feat: add memory usage health check

RELEASES.md

@@ -22,6 +22,8 @@ The plugin version is unchanged at `44` and is compatible with version `v1.14.0`
 - Added:
   - `--system-tracker-disk-required-available-space-percentage`
   - `--system-tracker-disk-warning-available-space-percentage`
+  - `--system-tracker-memory-warning-available-percentage` - Monitors available system memory and reports unhealthy when below threshold. Disabled by default (set to 0). Supports both bare-metal and containerized (K8s/Docker) environments via cgroup detection.
+
 - Deprecated:
   - `--system-tracker-disk-required-available-space`
   - `--system-tracker-disk-warning-threshold-available-space`

config/config.go

@@ -55,7 +55,8 @@ const (
 	chainUpgradeFileName = "upgrade"
 	subnetConfigFileExt  = ".json"
 
-	maxDiskSpaceThreshold = 50
+	maxDiskSpaceThreshold   = 50
+	maxMemorySpaceThreshold = 50
 )
 
 var (
@@ -88,6 +89,7 @@ var (
 	errInvalidSignerConfig                    = fmt.Errorf("only one of the following flags can be set: %s, %s, %s, %s", StakingEphemeralSignerEnabledKey, StakingSignerKeyContentKey, StakingSignerKeyPathKey, StakingRPCSignerEndpointKey)
 	errDiskSpaceOutOfRange                    = fmt.Errorf("out of range [0,%d]", maxDiskSpaceThreshold)
 	errDiskWarnAfterFatal                     = errors.New("warning disk space threshold cannot be greater than fatal threshold")
+	errMemorySpaceOutOfRange                  = fmt.Errorf("out of range [0,%d]", maxMemorySpaceThreshold)
 )
 
 func getConsensusConfig(v *viper.Viper) snowball.Parameters {
@@ -1159,6 +1161,16 @@ func getDiskSpaceConfig(v *viper.Viper) (
 	}
 }
 
+func getMemoryConfig(v *viper.Viper) (uint64, error) {
+	warnKey := SystemTrackerWarningAvailableMemoryPercentageKey
+	warn := v.GetUint64(warnKey)
+
+	if warn > maxMemorySpaceThreshold {
+		return 0, fmt.Errorf("%w: %q (%d)", errMemorySpaceOutOfRange, warnKey, warn)
+	}
+	return warn, nil
+}
+
 func getDiskTargeterConfig(v *viper.Viper) (tracker.TargeterConfig, error) {
 	vdrAlloc := v.GetFloat64(DiskVdrAllocKey)
 	maxNonVdrUsage := v.GetFloat64(DiskMaxNonVdrUsageKey)
@@ -1416,6 +1428,11 @@ func GetNodeConfig(v *viper.Viper) (node.Config, error) {
 		return node.Config{}, err
 	}
 
+	nodeConfig.WarningAvailableMemoryPercentage, err = getMemoryConfig(v)
+	if err != nil {
+		return node.Config{}, err
+	}
+
 	nodeConfig.CPUTargeterConfig, err = getCPUTargeterConfig(v)
 	if err != nil {
 		return node.Config{}, err

config/config_test.go

@@ -690,6 +690,47 @@ func TestGetDiskSpaceConfig(t *testing.T) {
 	}
 }
 
+func TestGetMemoryConfig(t *testing.T) {
+	tests := []struct {
+		name        string
+		config      map[string]uint64
+		expectedErr error
+	}{
+		{
+			name:   "empty config",
+			config: map[string]uint64{},
+		},
+		{
+			name: "valid config",
+			config: map[string]uint64{
+				SystemTrackerWarningAvailableMemoryPercentageKey: maxMemorySpaceThreshold,
+			},
+		},
+		{
+			name: "invalid config - warning too big",
+			config: map[string]uint64{
+				SystemTrackerWarningAvailableMemoryPercentageKey: maxMemorySpaceThreshold + 1,
+			},
+			expectedErr: errMemorySpaceOutOfRange,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			require := require.New(t)
+			v := setupViperFlags()
+
+			for key, value := range tt.config {
+				v.Set(key, value)
+			}
+
+			_, err := GetNodeConfig(v)
+
+			require.ErrorIs(err, tt.expectedErr)
+		})
+	}
+}
+
 // setups config json file and writes content
 func setupConfigJSON(t *testing.T, rootPath string, value string) string {
 	configFilePath := filepath.Join(rootPath, "config.json")

config/flags.go

@@ -361,6 +361,7 @@ func addNodeFlags(fs *pflag.FlagSet) {
 	fs.Uint64(SystemTrackerRequiredAvailableDiskSpacePercentageKey, 3, "Minimum percentage (between 0 and 50) of available disk space, under which the node will shutdown.")
 	fs.Uint64(SystemTrackerWarningThresholdAvailableDiskSpaceKey, 0, fmt.Sprintf("DEPRECATED: Warning threshold for the number of available bytes on disk, under which the node will be considered unhealthy.  Must be >= [%s]", SystemTrackerRequiredAvailableDiskSpaceKey))
 	fs.Uint64(SystemTrackerWarningAvailableDiskSpacePercentageKey, 10, fmt.Sprintf("Warning threshold for the percentage (between 0 and 50) of available disk space, under which the node will be considered unhealthy. Must be >= [%s]", SystemTrackerRequiredAvailableDiskSpacePercentageKey))
+	fs.Uint64(SystemTrackerWarningAvailableMemoryPercentageKey, 0, "Warning threshold percentage (between 0 and 50) for available system memory. When available memory drops below this, the node will report unhealthy. Set to 0 to disable, disabled by default.")
 
