// InspectOriginalAndGhostTables compares original and ghost tables to see whether the migration
// makes sense and is valid. It extracts the list of shared columns and the chosen migration unique key
func (this *Inspector) InspectOriginalAndGhostTables() (err error) {
this.migrationContext.GhostTableColumns, this.migrationContext.GhostTableUniqueKeys, err = this.InspectTableColumnsAndUniqueKeys(this.migrationContext.GetGhostTableName())
if err != nil {
return err
}
sharedUniqueKeys, err := this.getSharedUniqueKeys(this.migrationContext.OriginalTableUniqueKeys, this.migrationContext.GhostTableUniqueKeys)
if err != nil {
return err
}
if len(sharedUniqueKeys) == 0 {
return fmt.Errorf("No shared unique key can be found after ALTER! Bailing out")
}
this.migrationContext.UniqueKey = sharedUniqueKeys[0]
log.Infof("Chosen shared unique key is %s", this.migrationContext.UniqueKey.Name)
if this.migrationContext.UniqueKey.HasNullable {
if this.migrationContext.NullableUniqueKeyAllowed {
log.Warningf("Chosen key (%s) has nullable columns. You have supplied with --allow-nullable-unique-key and so this migration proceeds. As long as there aren't NULL values in this key's column, migration should be fine. NULL values will corrupt migration's data", this.migrationContext.UniqueKey)
} else {
return fmt.Errorf("Chosen key (%s) has nullable columns. Bailing out. To force this operation to continue, supply --allow-nullable-unique-key flag. Only do so if you are certain there are no actual NULL values in this key. As long as there aren't, migration should be fine. NULL values in columns of this key will corrupt migration's data", this.migrationContext.UniqueKey)
}
}
if !this.migrationContext.UniqueKey.IsPrimary() {
if this.migrationContext.OriginalBinlogRowImage != "FULL" {
return fmt.Errorf("binlog_row_image is '%s' and chosen key is %s, which is not the primary key. This operation cannot proceed. You may `set global binlog_row_image='full'` and try again", this.migrationContext.OriginalBinlogRowImage, this.migrationContext.UniqueKey)
}
}
this.migrationContext.SharedColumns, this.migrationContext.MappedSharedColumns = this.getSharedColumns(this.migrationContext.OriginalTableColumns, this.migrationContext.GhostTableColumns, this.migrationContext.ColumnRenameMap)
log.Infof("Shared columns are %s", this.migrationContext.SharedColumns)
// By fact that a non-empty unique key exists we also know the shared columns are non-empty
return nil
}
// Attempt to resolve a hostname. This may return a database cached hostname or otherwise
// it may resolve the hostname via CNAME
func ResolveHostname(hostname string) (string, error) {
hostname = strings.TrimSpace(hostname)
if hostname == "" {
return hostname, errors.New("Will not resolve empty hostname")
}
if strings.Contains(hostname, ",") {
return hostname, fmt.Errorf("Will not resolve multi-hostname: %+v", hostname)
}
if (&InstanceKey{Hostname: hostname}).IsDetached() {
// quietly abort. Nothign to do. The hostname is detached for a reason: it
// will not be resolved, for sure.
return hostname, nil
}
// First go to lightweight cache
if resolvedHostname, found := hostnameResolvesLightweightCache.Get(hostname); found {
return resolvedHostname.(string), nil
}
if !hostnameResolvesLightweightCacheLoadedOnceFromDB {
// A continuous-discovery will first make sure to load all resolves from DB.
// However cli does not do so.
// Anyway, it seems like the cache was not loaded from DB. Before doing real resolves,
// let's try and get the resolved hostname from database.
if !HostnameResolveMethodIsNone() {
if resolvedHostname, err := ReadResolvedHostname(hostname); err == nil && resolvedHostname != "" {
hostnameResolvesLightweightCache.Set(hostname, resolvedHostname, 0)
return resolvedHostname, nil
}
}
}
// Unfound: resolve!
log.Debugf("Hostname unresolved yet: %s", hostname)
resolvedHostname, err := resolveHostname(hostname)
if config.Config.RejectHostnameResolvePattern != "" {
// Reject, don't even cache
if matched, _ := regexp.MatchString(config.Config.RejectHostnameResolvePattern, resolvedHostname); matched {
log.Warningf("ResolveHostname: %+v resolved to %+v but rejected due to RejectHostnameResolvePattern '%+v'", hostname, resolvedHostname, config.Config.RejectHostnameResolvePattern)
return hostname, nil
}
}
if err != nil {
// Problem. What we'll do is cache the hostname for just one minute, so as to avoid flooding requests
// on one hand, yet make it refresh shortly on the other hand. Anyway do not write to database.
