Merge pull request #63 from github/cut-over-on-test-on-replica

test-on-replica to invoke cut-over swap

Author: Shlomi Noach

build.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 #
 #
-RELEASE_VERSION="0.8.8"
+RELEASE_VERSION="0.8.9"
 
 buildpath=/tmp/gh-ost
 target=gh-ost

go/base/context.go

@@ -31,9 +31,9 @@ const (
 type CutOver int
 
 const (
-	CutOverTwoStep CutOver = 1
-	CutOverVoluntaryLock
-	CutOverUdfWait
+	CutOverTwoStep       CutOver = iota
+	CutOverVoluntaryLock         = iota
+	CutOverUdfWait               = iota
 )
 
 const (

go/logic/applier.go

@@ -536,6 +536,31 @@ func (this *Applier) SwapTablesAtomic(sessionIdChan chan int64) error {
 	return nil
 }
 
+func (this *Applier) RenameTablesRollback() (renameError error) {
+
+	query := fmt.Sprintf(`rename /* gh-ost */ table %s.%s to %s.%s`,
+		sql.EscapeName(this.migrationContext.DatabaseName),
+		sql.EscapeName(this.migrationContext.OriginalTableName),
+		sql.EscapeName(this.migrationContext.DatabaseName),
+		sql.EscapeName(this.migrationContext.GetGhostTableName()),
+	)
+	log.Infof("Renaming back to ghost table")
+	if _, err := sqlutils.ExecNoPrepare(this.db, query); err != nil {
+		renameError = err
+	}
+	query = fmt.Sprintf(`rename /* gh-ost */ table %s.%s to %s.%s`,
+		sql.EscapeName(this.migrationContext.DatabaseName),
+		sql.EscapeName(this.migrationContext.GetOldTableName()),
+		sql.EscapeName(this.migrationContext.DatabaseName),
+		sql.EscapeName(this.migrationContext.OriginalTableName),
+	)
+	log.Infof("Renaming back to original table")
+	if _, err := sqlutils.ExecNoPrepare(this.db, query); err != nil {
+		renameError = err
+	}
+	return log.Errore(renameError)
+}
+
 // StopSlaveIOThread is applicable with --test-on-replica; it stops the IO thread, duh.
 // We need to keep the SQL thread active so as to complete processing received events,
 // and have them written to the binary log, so that we can then read them via streamer

go/logic/migrator.go

@@ -389,14 +389,25 @@ func (this *Migrator) stopWritesAndCompleteMigration() (err error) {
 	)
 
 	if this.migrationContext.TestOnReplica {
-		return this.stopWritesAndCompleteMigrationOnReplica()
+		// return this.stopWritesAndCompleteMigrationOnReplica()
+
+		// With `--test-on-replica` we stop replication thread, and then proceed to use
+		// the same cut-over phase as the master would use. That means we take locks
+		// and swap the tables.
+		// The difference is that we will later swap the tables back.
+		log.Debugf("testing on replica. Stopping replication IO thread")
+		if err := this.retryOperation(this.applier.StopSlaveNicely); err != nil {
+			return err
+		}
+		// We're merly testing, we don't want to keep this state. Rollback the renames as possible
+		defer this.applier.RenameTablesRollback()
 	}
-	// Running on master
+
 	if this.migrationContext.CutOverType == base.CutOverTwoStep {
 		return this.stopWritesAndCompleteMigrationOnMasterQuickAndBumpy()
 	}
 
-	{
+	if this.migrationContext.CutOverType == base.CutOverVoluntaryLock {
 		// Lock-based solution: we use low timeout and multiple attempts. But for
 		// each failed attempt, we throttle until replication lag is back to normal
 		if err := this.retryOperation(