cli/vtbackup.go

   Line 232 in 3038580

   go servenv.RunDefault()

3. VTB starts writing stats shortly after

   vitess/go/cmd/vtbackup/cli/vtbackup.go

   Line 258 in 3038580

   phase.Set(phaseName, int64(0))

This sequence means that sometimes VTB starts writing stats before servenv plugins have finished initializing, which makes the E2E test that verifies stats flaky.

This PR is a hacky way to deflake. There are less hacky ways we could do this, e.g. exposing some kind of readiness indicator directly from the servenv package.

Related Issue(s)

Fixes #14114

Checklist

• "Backport to:" labels have been added if this change should be back-ported
• Tests were added or are not required (this PR deflakes existing tests)
• Did the new or modified tests pass consistently locally and on the CI
• Documentation was added or is not required

Deployment Notes

[vitess-bot]

Review Checklist

Hello reviewers! 👋 Please follow this checklist when reviewing this Pull Request.

General

• Ensure that the Pull Request has a descriptive title.
• Ensure there is a link to an issue (except for internal cleanup and flaky test fixes), new features should have an RFC that documents use cases and test cases.

Tests

• Bug fixes should have at least one unit or end-to-end test, enhancement and new features should have a sufficient number of tests.

Documentation

• Apply the release notes (needs details) label if users need to know about this change.
• New features should be documented.
• There should be some code comments as to why things are implemented the way they are.
• There should be a comment at the top of each new or modified test to explain what the test does.

New flags

• Is this flag really necessary?
• Flag names must be clear and intuitive, use dashes (-), and have a clear help text.

If a workflow is added or modified:

• Each item in Jobs should be named in order to mark it as required.
• If the workflow needs to be marked as required, the maintainer team must be notified.

Backward compatibility

• Protobuf changes should be wire-compatible.
• Changes to _vt tables and RPCs need to be backward compatible.
• RPC changes should be compatible with vitess-operator
• If a flag is removed, then it should also be removed from vitess-operator and arewefastyet, if used there.
• vtctl command output order should be stable and awk-able.

[ajm188]

this seems fine to me; it might be a bit more readable (in absence of the comment) to do this:

diff --git a/go/stats/opentsdb/init.go b/go/stats/opentsdb/init.go
index 51186ad765..0782deb8a2 100644
--- a/go/stats/opentsdb/init.go
+++ b/go/stats/opentsdb/init.go
@@ -17,18 +17,24 @@ limitations under the License.
 package opentsdb
 
 import (
+	"context"
 	"encoding/json"
 	"fmt"
 	"net/http"
 	"net/url"
 	"sort"
+	"sync"
 
 	"vitess.io/vitess/go/stats"
 	"vitess.io/vitess/go/vt/log"
 	"vitess.io/vitess/go/vt/servenv"
 )
 
-var singletonBackend stats.PushBackend
+var (
+	singletonOnce    sync.Once
+	singletonInit    = make(chan struct{})
+	singletonBackend stats.PushBackend
+)
 
 // Init attempts to create a singleton *opentsdb.backend and register it as a PushBackend.
 // If it fails to create one, this is a noop. The prefix argument is an optional string
@@ -41,12 +47,23 @@ func Init(prefix string) {
 		if err != nil {
 			log.Infof("Failed to initialize singleton opentsdb backend: %v", err)
 		} else {
+			defer singletonOnce.Do(func() { close(singletonInit) })
+
 			singletonBackend = backend
 			log.Info("Initialized opentsdb backend.")
 		}
 	})
 }
 
+func AwaitInit(ctx context.Context) error {
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-singletonInit:
+		return nil
+	}
+}
+
 // InitWithoutServenv initializes the opentsdb without servenv
 func InitWithoutServenv(prefix string) (stats.PushBackend, error) {
 	b, err := newBackend(prefix)

and then here:

go servenv.RunDefault()
if err := opentsdb.AwaitInit(ctx); err != nil {
    return err
}

(or abstract a bit further so we don't expose the particular plugins (i.e. stats.AwaitPlugins(ctx)))

[maxenglander]

Thanks for the suggestion, I'll go with the middle option for now (opentsdb.AwaitInit).

[maxenglander]

Actually going with third option since we don't know if user will select --stats_backend opentsdb or something else.

[ajm188]

But opentsdbs OnRun still gets registered either way, no?

[maxenglander]

@ajm188 ah that's true 😳

Do you prefer I switch to that or are you OK with what I pushed up for stats.AwaitBackend?