hostnameResolvesLightweightCache.Set(hostname, resolvedHostname, time.Minute)
return hostname, err
}
// Good result! Cache it, also to DB
log.Debugf("Cache hostname resolve %s as %s", hostname, resolvedHostname)
UpdateResolvedHostname(hostname, resolvedHostname)
return resolvedHostname, nil
}
// discoverInstance will attempt discovering an instance (unless it is already up to date) and will
// list down its master and slaves (if any) for further discovery.
func discoverInstance(instanceKey inst.InstanceKey) {
start := time.Now()
instanceKey.Formalize()
if !instanceKey.IsValid() {
return
}
if existsInCacheError := recentDiscoveryOperationKeys.Add(instanceKey.DisplayString(), true, cache.DefaultExpiration); existsInCacheError != nil {
// Just recently attempted
return
}
instance, found, err := inst.ReadInstance(&instanceKey)
if found && instance.IsUpToDate && instance.IsLastCheckValid {
// we've already discovered this one. Skip!
return
}
discoveriesCounter.Inc(1)
// First we've ever heard of this instance. Continue investigation:
instance, err = inst.ReadTopologyInstance(&instanceKey)
// panic can occur (IO stuff). Therefore it may happen
// that instance is nil. Check it.
if instance == nil {
failedDiscoveriesCounter.Inc(1)
log.Warningf("discoverInstance(%+v) instance is nil in %.3fs, error=%+v", instanceKey, time.Since(start).Seconds(), err)
return
}
log.Debugf("Discovered host: %+v, master: %+v, version: %+v in %.3fs", instance.Key, instance.MasterKey, instance.Version, time.Since(start).Seconds())
if atomic.LoadInt64(&isElectedNode) == 0 {
// Maybe this node was elected before, but isn't elected anymore.
// If not elected, stop drilling up/down the topology
return
}
// Investigate slaves:
for _, slaveKey := range instance.SlaveHosts.GetInstanceKeys() {
slaveKey := slaveKey
if slaveKey.IsValid() {
discoveryQueue.Push(slaveKey)
}
}
// Investigate master:
if instance.MasterKey.IsValid() {
discoveryQueue.Push(instance.MasterKey)
}
}
// pop the entry and remove it from known keys;
// blocks if queue is empty.
func (q *Queue) Pop() inst.InstanceKey {
q.Lock()
queue := q.queue
q.Unlock()
key := <-queue
q.Lock()
defer q.Unlock()
// alarm if have been waiting for too long
timeOnQueue := time.Since(q.knownKeys[key])
if timeOnQueue > time.Duration(config.Config.InstancePollSeconds)*time.Second {
log.Warningf("key %v spent %.4fs waiting on a discoveryQueue", key, timeOnQueue.Seconds())
}
delete(q.knownKeys, key)
return key
}
// discoverInstance will attempt discovering an instance (unless it is already up to date) and will
// list down its master and slaves (if any) for further discovery.
func discoverInstance(instanceKey inst.InstanceKey) {
instanceKey.Formalize()
if !instanceKey.IsValid() {
return
}
if existsInCacheError := recentDiscoveryOperationKeys.Add(instanceKey.DisplayString(), true, cache.DefaultExpiration); existsInCacheError != nil {
// Just recently attempted
return
}
instance, found, err := inst.ReadInstance(&instanceKey)
if found && instance.IsUpToDate && instance.IsLastCheckValid {
// we've already discovered this one. Skip!
return
}
discoveriesCounter.Inc(1)
// First we've ever heard of this instance. Continue investigation:
instance, err = inst.ReadTopologyInstance(&instanceKey)
// panic can occur (IO stuff). Therefore it may happen
// that instance is nil. Check it.
if instance == nil {
failedDiscoveriesCounter.Inc(1)
log.Warningf("instance is nil in discoverInstance. key=%+v, error=%+v", instanceKey, err)
return
}
log.Debugf("Discovered host: %+v, master: %+v", instance.Key, instance.MasterKey)
if !isElectedNode {
// Maybe this node was elected before, but isn't elected anymore.