 	// CPU management
 	fs.Float64(CPUVdrAllocKey, float64(runtime.NumCPU()), "Maximum number of CPUs to allocate for use by validators. Value should be in range [0, total core count]")

config/keys.go

@@ -195,6 +195,7 @@ const (
 	SystemTrackerRequiredAvailableDiskSpacePercentageKey = "system-tracker-disk-required-available-space-percentage"
 	SystemTrackerWarningThresholdAvailableDiskSpaceKey   = "system-tracker-disk-warning-threshold-available-space"
 	SystemTrackerWarningAvailableDiskSpacePercentageKey  = "system-tracker-disk-warning-available-space-percentage"
+	SystemTrackerWarningAvailableMemoryPercentageKey     = "system-tracker-memory-warning-available-percentage"
 	DiskVdrAllocKey                                      = "throttler-inbound-disk-validator-alloc"
 	DiskMaxNonVdrUsageKey                                = "throttler-inbound-disk-max-non-validator-usage"
 	DiskMaxNonVdrNodeUsageKey                            = "throttler-inbound-disk-max-non-validator-node-usage"

config/node/config.go

@@ -215,6 +215,8 @@ type Config struct {
 	RequiredAvailableDiskSpacePercentage uint64 `json:"requiredAvailableDiskSpacePercentage"`
 	WarningAvailableDiskSpacePercentage  uint64 `json:"warningAvailableDiskSpacePercentage"`
 
+	WarningAvailableMemoryPercentage uint64 `json:"warningAvailableMemoryPercentage"`
+
 	TraceConfig trace.Config `json:"traceConfig"`
 
 	// See comment on [UseCurrentHeight] in platformvm.Config

node/node.go

@@ -1485,6 +1485,28 @@ func (n *Node) initHealthAPI() error {
 		return fmt.Errorf("couldn't register resource health check: %w", err)
 	}
 
+	memoryCheck := health.CheckerFunc(func(context.Context) (interface{}, error) {
+		// Monitor available memory and report unhealthy if below threshold.
+
+		availableMemoryBytes := n.resourceTracker.MemoryTracker().AvailableMemoryBytes()
+		availableMemoryPercentage := n.resourceTracker.MemoryTracker().AvailableMemoryPercentage()
+
+		var err error
+		if n.Config.WarningAvailableMemoryPercentage > 0 && availableMemoryPercentage < n.Config.WarningAvailableMemoryPercentage {
+			err = fmt.Errorf("remaining available memory percentage (%d%%) is below warning threshold (%d%%)", availableMemoryPercentage, n.Config.WarningAvailableMemoryPercentage)
+		}
+
+		return map[string]interface{}{
+			"availableMemoryBytes":      availableMemoryBytes,
+			"availableMemoryPercentage": availableMemoryPercentage,
+		}, err
+	})
+
+	err = n.health.RegisterHealthCheck("memory", memoryCheck, health.ApplicationTag)
+	if err != nil {
+		return fmt.Errorf("couldn't register memory health check: %w", err)
+	}
+
 	wrongBLSKeyCheck := health.CheckerFunc(func(context.Context) (interface{}, error) {
 		vdr, ok := n.vdrs.GetValidator(constants.PrimaryNetworkID, n.ID)
 		if !ok {

snow/networking/tracker/resource_tracker.go

@@ -39,10 +39,16 @@ type DiskTracker interface {
 	AvailableDiskPercentage() uint64
 }
 
+type MemoryTracker interface {
+	AvailableMemoryBytes() uint64
+	AvailableMemoryPercentage() uint64
+}
+
 // ResourceTracker is an interface for tracking peers' usage of resources
 type ResourceTracker interface {
 	CPUTracker() Tracker
 	DiskTracker() DiskTracker
+	MemoryTracker() MemoryTracker
 	// Registers that the given node started processing at the given time.
 	StartProcessing(ids.NodeID, time.Time)
 	// Registers that the given node stopped processing at the given time.
@@ -198,6 +204,30 @@ func (t *diskResourceTracker) TimeUntilUsage(nodeID ids.NodeID, now time.Time, v
 	return m.TimeUntil(now, value/scale)
 }
 
