diff --git a/go/base/context.go b/go/base/context.go index 5a06d0d4d..4d0c3b09d 100644 --- a/go/base/context.go +++ b/go/base/context.go @@ -238,7 +238,7 @@ type MigrationContext struct { AbortError error abortMutex *sync.Mutex - Metrics metrics.MemStatsGaugeEmitter + Metrics metrics.Emitter OriginalTableColumnsOnApplier *sql.ColumnList OriginalTableColumns *sql.ColumnList diff --git a/go/logic/migrator.go b/go/logic/migrator.go index 5ed2e2d9a..0737e67b4 100644 --- a/go/logic/migrator.go +++ b/go/logic/migrator.go @@ -652,7 +652,7 @@ func (mgtr *Migrator) Migrate() (err error) { } else { retrier = mgtr.retryOperation } - if err := retrier(mgtr.cutOver); err != nil { + if err := mgtr.cutOverWithMetrics(retrier); err != nil { return err } atomic.StoreInt64(&mgtr.migrationContext.CutOverCompleteFlag, 1) @@ -777,7 +777,7 @@ func (mgtr *Migrator) Revert() error { if err := mgtr.hooksExecutor.OnBeforeCutOver(); err != nil { return err } - if err := retrier(mgtr.cutOver); err != nil { + if err := mgtr.cutOverWithMetrics(retrier); err != nil { return err } atomic.StoreInt64(&mgtr.migrationContext.CutOverCompleteFlag, 1) @@ -797,6 +797,28 @@ func (mgtr *Migrator) ExecOnFailureHook() (err error) { return mgtr.hooksExecutor.OnFailure() } +func (mgtr *Migrator) cutOverWithMetrics(retrier func(func() error, ...bool) error) error { + return mgtr.cutOverOperationWithMetrics(retrier, mgtr.cutOver) +} + +func (mgtr *Migrator) cutOverOperationWithMetrics(retrier func(func() error, ...bool) error, operation func() error) error { + cutOverStartTime := time.Now() + err := retrier(func() error { + err := operation() + if err != nil { + metrics.RecordCutOverAttempt(mgtr.migrationContext.Metrics, metrics.CutOverOutcomeRetry) + return err + } + metrics.RecordCutOverAttempt(mgtr.migrationContext.Metrics, metrics.CutOverOutcomeSuccess) + return nil + }) + if err != nil { + metrics.RecordCutOverAttempt(mgtr.migrationContext.Metrics, metrics.CutOverOutcomeAbort) + } + metrics.RecordCutOverTotal(mgtr.migrationContext.Metrics, time.Since(cutOverStartTime), err) + return err +} + func (mgtr *Migrator) handleCutOverResult(cutOverError error) (err error) { if mgtr.migrationContext.TestOnReplica { // We're merely testing, we don't want to keep this state. Rollback the renames as possible @@ -959,9 +981,12 @@ func (mgtr *Migrator) cutOverTwoStep() (err error) { defer atomic.StoreInt64(&mgtr.migrationContext.InCutOverCriticalSectionFlag, 0) atomic.StoreInt64(&mgtr.migrationContext.AllEventsUpToLockProcessedInjectedFlag, 0) + phaseStartTime := time.Now() if err := mgtr.retryOperation(mgtr.applier.LockOriginalTable); err != nil { + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseOriginalTableLock, time.Since(phaseStartTime), err) return err } + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseOriginalTableLock, time.Since(phaseStartTime), nil) if err := mgtr.retryOperation(mgtr.waitForEventsUpToLock); err != nil { return err @@ -972,12 +997,19 @@ func (mgtr *Migrator) cutOverTwoStep() (err error) { return err } } + phaseStartTime = time.Now() if err := mgtr.retryOperation(mgtr.applier.SwapTablesQuickAndBumpy); err != nil { + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseMagicRename, time.Since(phaseStartTime), err) return err } + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseMagicRename, time.Since(phaseStartTime), nil) + + phaseStartTime = time.Now() if err := mgtr.retryOperation(mgtr.applier.UnlockTables); err != nil { + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseUnlock, time.Since(phaseStartTime), err) return err } + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseUnlock, time.Since(phaseStartTime), nil) lockAndRenameDuration := mgtr.migrationContext.RenameTablesEndTime.Sub(mgtr.migrationContext.LockTablesStartTime) renameDuration := mgtr.migrationContext.RenameTablesEndTime.Sub(mgtr.migrationContext.RenameTablesStartTime) @@ -1001,14 +1033,17 @@ func (mgtr *Migrator) atomicCutOver() (err error) { tableLocked := make(chan error, 2) tableUnlocked := make(chan error, 2) var renameLockSessionId int64 + phaseStartTime := time.Now() go func() { if err := mgtr.applier.AtomicCutOverMagicLock(lockOriginalSessionIdChan, tableLocked, okToUnlockTable, tableUnlocked, &renameLockSessionId); err != nil { mgtr.migrationContext.Log.Errore(err) } }() if err := <-tableLocked; err != nil { + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseMagicLock, time.Since(phaseStartTime), err) return mgtr.migrationContext.Log.Errore(err) } + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseMagicLock, time.Since(phaseStartTime), nil) lockOriginalSessionId := <-lockOriginalSessionIdChan mgtr.migrationContext.Log.Infof("Session locking original & magic tables is %+v", lockOriginalSessionId) // At this point we know the original table is locked. @@ -1026,7 +1061,8 @@ func (mgtr *Migrator) atomicCutOver() (err error) { // Step 2 // We now attempt an atomic RENAME on original & ghost tables, and expect it to block. - mgtr.migrationContext.RenameTablesStartTime = time.Now() + phaseStartTime = time.Now() + mgtr.migrationContext.RenameTablesStartTime = phaseStartTime var tableRenameKnownToHaveFailed int64 renameSessionIdChan := make(chan int64, 2) @@ -1051,6 +1087,7 @@ func (mgtr *Migrator) atomicCutOver() (err error) { } // Wait for the RENAME to appear in PROCESSLIST if err := mgtr.retryOperation(waitForRename, true); err != nil { + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseMagicRename, time.Since(phaseStartTime), err) // Abort! Release the lock okToUnlockTable <- true return err @@ -1059,21 +1096,27 @@ func (mgtr *Migrator) atomicCutOver() (err error) { mgtr.migrationContext.Log.Infof("Found atomic RENAME to be blocking, as expected. Double checking the lock is still in place (though I don't strictly have to)") } if err := mgtr.applier.ExpectUsedLock(lockOriginalSessionId); err != nil { + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseMagicRename, time.Since(phaseStartTime), err) // Abort operation. Just make sure to drop the magic table. return mgtr.migrationContext.Log.Errore(err) } mgtr.migrationContext.Log.Infof("Connection holding lock on original table still exists") + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseMagicRename, time.Since(phaseStartTime), nil) + // Now that we've found the RENAME blocking, AND the locking connection still alive, // we know it is safe to proceed to release the lock renameLockSessionId = renameSessionId + unlockStartTime := time.Now() okToUnlockTable <- true // BAM! magic table dropped, original table lock is released // -> RENAME released -> queries on original are unblocked. if err := <-tableUnlocked; err != nil { + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseUnlock, time.Since(unlockStartTime), err) return mgtr.migrationContext.Log.Errore(err) } + metrics.RecordCutOverPhase(mgtr.migrationContext.Metrics, metrics.CutOverPhaseUnlock, time.Since(unlockStartTime), nil) if err := <-tablesRenamed; err != nil { return mgtr.migrationContext.Log.Errore(err) } @@ -1168,7 +1211,7 @@ func (mgtr *Migrator) initiateInspector() (err error) { return nil } -// emitProgressMetrics emits StatsD gauges from a progress snapshot. +// emitProgressMetrics emits StatsD gauges func (mgtr *Migrator) emitProgressMetrics(snap migrationProgressSnapshot) { metrics.EmitProgressGauges( mgtr.migrationContext.Metrics, @@ -1181,6 +1224,13 @@ func (mgtr *Migrator) emitProgressMetrics(snap migrationProgressSnapshot) { snap.applyEventsBacklog, snap.applyEventsCapacity, ) + isThrottled, _, _ := mgtr.migrationContext.IsThrottled() + metrics.EmitLagGauges( + mgtr.migrationContext.Metrics, + snap.replicationLagSeconds, + snap.heartbeatLagSeconds, + isThrottled, + ) } // reportStatus samples progress, emits metrics, and optionally prints status output. diff --git a/go/logic/migrator_test.go b/go/logic/migrator_test.go index 31e161b5e..a8c555220 100644 --- a/go/logic/migrator_test.go +++ b/go/logic/migrator_test.go @@ -29,6 +29,7 @@ import ( "github.com/github/gh-ost/go/base" "github.com/github/gh-ost/go/binlog" + "github.com/github/gh-ost/go/metrics" "github.com/github/gh-ost/go/mysql" "github.com/github/gh-ost/go/sql" "github.com/testcontainers/testcontainers-go" @@ -410,11 +411,92 @@ func TestMigratorGetMigrationStateAndETA(t *testing.T) { type progressGaugeSpy struct { names []string values []float64 + tags [][]string } -func (s *progressGaugeSpy) Gauge(name string, value float64, _ ...string) { +func (s *progressGaugeSpy) Gauge(name string, value float64, tags ...string) { s.names = append(s.names, name) s.values = append(s.values, value) + s.tags = append(s.tags, append([]string(nil), tags...)) +} + +func (s *progressGaugeSpy) Count(name string, value int64, tags ...string) { +} + +func (s *progressGaugeSpy) Histogram(name string, value float64, tags ...string) { +} + +type cutOverMetricsSpy struct { + countNames []string + countTags [][]string + histogramNames []string + histogramTags [][]string +} + +func (s *cutOverMetricsSpy) Gauge(_ string, _ float64, _ ...string) {} + +func (s *cutOverMetricsSpy) Count(name string, _ int64, tags ...string) { + s.countNames = append(s.countNames, name) + s.countTags = append(s.countTags, append([]string(nil), tags...)) +} + +func (s *cutOverMetricsSpy) Histogram(name string, _ float64, tags ...string) { + s.histogramNames = append(s.histogramNames, name) + s.histogramTags = append(s.histogramTags, append([]string(nil), tags...)) +} + +func TestCutOverOperationWithMetricsRetryThenSuccess(t *testing.T) { + spy := &cutOverMetricsSpy{} + ctx := base.NewMigrationContext() + ctx.Metrics = spy + migrator := NewMigrator(ctx, "test") + + attempts := 0 + retrier := func(operation func() error, _ ...bool) error { + for { + err := operation() + if err == nil { + return nil + } + if attempts >= 2 { + return err + } + } + } + operation := func() error { + attempts++ + if attempts == 1 { + return errors.New("transient cutover failure") + } + return nil + } + + require.NoError(t, migrator.cutOverOperationWithMetrics(retrier, operation)) + assert.Equal(t, []string{"cut_over.attempts_total", "cut_over.attempts_total"}, spy.countNames) + assert.Equal(t, [][]string{{"outcome:" + metrics.CutOverOutcomeRetry}, {"outcome:" + metrics.CutOverOutcomeSuccess}}, spy.countTags) + assert.Equal(t, []string{"cut_over.total_duration_milliseconds"}, spy.histogramNames) + assert.Equal(t, [][]string{{"outcome:" + metrics.CutOverOutcomeSuccess}}, spy.histogramTags) +} + +func TestCutOverOperationWithMetricsAbort(t *testing.T) { + spy := &cutOverMetricsSpy{} + ctx := base.NewMigrationContext() + ctx.Metrics = spy + migrator := NewMigrator(ctx, "test") + cutOverErr := errors.New("cutover failed") + + retrier := func(operation func() error, _ ...bool) error { + return operation() + } + operation := func() error { + return cutOverErr + } + + require.ErrorIs(t, migrator.cutOverOperationWithMetrics(retrier, operation), cutOverErr) + assert.Equal(t, []string{"cut_over.attempts_total", "cut_over.attempts_total"}, spy.countNames) + assert.Equal(t, [][]string{{"outcome:" + metrics.CutOverOutcomeRetry}, {"outcome:" + metrics.CutOverOutcomeAbort}}, spy.countTags) + assert.Equal(t, []string{"cut_over.total_duration_milliseconds"}, spy.histogramNames) + assert.Equal(t, [][]string{{"outcome:" + metrics.CutOverOutcomeAbort}}, spy.histogramTags) } func TestReportStatusEmitsProgressGaugesEveryTick(t *testing.T) { @@ -427,7 +509,7 @@ func TestReportStatusEmitsProgressGaugesEveryTick(t *testing.T) { migrator := NewMigrator(ctx, "test") migrator.reportStatus(NoPrintStatusRule, io.Discard) - require.Len(t, spy.names, 6) + require.Len(t, spy.names, 8) assert.Equal(t, []string{ "row_copy.rows_copied", "row_copy.rows_estimate", @@ -435,16 +517,18 @@ func TestReportStatusEmitsProgressGaugesEveryTick(t *testing.T) { "binlog.backlog_size", "binlog.backlog_capacity", "binlog.backlog_utilization", + "lag.replication_seconds", + "lag.heartbeat_seconds", }, spy.names) - assert.Equal(t, []float64{1000, 5000, 42, 0, float64(cap(migrator.applyEventsQueue)), 0}, spy.values) + assert.Equal(t, []float64{1000, 5000, 42, 0, float64(cap(migrator.applyEventsQueue)), 0}, spy.values[:6]) assert.InDelta(t, 20.0, ctx.GetProgressPct(), 0.01) spy.names = nil spy.values = nil atomic.StoreInt64(&ctx.TotalDMLEventsApplied, 100) migrator.reportStatus(NoPrintStatusRule, io.Discard) - require.Len(t, spy.names, 6) - assert.Equal(t, []float64{1000, 5000, 100, 0, float64(cap(migrator.applyEventsQueue)), 0}, spy.values) + require.Len(t, spy.names, 8) + assert.Equal(t, []float64{1000, 5000, 100, 0, float64(cap(migrator.applyEventsQueue)), 0}, spy.values[:6]) } func TestReportStatusEmitsBinlogBacklogGauges(t *testing.T) { @@ -463,7 +547,7 @@ func TestReportStatusEmitsBinlogBacklogGauges(t *testing.T) { migrator.reportStatus(NoPrintStatusRule, io.Discard) capacity := float64(cap(migrator.applyEventsQueue)) - require.Len(t, spy.names, 6) + require.Len(t, spy.names, 8) assert.Equal(t, float64(2), spy.values[3]) assert.Equal(t, capacity, spy.values[4]) assert.InDelta(t, 2/capacity, spy.values[5], 1e-9) @@ -480,7 +564,7 @@ func TestReportStatusEmitsGaugesWhenRowCopyComplete(t *testing.T) { atomic.StoreInt64(&migrator.rowCopyCompleteFlag, 1) migrator.reportStatus(NoPrintStatusRule, io.Discard) - require.Len(t, spy.names, 6) + require.Len(t, spy.names, 8) assert.Equal(t, float64(5000), spy.values[0]) assert.Equal(t, float64(5000), spy.values[1], "rows_estimate tracks rows_copied when row copy is complete") } @@ -500,8 +584,27 @@ func TestReportStatusEmitsGaugesWhenPrintSuppressed(t *testing.T) { require.False(t, migrator.shouldPrintStatus(HeuristicPrintStatusRule, snap.elapsedSeconds, etaDuration)) migrator.reportStatus(HeuristicPrintStatusRule, io.Discard) - require.Len(t, spy.names, 6) - assert.Equal(t, []float64{1000, 5000, 0, 