// If not elected, stop drilling down to further investigate slaves.
return
}
// Investigate slaves:
for _, slaveKey := range instance.SlaveHosts.GetInstanceKeys() {
discoveryInstanceKeys <- slaveKey
}
// Investigate master:
discoveryInstanceKeys <- instance.MasterKey
}
// DiscoverInstance will attempt discovering an instance (unless it is already up to date) and will
// list down its master and slaves (if any) for further discovery.
func DiscoverInstance(instanceKey inst.InstanceKey) {
instanceKey.Formalize()
if !instanceKey.IsValid() {
return
}
instance, found, err := inst.ReadInstance(&instanceKey)
if found && instance.IsUpToDate && instance.IsLastCheckValid {
// we've already discovered this one. Skip!
goto Cleanup
}
// First we've ever heard of this instance. Continue investigation:
instance, err = inst.ReadTopologyInstance(&instanceKey)
// panic can occur (IO stuff). Therefore it may happen
// that instance is nil. Check it.
if err != nil || instance == nil {
log.Warningf("instance is nil in DiscoverInstance. key=%+v, error=%+v", instanceKey, err)
goto Cleanup
}
log.Debugf("Discovered host: %+v, master: %+v", instance.Key, instance.MasterKey)
// Investigate slaves:
for _, slaveKey := range instance.SlaveHosts.GetInstanceKeys() {
discoveryInstanceKeys <- slaveKey
}
// Investigate master:
discoveryInstanceKeys <- instance.MasterKey
Cleanup:
}
// Start discovery begins a one time asynchronuous discovery process for the given
// instance and all of its topology connected instances.
// That is, the instance will be investigated for master and slaves, and the routines will follow on
// each and every such found master/slave.
// In essense, assuming all slaves in a replication topology are running, and given a single instance
// in such topology, this function will detect the entire topology.
func StartDiscovery(instanceKey inst.InstanceKey) {
log.Infof("Starting discovery at %+v", instanceKey)
pendingTokens := make(chan bool, maxConcurrency)
completedTokens := make(chan bool, maxConcurrency)
AccountedDiscoverInstance(instanceKey, pendingTokens, completedTokens)
go handleDiscoveryRequests(pendingTokens, completedTokens)
// Block until all are complete
for {
select {
case <-pendingTokens:
<-completedTokens
default:
inst.AuditOperation("start-discovery", &instanceKey, "")
return
}
}
}
// ContinuousDiscovery starts an asynchronuous infinite discovery process where instances are
// periodically investigated and their status captured, and long since unseen instances are
// purged and forgotten.
func ContinuousDiscovery() {
log.Infof("Starting continuous discovery")
inst.LoadHostnameResolveCacheFromDatabase()
go handleDiscoveryRequests(nil, nil)
tick := time.Tick(time.Duration(config.Config.DiscoveryPollSeconds) * time.Second)
forgetUnseenTick := time.Tick(time.Minute)
recoverTick := time.Tick(10 * time.Second)
var snapshotTopologiesTick <-chan time.Time
if config.Config.SnapshotTopologiesIntervalHours > 0 {
snapshotTopologiesTick = time.Tick(time.Duration(config.Config.SnapshotTopologiesIntervalHours) * time.Hour)
}
elected := false
_ = CreateElectionAnchor(false)
for {
select {
case <-tick:
if elected, _ = AttemptElection(); elected {
instanceKeys, _ := inst.ReadOutdatedInstanceKeys()
log.Debugf("outdated keys: %+v", instanceKeys)
for _, instanceKey := range instanceKeys {
discoveryInstanceKeys <- instanceKey
}
} else {
log.Debugf("Not elected as active node; polling")
}
case <-forgetUnseenTick:
// See if we should also forget objects (lower frequency)
go func() {
if elected {
inst.ForgetLongUnseenInstances()
inst.ForgetUnseenInstancesDifferentlyResolved()
inst.ForgetExpiredHostnameResolves()
inst.DeleteInvalidHostnameResolves()
inst.ReviewUnseenInstances()
inst.InjectUnseenMasters()
inst.ResolveUnknownMasterHostnameResolves()
inst.ExpireMaintenance()
inst.ExpireDowntime()
inst.ExpireCandidateInstances()
inst.ExpireHostnameUnresolve()
inst.ExpireClusterDomainName()
}
if !elected {
// Take this opportunity to refresh yourself
inst.LoadHostnameResolveCacheFromDatabase()
}
inst.ReadClusterAliases()
HealthTest()
}()
case <-recoverTick:
go func() {
if elected {
ClearActiveFailureDetections()
ClearActiveRecoveries()
CheckAndRecover(nil, nil, false, false)
}
}()
case <-snapshotTopologiesTick:
go func() {
inst.SnapshotTopologies()
}()
}
}
}
func pollAgent(hostname string) error {
polledAgent, err := agent.GetAgent(hostname)
agent.UpdateAgentLastChecked(hostname)
if err != nil {
return log.Errore(err)
}
err = agent.UpdateAgentInfo(hostname, polledAgent)
if err != nil {
return log.Errore(err)
}
return nil
}
// ContinuousAgentsPoll starts an asynchronuous infinite process where agents are
// periodically investigated and their status captured, and long since unseen agents are
// purged and forgotten.