+type memoryResourceTracker struct {
+	t *resourceTracker
+}
+
+func (t *memoryResourceTracker) AvailableMemoryBytes() uint64 {
+	rt := t.t
+	rt.lock.Lock()
+	defer rt.lock.Unlock()
+
+	bytesAvailable := rt.resources.AvailableMemoryBytes()
+	rt.metrics.memoryAvailable.Set(float64(bytesAvailable))
+	return bytesAvailable
+}
+
+func (t *memoryResourceTracker) AvailableMemoryPercentage() uint64 {
+	rt := t.t
+	rt.lock.Lock()
+	defer rt.lock.Unlock()
+
+	percentageAvailable := rt.resources.AvailableMemoryPercentage()
+	rt.metrics.memoryPercentageAvailable.Set(float64(percentageAvailable))
+	return percentageAvailable
+}
+
 type resourceTracker struct {
 	lock sync.RWMutex
 
@@ -244,6 +274,10 @@ func (rt *resourceTracker) DiskTracker() DiskTracker {
 	return &diskResourceTracker{t: rt}
 }
 
+func (rt *resourceTracker) MemoryTracker() MemoryTracker {
+	return &memoryResourceTracker{t: rt}
+}
+
 func (rt *resourceTracker) StartProcessing(nodeID ids.NodeID, now time.Time) {
 	rt.lock.Lock()
 	defer rt.lock.Unlock()
@@ -297,12 +331,14 @@ func (rt *resourceTracker) prune(now time.Time) {
 }
 
 type trackerMetrics struct {
-	processingTimeMetric    prometheus.Gauge
-	cpuMetric               prometheus.Gauge
-	diskReadsMetric         prometheus.Gauge
-	diskWritesMetric        prometheus.Gauge
-	diskSpaceAvailable      prometheus.Gauge
-	diskPercentageAvailable prometheus.Gauge
+	processingTimeMetric      prometheus.Gauge
+	cpuMetric                 prometheus.Gauge
+	diskReadsMetric           prometheus.Gauge
+	diskWritesMetric          prometheus.Gauge
+	diskSpaceAvailable        prometheus.Gauge
+	diskPercentageAvailable   prometheus.Gauge
+	memoryAvailable           prometheus.Gauge
+	memoryPercentageAvailable prometheus.Gauge
 }
 
 func newCPUTrackerMetrics(reg prometheus.Registerer) (*trackerMetrics, error) {
@@ -324,13 +360,21 @@ func newCPUTrackerMetrics(reg prometheus.Registerer) (*trackerMetrics, error) {
 			Help: "Disk writes (bytes/sec) tracked by the resource manager",
 		}),
 		diskSpaceAvailable: prometheus.NewGauge(prometheus.GaugeOpts{
-			Name: "disk_available_space",
+			Name: "disk_available",
 			Help: "Available space remaining (bytes) on the database volume",
 		}),
 		diskPercentageAvailable: prometheus.NewGauge(prometheus.GaugeOpts{
-			Name: "disk_available_percentage",
+			Name: "disk_percentage_available",
 			Help: "Percentage of database volume available",
 		}),
+		memoryAvailable: prometheus.NewGauge(prometheus.GaugeOpts{
+			Name: "memory_available",
+			Help: "Available memory remaining (bytes) on the system",
+		}),
+		memoryPercentageAvailable: prometheus.NewGauge(prometheus.GaugeOpts{
+			Name: "memory_percentage_available",
+			Help: "Percentage of system memory available",
+		}),
 	}
 	err := errors.Join(
 		reg.Register(m.processingTimeMetric),
@@ -339,6 +383,8 @@ func newCPUTrackerMetrics(reg prometheus.Registerer) (*trackerMetrics, error) {
 		reg.Register(m.diskWritesMetric),
 		reg.Register(m.diskSpaceAvailable),
 		reg.Register(m.diskPercentageAvailable),
+		reg.Register(m.memoryAvailable),
+		reg.Register(m.memoryPercentageAvailable),
 	)
 	return m, err
 }

tests/fixture/tmpnet/defaults.go

@@ -43,10 +43,11 @@ func DefaultTmpnetFlags() FlagsMap {
 		config.HealthCheckFreqKey:               "2s",
 		config.AdminAPIEnabledKey:               "true",
 		config.IndexEnabledKey:                  "true",
-		// Disable disk checks by default since temporary networks often run in
-		// resource-constrained environments that commonly have low disk space.
+		// Disable disk and memory checks by default since temporary networks often run in
+		// resource-constrained environments that commonly have low disk space and memory.
 		config.SystemTrackerRequiredAvailableDiskSpacePercentageKey: "0",
 		config.SystemTrackerWarningAvailableDiskSpacePercentageKey:  "0",
+		config.SystemTrackerWarningAvailableMemoryPercentageKey:     "0",
 	}
 }
 

utils/resource/no_usage.go

@@ -25,3 +25,11 @@ func (noUsage) DiskUsage() (float64, float64) {
 func (noUsage) AvailableDiskBytes() uint64 {
 	return math.MaxUint64
 }
+
+func (noUsage) AvailableMemoryBytes() uint64 {
+	return math.MaxUint64
+}
+
+func (noUsage) AvailableMemoryPercentage() uint64 {
+	return 100
+}