[ajm188]

so i like yours structurally (since it's possible of course to compile vitess without a particular plugin_ file), however, the problem is this close happens during init() and not during servenv.OnRun

we need some mechanism for a plugin to signal back to the main stats package that it's been fully initialized.

[ajm188]

something maybe like this:

diff --git a/go/stats/export.go b/go/stats/export.go
index 8bda85c87b..07381f9929 100644
--- a/go/stats/export.go
+++ b/go/stats/export.go
@@ -29,6 +29,7 @@ package stats
 
 import (
 	"bytes"
+	"context"
 	"expvar"
 	"fmt"
 	"strconv"
@@ -218,6 +219,12 @@ func RegisterPushBackend(name string, backend PushBackend) {
 		log.Fatalf("PushBackend %s already exists; can't register the same name multiple times", name)
 	}
 	pushBackends[name] = backend
+	backends[name] = &struct {
+		once sync.Once
+		ch   chan struct{}
+	}{
+		ch: make(chan struct{}),
+	}
 	if emitStats {
 		// Start a single goroutine to emit stats periodically
 		once.Do(func() {
@@ -226,6 +233,36 @@ func RegisterPushBackend(name string, backend PushBackend) {
 	}
 }
 
+var (
+	backends = map[string]*struct {
+		once sync.Once
+		ch   chan struct{}
+	}{}
+)
+
+func InitializeBackend(backend string) {
+	if d, ok := backends[backend]; ok {
+		d.once.Do(func() { close(d.ch) })
+	}
+}
+
+func AwaitBackend(ctx context.Context) error {
+	pushBackendsLock.Lock()
+	defer pushBackendsLock.Unlock()
+
+	d, ok := backends[statsBackend]
+	if !ok {
+		return fmt.Errorf("...")
+	}
+
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-d.ch:
+		return nil
+	}
+}
+
 // emitToBackend does a periodic emit to the selected PushBackend. If a push fails,
 // it will be logged as a warning (but things will otherwise proceed as normal).
 func emitToBackend(emitPeriod *time.Duration) {
diff --git a/go/stats/opentsdb/init.go b/go/stats/opentsdb/init.go
index 51186ad765..462a442501 100644
--- a/go/stats/opentsdb/init.go
+++ b/go/stats/opentsdb/init.go
@@ -41,6 +41,8 @@ func Init(prefix string) {
 		if err != nil {
 			log.Infof("Failed to initialize singleton opentsdb backend: %v", err)
 		} else {
+			defer stats.Initialize("opentsdb")
+
 			singletonBackend = backend
 			log.Info("Initialized opentsdb backend.")
 		}

[maxenglander]

@ajm188 and I chatted about this part:

the problem is this close happens during init() and not during servenv.OnRun

It turns out opentsdb is a bit special in this regard in that it exposes a Init function that uses servenv.OnRun under the hood. So the full sequence is:

1. vitess/go/cmd/vtbackup/cli/plugin_opentsdb.go

   Line 24 in 83932bc

   opentsdb.Init("vtbackup")

2. vitess/go/stats/opentsdb/init.go

   Lines 36 to 38 in 83932bc

   	func Init(prefix string) {
   	// Needs to happen in servenv.OnRun() instead of init because it requires flag parsing and logging
   	servenv.OnRun(func() {

3. vitess/go/vt/servenv/run.go

   Lines 40 to 43 in 83932bc

   	func Run(port int) {
   	populateListeningURL(int32(port))
   	createGRPCServer()
   	onRunHooks.Fire()

4. vitess/go/stats/opentsdb/init.go

   Lines 51 to 58 in 83932bc

   	func InitWithoutServenv(prefix string) (stats.PushBackend, error) {
   	b, err := newBackend(prefix)
   	if err != nil {
   	return nil, err
   	}
   	stats.RegisterPushBackend("opentsdb", b)

5. vitess/go/stats/export.go

   Lines 228 to 236 in 83932bc

   	func RegisterPushBackend(name string, backend PushBackend) {
   	pushBackendsLock.Lock()
   	defer pushBackendsLock.Unlock()
   	if _, ok := pushBackends[name]; ok {
   	log.Fatalf("PushBackend %s already exists; can't register the same name multiple times", name)
   	}
   	pushBackends[name] = backend
   	if name == statsBackend {
   	close(statsBackendInit)

It's probably worth making the opentsdb package and consumer plugins more like the rest of the code base at some point, but will leave things as they are for now if that's alright.

[ajm188]

yep, you're totally right! this is looking good to me now