func ContinuousAgentsPoll() {
log.Infof("Starting continuous agents poll")
go discoverSeededAgents()
tick := time.Tick(time.Duration(config.Config.DiscoveryPollSeconds) * time.Second)
forgetUnseenTick := time.Tick(time.Hour)
for _ = range tick {
agentsHosts, _ := agent.ReadOutdatedAgentsHosts()
log.Debugf("outdated agents hosts: %+v", agentsHosts)
for _, hostname := range agentsHosts {
go pollAgent(hostname)
}
// See if we should also forget agents (lower frequency)
select {
case <-forgetUnseenTick:
agent.ForgetLongUnseenAgents()
agent.FailStaleSeeds()
default:
}
}
}
func discoverSeededAgents() {
for seededAgent := range agent.SeededAgents {
instanceKey := inst.InstanceKey{Hostname: seededAgent.Hostname, Port: int(seededAgent.MySQLPort)}
go StartDiscovery(instanceKey)
}
}
// discoverInstance will attempt discovering an instance (unless it is already up to date) and will
// list down its master and slaves (if any) for further discovery.
func discoverInstance(instanceKey inst.InstanceKey) {
instanceKey.Formalize()
if !instanceKey.IsValid() {
return
}
instance, found, err := inst.ReadInstance(&instanceKey)
if found && instance.IsUpToDate && instance.IsLastCheckValid {
// we've already discovered this one. Skip!
goto Cleanup
}
discoveriesCounter.Inc(1)
// First we've ever heard of this instance. Continue investigation:
instance, err = inst.ReadTopologyInstance(&instanceKey)
// panic can occur (IO stuff). Therefore it may happen
// that instance is nil. Check it.
if err != nil || instance == nil {
failedDiscoveriesCounter.Inc(1)
log.Warningf("instance is nil in discoverInstance. key=%+v, error=%+v", instanceKey, err)
goto Cleanup
}
log.Debugf("Discovered host: %+v, master: %+v", instance.Key, instance.MasterKey)
if !isElectedNode {
// Maybe this node was elected before, but isn't elected anymore.
// If not elected, stop drilling down to further investigate slaves.
return
}
// Investigate slaves:
for _, slaveKey := range instance.SlaveHosts.GetInstanceKeys() {
discoveryInstanceKeys <- slaveKey
}
// Investigate master:
discoveryInstanceKeys <- instance.MasterKey
Cleanup:
}
// Start discovery begins a one time asynchronuous discovery process for the given
// instance and all of its topology connected instances.
// That is, the instance will be investigated for master and slaves, and the routines will follow on
// each and every such found master/slave.
// In essense, assuming all slaves in a replication topology are running, and given a single instance
// in such topology, this function will detect the entire topology.
func StartDiscovery(instanceKey inst.InstanceKey) {
log.Infof("Starting discovery at %+v", instanceKey)
pendingTokens := make(chan bool, maxConcurrency)
completedTokens := make(chan bool, maxConcurrency)
accountedDiscoverInstance(instanceKey, pendingTokens, completedTokens)
go handleDiscoveryRequests(pendingTokens, completedTokens)
// Block until all are complete
for {
select {
case <-pendingTokens:
<-completedTokens
default:
inst.AuditOperation("start-discovery", &instanceKey, "")
return
}
}
}
func initGraphiteMetrics() error {
if config.Config.GraphiteAddr == "" {
return nil
}
if config.Config.GraphitePath == "" {
return log.Errorf("No graphite path provided (see GraphitePath config variable). Will not log to graphite")
}
addr, err := net.ResolveTCPAddr("tcp", config.Config.GraphiteAddr)
if err != nil {
return log.Errore(err)
}
graphitePathHostname := ThisHostname
if config.Config.GraphiteConvertHostnameDotsToUnderscores {
graphitePathHostname = strings.Replace(graphitePathHostname, ".", "_", -1)
}
graphitePath := config.Config.GraphitePath
graphitePath = strings.Replace(graphitePath, "{hostname}", graphitePathHostname, -1)
log.Debugf("Will log to graphite on %+v, %+v", config.Config.GraphiteAddr, graphitePath)
go graphite.Graphite(metrics.DefaultRegistry, 1*time.Minute, graphitePath, addr)
return nil
}
// ContinuousDiscovery starts an asynchronuous infinite discovery process where instances are
// periodically investigated and their status captured, and long since unseen instances are
// purged and forgotten.