0, float64(cap(migrator.applyEventsQueue)), 0}, spy.values) + require.Len(t, spy.names, 8) + assert.Equal(t, []float64{1000, 5000, 0, 0, float64(cap(migrator.applyEventsQueue)), 0}, spy.values[:6]) +} + +func TestReportStatusEmitsLagGaugesWhenThrottled(t *testing.T) { + spy := &progressGaugeSpy{} + ctx := base.NewMigrationContext() + ctx.Metrics = spy + ctx.SetThrottled(true, "max-lag-millis", base.NoThrottleReasonHint) + atomic.StoreInt64(&ctx.CurrentLag, int64(5*time.Second)) + ctx.SetLastHeartbeatOnChangelogTime(time.Now().Add(-4 * time.Second)) + + migrator := NewMigrator(ctx, "test") + migrator.reportStatus(NoPrintStatusRule, io.Discard) + + require.GreaterOrEqual(t, len(spy.names), 8) + assert.Equal(t, "lag.replication_seconds", spy.names[6]) + assert.Equal(t, "lag.heartbeat_seconds", spy.names[7]) + require.Len(t, spy.tags[6], 1) + assert.Equal(t, "throttled:true", spy.tags[6][0]) + assert.Equal(t, "throttled:true", spy.tags[7][0]) } func TestMigratorShouldPrintStatus(t *testing.T) { diff --git a/go/metrics/binlog_backlog.go b/go/metrics/binlog_backlog.go deleted file mode 100644 index 2d6ea5a4b..000000000 --- a/go/metrics/binlog_backlog.go +++ /dev/null @@ -1,31 +0,0 @@ -/* - Copyright 2026 GitHub Inc. - See https://github.com/github/gh-ost/blob/master/LICENSE -*/ - -package metrics - -// EmitBinlogBacklogGauges emits apply-events queue depth gauges (namespace is applied by the client): -// gh_ost.binlog.backlog_size, gh_ost.binlog.backlog_capacity, gh_ost.binlog.backlog_utilization. -func EmitBinlogBacklogGauges(emit MemStatsGaugeEmitter, backlogSize, backlogCapacity int) { - if emit == nil { - return - } - emit.Gauge("binlog.backlog_size", float64(backlogSize)) - emit.Gauge("binlog.backlog_capacity", float64(backlogCapacity)) - emit.Gauge("binlog.backlog_utilization", binlogBacklogUtilization(backlogSize, backlogCapacity)) -} - -func binlogBacklogUtilization(backlogSize, backlogCapacity int) float64 { - if backlogCapacity <= 0 { - return 0 - } - utilization := float64(backlogSize) / float64(backlogCapacity) - if utilization > 1 { - return 1 - } - if utilization < 0 { - return 0 - } - return utilization -} diff --git a/go/metrics/binlog_backlog_test.go b/go/metrics/binlog_backlog_test.go deleted file mode 100644 index 4f99dc1b1..000000000 --- a/go/metrics/binlog_backlog_test.go +++ /dev/null @@ -1,53 +0,0 @@ -/* - Copyright 2026 GitHub Inc. - See https://github.com/github/gh-ost/blob/master/LICENSE -*/ - -package metrics - -import "testing" - -func TestEmitBinlogBacklogGauges(t *testing.T) { - spy := &gaugeSpy{} - EmitBinlogBacklogGauges(spy, 250, 1000) - - wantNames := []string{ - "binlog.backlog_size", - "binlog.backlog_capacity", - "binlog.backlog_utilization", - } - wantVals := []float64{250, 1000, 0.25} - - if len(spy.names) != len(wantNames) { - t.Fatalf("got %d gauges, want %d", len(spy.names), len(wantNames)) - } - for i := range wantNames { - if spy.names[i] != wantNames[i] || spy.values[i] != wantVals[i] { - t.Fatalf("[%d] got %s=%v want %s=%v", i, spy.names[i], spy.values[i], wantNames[i], wantVals[i]) - } - } -} - -func TestEmitBinlogBacklogGauges_nilSafe(t *testing.T) { - EmitBinlogBacklogGauges(nil, 1, 2) -} - -func TestBinlogBacklogUtilization(t *testing.T) { - tests := []struct { - size, capacity int - want float64 - }{ - {0, 1000, 0}, - {250, 1000, 0.25}, - {1000, 1000, 1}, - {1500, 1000, 1}, - {-1, 1000, 0}, - {10, 0, 0}, - } - for _, tt := range tests { - got := binlogBacklogUtilization(tt.size, tt.capacity) - if got != tt.want { - t.Fatalf("utilization(%d, %d) = %v, want %v", tt.size, tt.capacity, got, tt.want) - } - } -} diff --git a/go/metrics/client.go b/go/metrics/client.go index ed6acc096..1d78df216 100644 --- a/go/metrics/client.go +++ b/go/metrics/client.go @@ -21,6 +21,13 @@ type Client struct { sd *statsd.Client } +// Emitter is implemented by *Client; used for tests without UDP. +type Emitter interface { + Gauge(name string, value float64, tags ...string) + Count(name string, value int64, tags ...string) + Histogram(name string, value float64, tags ...string) +} + // NewClient connects to addr for StatsD. If addr is empty, returns Noop and nil error. // namespace is typically "gh_ost." (metrics are named namespace + short name, e.g. gh_ost.startup). // tags are global tags applied to every metric (repeatable --statsd-tags). @@ -60,6 +67,13 @@ func (c *Client) Count(name string, value int64, tags ...string) { _ = c.sd.Count(name, value, tags, 1.0) } +func (c *Client) Histogram(name string, value float64, tags ...string) { + if c.sd == nil { + return + } + _ = c.sd.Histogram(name, value, tags, 1.0) +} + // Close flushes