func ContinuousDiscovery() {
log.Infof("Starting continuous discovery")
inst.LoadHostnameResolveCacheFromDatabase()
go handleDiscoveryRequests(nil, nil)
tick := time.Tick(time.Duration(config.Config.DiscoveryPollSeconds) * time.Second)
forgetUnseenTick := time.Tick(time.Minute)
recoverTick := time.Tick(10 * time.Second)
var snapshotTopologiesTick <-chan time.Time
if config.Config.SnapshotTopologiesIntervalHours > 0 {
snapshotTopologiesTick = time.Tick(time.Duration(config.Config.SnapshotTopologiesIntervalHours) * time.Hour)
}
go initGraphiteMetrics()
for {
select {
case <-tick:
go func() {
if isElectedNode, _ = attemptElection(); isElectedNode {
instanceKeys, _ := inst.ReadOutdatedInstanceKeys()
log.Debugf("outdated keys: %+v", instanceKeys)
for _, instanceKey := range instanceKeys {
discoveryInstanceKeys <- instanceKey
}
} else {
log.Debugf("Not elected as active node; polling")
}
discoveryQueueLengthGauge.Update(int64(len(discoveryInstanceKeys)))
}()
case <-forgetUnseenTick:
// See if we should also forget objects (lower frequency)
go func() {
if isElectedNode {
inst.ForgetLongUnseenInstances()
inst.ForgetUnseenInstancesDifferentlyResolved()
inst.ForgetExpiredHostnameResolves()
inst.DeleteInvalidHostnameResolves()
inst.ReviewUnseenInstances()
inst.InjectUnseenMasters()
inst.ResolveUnknownMasterHostnameResolves()
inst.ExpireMaintenance()
inst.ExpireDowntime()
inst.ExpireCandidateInstances()
inst.ExpireHostnameUnresolve()
inst.ExpireClusterDomainName()
inst.ExpireAudit()
inst.ExpireMasterPositionEquivalence()
}
if !isElectedNode {
// Take this opportunity to refresh yourself
inst.LoadHostnameResolveCacheFromDatabase()
}
inst.ReadClusterAliases()
HealthTest()
}()
case <-recoverTick:
go func() {
if isElectedNode {
ClearActiveFailureDetections()
ClearActiveRecoveries()
CheckAndRecover(nil, nil, false, false)
}
}()
case <-snapshotTopologiesTick:
go func() {
inst.SnapshotTopologies()
}()
}
}
}
func pollAgent(hostname string) error {
polledAgent, err := agent.GetAgent(hostname)
agent.UpdateAgentLastChecked(hostname)
if err != nil {
return log.Errore(err)
}
err = agent.UpdateAgentInfo(hostname, polledAgent)
if err != nil {
return log.Errore(err)
}
return nil
}
// ContinuousAgentsPoll starts an asynchronuous infinite process where agents are
// periodically investigated and their status captured, and long since unseen agents are
// purged and forgotten.
func ContinuousAgentsPoll() {
log.Infof("Starting continuous agents poll")
go discoverSeededAgents()
tick := time.Tick(time.Duration(config.Config.DiscoveryPollSeconds) * time.Second)
forgetUnseenTick := time.Tick(time.Hour)
for range tick {
agentsHosts, _ := agent.ReadOutdatedAgentsHosts()
log.Debugf("outdated agents hosts: %+v", agentsHosts)
for _, hostname := range agentsHosts {
go pollAgent(hostname)
}
// See if we should also forget agents (lower frequency)
select {
case <-forgetUnseenTick:
agent.ForgetLongUnseenAgents()
agent.FailStaleSeeds()
default:
}
}
}
func discoverSeededAgents() {
for seededAgent := range agent.SeededAgents {
instanceKey := inst.InstanceKey{Hostname: seededAgent.Hostname, Port: int(seededAgent.MySQLPort)}
go StartDiscovery(instanceKey)
}
}