buffered metrics; safe for Noop. func (c *Client) Close() error { if c.sd == nil { diff --git a/go/metrics/emit.go b/go/metrics/emit.go new file mode 100644 index 000000000..7c359edab --- /dev/null +++ b/go/metrics/emit.go @@ -0,0 +1,170 @@ +/* + Copyright 2026 GitHub Inc. + See https://github.com/github/gh-ost/blob/master/LICENSE +*/ + +package metrics + +import ( + "context" + "fmt" + "runtime" + "time" +) + +// EmitProgressGauges emits row-copy and DML progress gauges (namespace is applied by the client): +// gh_ost.row_copy.rows_copied, gh_ost.row_copy.rows_estimate, gh_ost.dml.events_applied. +func EmitProgressGauges(emit Emitter, rowsCopied, rowsEstimate, dmlEventsApplied int64) { + if emit == nil { + return + } + emit.Gauge("row_copy.rows_copied", float64(rowsCopied)) + emit.Gauge("row_copy.rows_estimate", float64(rowsEstimate)) + emit.Gauge("dml.events_applied", float64(dmlEventsApplied)) +} + +// EmitBinlogBacklogGauges emits apply-events queue depth gauges (namespace is applied by the client): +// gh_ost.binlog.backlog_size, gh_ost.binlog.backlog_capacity, gh_ost.binlog.backlog_utilization. +func EmitBinlogBacklogGauges(emit Emitter, backlogSize, backlogCapacity int) { + if emit == nil { + return + } + emit.Gauge("binlog.backlog_size", float64(backlogSize)) + emit.Gauge("binlog.backlog_capacity", float64(backlogCapacity)) + emit.Gauge("binlog.backlog_utilization", binlogBacklogUtilization(backlogSize, backlogCapacity)) +} + +func binlogBacklogUtilization(backlogSize, backlogCapacity int) float64 { + if backlogCapacity <= 0 { + return 0 + } + utilization := float64(backlogSize) / float64(backlogCapacity) + if utilization > 1 { + return 1 + } + if utilization < 0 { + return 0 + } + return utilization +} + +// EmitLagGauges emits replication and heartbeat lag gauges (namespace is applied by the client): +// gh_ost.lag.replication_seconds, gh_ost.lag.heartbeat_seconds, each tagged throttled:true|false. +// +// These are point-in-time readings each status tick (not a distribution), so gauges are used +// rather than histograms; DogStatsD histogram aggregation exposes count/max series that do not +// match the log line lag values in Prometheus/Grafana. +func EmitLagGauges(emit Emitter, replicationLagSeconds, heartbeatLagSeconds float64, throttled bool) { + if emit == nil { + return + } + tags := []string{fmt.Sprintf("throttled:%t", throttled)} + emit.Gauge("lag.replication_seconds", replicationLagSeconds, tags...) + emit.Gauge("lag.heartbeat_seconds", heartbeatLagSeconds, tags...) +} + +// EmitGoRuntimeGauges emits gh_ost.go_runtime.* gauges (namespace is applied by the client). +// m and numGoroutine are typically from runtime.ReadMemStats and runtime.NumGoroutine. +func EmitGoRuntimeGauges(emit Emitter, m *runtime.MemStats, numGoroutine int) { + if emit == nil || m == nil { + return + } + emit.Gauge("go_runtime.alloc_bytes", float64(m.Alloc)) + emit.Gauge("go_runtime.sys_bytes", float64(m.Sys)) + emit.Gauge("go_runtime.heap_inuse_bytes", float64(m.HeapInuse)) + emit.Gauge("go_runtime.num_gc", float64(m.NumGC)) + emit.Gauge("go_runtime.gc_pause_total_ns", float64(m.PauseTotalNs)) + emit.Gauge("go_runtime.goroutines", float64(numGoroutine)) +} + +// StartGoRuntimeReporter periodically samples runtime memory and goroutines and emits gauges +// until ctx is cancelled. It is a no-op when interval <= 0, client is nil, or StatsD is disabled +// (noop client). +func StartGoRuntimeReporter(ctx context.Context, client *Client, interval time.Duration) { + if ctx == nil || client == nil || interval <= 0 || client.sd == nil { + return + } + + emit := func() { + var m runtime.MemStats + runtime.ReadMemStats(&m) + EmitGoRuntimeGauges(client, &m, runtime.NumGoroutine()) + } + + go func() { + ticker := time.NewTicker(interval) + defer ticker.Stop() + + emit() + for { + select { + case <-ctx.Done(): + return + case <-ticker.C: + emit() + } + } + }() +} + +const ( + CutOverOutcomeSuccess = "success" + CutOverOutcomeRetry = "retry" + CutOverOutcomeAbort = "abort" + + CutOverPhaseMagicLock = "magic_lock" + CutOverPhaseOriginalTableLock = "original_table_lock" + CutOverPhaseMagicRename = "magic_rename" + CutOverPhaseUnlock = "unlock" +) + +type cutOverHistogramEmitter interface { + Histogram(name string, value float64, tags ...string) +} + +type cutOverCountEmitter interface { + Count(name string, value int64, tags ...string) +} + +// RecordCutOverPhase emits gh_ost.cut_over.phase_duration_milliseconds. +func RecordCutOverPhase(emit any, phase string, duration time.Duration, err error) { + if emit == nil || phase == "" || duration < 0 { + return + } + histogramEmitter, ok := emit.(cutOverHistogramEmitter) + if !ok { + return + } + histogramEmitter.Histogram("cut_over.phase_duration_milliseconds", float64(duration.Milliseconds()), "phase:"+phase, "outcome:"+cutOverOutcomeFromError(err)) +} + +// RecordCutOverAttempt emits gh_ost.cut_over.attempts_total. +func RecordCutOverAttempt(emit any, outcome string) { + if emit == nil || outcome == "" { + return + } + countEmitter, ok := emit.(cutOverCountEmitter) + if !ok { + return + } + countEmitter.Count("cut_over.attempts_total", 1, "outcome:"+outcome) +} + +// RecordCutOverTotal emits gh_ost.cut_over.total_duration_milliseconds for terminal cut-over outcomes. +func RecordCutOverTotal(emit any, duration time.Duration, err error) { + if emit == nil || duration < 0 { + return + } + histogramEmitter, ok := emit.(cutOverHistogramEmitter) + if !ok { + return + } + histogramEmitter.Histogram("cut_over.total_duration_milliseconds", float64(duration.Milliseconds()), "outcome:"+cutOverOutcomeFromError(err)) +} + +func cutOverOutcomeFromError(err error) string { + if err != nil { + return CutOverOutcomeAbort + } + return CutOverOutcomeSuccess +} diff --git a/go/metrics/emit_test.go b/go/metrics/emit_test.go new file mode 100644 index 000000000..835b71b81 --- /dev/null +++ b/go/metrics/emit_test.go @@ -0,0 +1,243 @@ +/* + Copyright 2026 GitHub Inc. + See https://github.com/github/gh-ost/blob/master/LICENSE +*/ + +package metrics + +import ( + "context" + "errors" + "runtime" + "slices" + "testing" + "time" +) + +type gaugeSpy struct { + names []string + values []float64 + tags [][]string +} + +func (g *gaugeSpy) Gauge(name string, value float64, tags ...string) { + g.names = append(g.names, name) + g.values = append(g.values, value) + g.tags = append(g.tags, append([]string(nil), tags...)) +} + +func (g *gaugeSpy) Count(name string, value int64, tags ...string) { +} + +func (g *gaugeSpy) Histogram(name string, value float64, tags ...string) { +} + +func TestEmitProgressGauges(t *testing.T) { + spy := &gaugeSpy{} + EmitProgressGauges(spy, 1000, 5000, 42) + + wantNames := []string{ + "row_copy.rows_copied", + "row_copy.rows_estimate", + "dml.events_applied", + } + wantVals := []float64{1000, 5000, 42} + + if len(spy.names) != len(wantNames) { + t.Fatalf("got %d gauges, want %d", len(spy.names), len(wantNames)) + } + for i := range wantNames { + if spy.names[i] != wantNames[i] || spy.values[i] != wantVals[i] { + t.Fatalf("[%d] got %s=%v want %s=%v", i, spy.names[i], spy.values[i], wantNames[i], wantVals[i]) + } + } +} + +func TestEmitProgressGauges_nilSafe(t *testing.T) { + EmitProgressGauges(nil, 1, 2, 3) +} + +func TestEmitBinlogBacklogGauges(t *testing.T) { + spy := &gaugeSpy{} + EmitBinlogBacklogGauges(spy, 250, 1000) + + wantNames := []string{ + "binlog.backlog_size", + "binlog.backlog_capacity", + "binlog.backlog_utilization", + } + wantVals := []float64{250, 1000, 0.25} + + if len(spy.names) != len(wantNames) { + t.Fatalf("got %d gauges, want %d", len(spy.names), len(wantNames)) + } + for i := range wantNames { + if spy.names[i] != wantNames[i] || spy.values[i] != wantVals[i] { + t.Fatalf("[%d] got %s=%v want %s=%v", i, spy.names[i], spy.values[i], wantNames[i], wantVals[i]) + } + } +} + +func TestEmitBinlogBacklogGauges_nilSafe(t *testing.T) { + EmitBinlogBacklogGauges(nil, 1, 2) +} + +func TestBinlogBacklogUtilization(t *testing.T) { + tests := []struct { + size, capacity int + want float64 + }{ + {0, 1000, 0}, + {250, 1000, 0.25}, + {1000, 1000, 1}, + {1500, 1000, 1}, + {-1, 1000, 0}, + {10, 0, 0}, + } + for _, tt := range tests { + got := binlogBacklogUtilization(tt.size, tt.capacity) + if got != tt.want { + t.Fatalf("utilization(%d, %d) = %v, want %v", tt.size, tt.capacity, got, tt.want) + } + } +} + +func TestEmitLagGauges_notThrottled(t *testing.T) { + spy := &gaugeSpy{} + EmitLagGauges(spy, 2.5, 1.25, false) + + wantNames := []string{"lag.replication_seconds", "lag.heartbeat_seconds"} + wantVals := []float64{2.5, 1.25} + wantTags := []string{"throttled:false"} + + if len(spy.names) != len(wantNames) { + t.Fatalf("got %d gauges, want %d", len(spy.names), len(wantNames)) + } + for i := range wantNames { + if spy.names[i] != wantNames[i] || spy.values[i] != wantVals[i] { + t.Fatalf("[%d] got %s=%v want %s=%v", i, spy.names[i], spy.values[i], wantNames[i], wantVals[i]) + } + if len(spy.tags[i]) != 1 || spy.tags[i][0] != wantTags[0] { + t.Fatalf("[%d] got tags %v want [%s]", i, spy.tags[i], wantTags[0]) + } + } +} + +func TestEmitLagGauges_throttled(t *testing.T) { + spy := &gaugeSpy{} + EmitLagGauges(spy, 4.0, 3.0, true) + + if len(spy.names) != 2 { + t.Fatalf("got %d gauges, want 2", len(spy.names)) + } + for i := range spy.names { + if len(spy.tags[i]) != 1 || spy.tags[i][0] != "throttled:true" { + t.Fatalf("[%d] got tags %v want [throttled:true]", i, spy.tags[i]) + } + } +} + +func TestEmitLagGauges_nilSafe(t *testing.T) { + EmitLagGauges(nil, 1, 2, false) +} + +func TestEmitGoRuntimeGauges(t *testing.T) { + spy := &gaugeSpy{} + m := &runtime.MemStats{ + Alloc: 100, + Sys: 200, + HeapInuse: 300, + NumGC: 7, + PauseTotalNs: 42, + } + EmitGoRuntimeGauges(spy, m, 123) + + wantNames := []string{ + "go_runtime.alloc_bytes", + "go_runtime.sys_bytes", + "go_runtime.heap_inuse_bytes", + "go_runtime.num_gc", + "go_runtime.gc_pause_total_ns", + "go_runtime.goroutines", + } + wantVals := []float64{100, 200, 300, 7, 42, 123} + + if len(spy.names) != len(wantNames) { + t.Fatalf("got %d gauges, want %d", len(spy.names), len(wantNames)) + } + for i := range wantNames { + if spy.names[i] != wantNames[i] || spy.values[i] != wantVals[i] { + t.Fatalf("[%d] got %s=%v want %s=%v", i, spy.names[i], spy.values[i], wantNames[i], wantVals[i]) + } + } +} + +func TestEmitGoRuntimeGauges_nilSafe(t *testing.T) { + EmitGoRuntimeGauges(nil, &runtime.MemStats{}, 1) + EmitGoRuntimeGauges(&gaugeSpy{}, nil, 1) +} + +func TestStartGoRuntimeReporter_stopsOnCancel(t *testing.T) { + ctx, cancel := context.WithCancel(context.Background()) + c := &Client{} // sd nil, so the reporter should not start. + StartGoRuntimeReporter(ctx, c, time.Millisecond) + cancel() + time.Sleep(20 * time.Millisecond) +} + +type cutOverSpy struct { + histogramNames []string + histogramValues []float64 + histogramTags [][]string + countNames []string + countValues []int64 + countTags [][]string +} + +func (s *cutOverSpy) Histogram(name string, value float64, tags ...string) { + s.histogramNames = append(s.histogramNames, name) + s.histogramValues = append(s.histogramValues, value) + s.histogramTags = append(s.histogramTags, tags) +} + +func (s *cutOverSpy) Count(name string, value int64, tags ...string) { + s.countNames = append(s.countNames, name) + s.countValues = append(s.countValues, value) + s.countTags = append(s.countTags, tags) +} + +func TestRecordCutOverMetrics(t *testing.T) { + spy := &cutOverSpy{} + + RecordCutOverPhase(spy, CutOverPhaseMagicLock, 1500*time.Millisecond, nil) + RecordCutOverAttempt(spy, CutOverOutcomeSuccess) + RecordCutOverTotal(spy, 2*time.Second, errors.New("boom")) + + if len(spy.histogramNames) != 2 { + t.Fatalf("got %d histograms, want 2", len(spy.histogramNames)) + } + if spy.histogramNames[0] != "cut_over.phase_duration_milliseconds" || spy.histogramValues[0] != 1500 { + t.Fatalf("got first histogram %s=%v", spy.histogramNames[0], spy.histogramValues[0]) + } + if !slices.Equal(spy.histogramTags[0], []string{"phase:magic_lock", "outcome:success"}) { + t.Fatalf("got phase tags %#v", spy.histogramTags[0]) + } + if spy.histogramNames[1] != "cut_over.total_duration_milliseconds" || spy.histogramValues[1] != 2000 { + t.Fatalf("got second histogram %s=%v", spy.histogramNames[1], spy.histogramValues[1]) + } + if !slices.Equal(spy.histogramTags[1], []string{"outcome:abort"}) { + t.Fatalf("got total tags %#v", spy.histogramTags[1]) + } + if len(spy.countNames) != 1 || spy.countNames[0] != "cut_over.attempts_total" || spy.countValues[0] != 1 { + t.Fatalf("got counts %#v values %#v", spy.countNames, spy.countValues) + } + if !slices.Equal(spy.countTags[0], []string{"outcome:success"}) { + t.Fatalf("got count tags %#v", spy.countTags[0]) + } +} + +func TestRecordCutOverMetricsNilSafe(t *testing.T) { + RecordCutOverPhase(nil, CutOverPhaseMagicLock, time.Second, nil) + RecordCutOverAttempt(nil, CutOverOutcomeSuccess) + RecordCutOverTotal(nil, time.Second, nil) +} diff --git a/go/metrics/go_runtime.go b/go/metrics/go_runtime.go deleted file mode 100644 index 24ae2c6b5..000000000 --- a/go/metrics/go_runtime.go +++ /dev/null @@ -1,61 +0,0 @@ -/* - Copyright 2022 GitHub Inc. - See https://github.com/github/gh-ost/blob/master/LICENSE -*/ - -package metrics - -import ( - "context" - "runtime" - "time" -) - -// MemStatsGaugeEmitter is implemented by *Client; used for tests without UDP. -type MemStatsGaugeEmitter interface { - Gauge(name string, value float64, tags ...string) -} - -// EmitGoRuntimeGauges emits gh_ost.go_runtime.* gauges (namespace is applied by the client). -// m and numGoroutine are typically from runtime.ReadMemStats and runtime.NumGoroutine. -func EmitGoRuntimeGauges(emit MemStatsGaugeEmitter, m *runtime.MemStats, numGoroutine int) { - if emit == nil || m == nil { - return - } - emit.Gauge("go_runtime.alloc_bytes", float64(m.Alloc)) - emit.Gauge("go_runtime.sys_bytes", float64(m.Sys)) - emit.Gauge("go_runtime.heap_inuse_bytes", float64(m.HeapInuse)) - emit.Gauge("go_runtime.num_gc", float64(m.NumGC)) - emit.Gauge("go_runtime.gc_pause_total_ns", float64(m.PauseTotalNs)) - emit.Gauge("go_runtime.goroutines", float64(numGoroutine)) -} - -// StartGoRuntimeReporter periodically samples runtime memory and goroutines and emits gauges -// until ctx is cancelled. It is a no-op when interval <= 0, client is nil, or StatsD is disabled -// (noop client). -func StartGoRuntimeReporter(ctx context.Context, client *Client, interval time.Duration) { - if ctx == nil || client == nil || interval <= 0 || client.sd == nil { - return - } - - emit := func() { - var m runtime.MemStats - runtime.ReadMemStats(&m) - EmitGoRuntimeGauges(client, &m, runtime.NumGoroutine()) - } - - go func() { - ticker := time.NewTicker(interval) - defer ticker.Stop() - - emit() - for { - select { - case <-ctx.Done(): - return - case <-ticker.C: - emit() - } - } - }() -} diff --git a/go/metrics/go_runtime_test.go b/go/metrics/go_runtime_test.go deleted file mode 100644 index 24811206b..000000000 --- a/go/metrics/go_runtime_test.go +++ /dev/null @@ -1,67 +0,0 @@ -/* - Copyright 2022 GitHub Inc. - See https://github.com/github/gh-ost/blob/master/LICENSE -*/ - -package metrics - -import ( - "context" - "runtime" - "testing" - "time" -) - -type gaugeSpy struct { - names []string - values []float64 -} - -func (g *gaugeSpy) Gauge(name string, value float64, _ ...string) { - g.names = append(g.names, name) - g.values = append(g.values, value) -} - -func TestEmitGoRuntimeGauges(t *testing.T) { - spy := &gaugeSpy{} - m := &runtime.MemStats{ - Alloc: 100, - Sys: 200, - HeapInuse: 300, - NumGC: 7, - PauseTotalNs: 42, - } - EmitGoRuntimeGauges(spy, m, 123) - - wantNames := []string{ - "go_runtime.alloc_bytes", - "go_runtime.sys_bytes", - "go_runtime.heap_inuse_bytes", - "go_runtime.num_gc", - "go_runtime.gc_pause_total_ns", - "go_runtime.goroutines", - } - wantVals := []float64{100, 200, 300, 7, 42, 123} - - if len(spy.names) != len(wantNames) { - t.Fatalf("got %d gauges, want %d", len(spy.names), len(wantNames)) - } - for i := range wantNames { - if spy.names[i] != wantNames[i] || spy.values[i] != wantVals[i] { - t.Fatalf("[%d] got %s=%v want %s=%v", i, spy.names[i], spy.values[i], wantNames[i], wantVals[i]) - } - } -} - -func TestEmitGoRuntimeGauges_nilSafe(t *testing.T) { - EmitGoRuntimeGauges(nil, &runtime.MemStats{}, 1) - EmitGoRuntimeGauges(&gaugeSpy{}, nil, 1) -} - -func TestStartGoRuntimeReporter_stopsOnCancel(t *testing.T) { - ctx, cancel := context.WithCancel(context.Background()) - c := &Client{} // sd nil — should not start - StartGoRuntimeReporter(ctx, c, time.Millisecond) - cancel() - time.Sleep(20 * time.Millisecond) -} diff --git a/go/metrics/progress.go b/go/metrics/progress.go deleted file mode 100644 index b6b3361b6..000000000 --- a/go/metrics/progress.go +++ /dev/null @@ -1,17 +0,0 @@ -/* - Copyright 2022 GitHub Inc. - See https://github.com/github/gh-ost/blob/master/LICENSE -*/ - -package metrics - -// EmitProgressGauges emits row-copy and DML progress gauges (namespace is applied by the client): -// gh_ost.row_copy.rows_copied, gh_ost.row_copy.rows_estimate, gh_ost.dml.events_applied. -func EmitProgressGauges(emit MemStatsGaugeEmitter, rowsCopied, rowsEstimate, dmlEventsApplied int64) { - if emit == nil { - return - } - emit.Gauge("row_copy.rows_copied", float64(rowsCopied)) - emit.Gauge("row_copy.rows_estimate", float64(rowsEstimate)) - emit.Gauge("dml.events_applied", float64(dmlEventsApplied)) -} diff --git a/go/metrics/progress_test.go b/go/metrics/progress_test.go deleted file mode 100644 index 4a6ce5d3b..000000000 --- a/go/metrics/progress_test.go +++ /dev/null @@ -1,33 +0,0 @@ -/* - Copyright 2022 GitHub Inc. - See https://github.com/github/gh-ost/blob/master/LICENSE -*/ - -package metrics - -import "testing" - -func TestEmitProgressGauges(t *testing.T) { - spy := &gaugeSpy{} - EmitProgressGauges(spy, 1000, 5000, 42) - - wantNames := []string{ - "row_copy.rows_copied", - "row_copy.rows_estimate", - "dml.events_applied", - } - wantVals := []float64{1000, 5000, 42} - - if len(spy.names) != len(wantNames) { - t.Fatalf("got %d gauges, want %d", len(spy.names), len(wantNames)) - } - for i := range wantNames { - if spy.names[i] != wantNames[i] || spy.values[i] != wantVals[i] { - t.Fatalf("[%d] got %s=%v want %s=%v", i, spy.names[i], spy.values[i], wantNames[i], wantVals[i]) - } - } -} - -func TestEmitProgressGauges_nilSafe(t *testing.T) { - EmitProgressGauges(nil, 1, 2, 3) -}