Merge pull request #1886 from kube-logging/force-hotreload-grace

pepov · web-flow · commit 697833114905 · 2024-12-17T15:24:14.000+01:00
feat: force HotReload after grace period instead of blocking indefinitely
diff --git a/charts/logging-operator/charts/logging-operator-crds/templates/logging.banzaicloud.io_fluentbitagents.yaml b/charts/logging-operator/charts/logging-operator-crds/templates/logging.banzaicloud.io_fluentbitagents.yaml
@@ -1480,6 +1480,8 @@ spec:
               flush:
                 format: int32
                 type: integer
+              forceHotReloadAfterGrace:
+                type: boolean
               forwardOptions:
                 properties:
                   Require_ack_response:
diff --git a/charts/logging-operator/charts/logging-operator-crds/templates/logging.banzaicloud.io_loggings.yaml b/charts/logging-operator/charts/logging-operator-crds/templates/logging.banzaicloud.io_loggings.yaml
@@ -2325,6 +2325,8 @@ spec:
                   flush:
                     format: int32
                     type: integer
+                  forceHotReloadAfterGrace:
+                    type: boolean
                   forwardOptions:
                     properties:
                       Require_ack_response:
diff --git a/charts/logging-operator/crds/logging.banzaicloud.io_fluentbitagents.yaml b/charts/logging-operator/crds/logging.banzaicloud.io_fluentbitagents.yaml
@@ -1477,6 +1477,8 @@ spec:
               flush:
                 format: int32
                 type: integer
+              forceHotReloadAfterGrace:
+                type: boolean
               forwardOptions:
                 properties:
                   Require_ack_response:
diff --git a/charts/logging-operator/crds/logging.banzaicloud.io_loggings.yaml b/charts/logging-operator/crds/logging.banzaicloud.io_loggings.yaml
@@ -2322,6 +2322,8 @@ spec:
                   flush:
                     format: int32
                     type: integer
+                  forceHotReloadAfterGrace:
+                    type: boolean
                   forwardOptions:
                     properties:
                       Require_ack_response:
diff --git a/config/crd/bases/logging.banzaicloud.io_fluentbitagents.yaml b/config/crd/bases/logging.banzaicloud.io_fluentbitagents.yaml
@@ -1477,6 +1477,8 @@ spec:
               flush:
                 format: int32
                 type: integer
+              forceHotReloadAfterGrace:
+                type: boolean
               forwardOptions:
                 properties:
                   Require_ack_response:
diff --git a/config/crd/bases/logging.banzaicloud.io_loggings.yaml b/config/crd/bases/logging.banzaicloud.io_loggings.yaml
@@ -2322,6 +2322,8 @@ spec:
                   flush:
                     format: int32
                     type: integer
+                  forceHotReloadAfterGrace:
+                    type: boolean
                   forwardOptions:
                     properties:
                       Require_ack_response:
diff --git a/config/samples/multitenant-routing/logging/tenant-infra-logging.yaml b/config/samples/multitenant-routing/logging/tenant-infra-logging.yaml
@@ -11,7 +11,8 @@ metadata:
     tenant: infra
 spec:
   loggingRef: infra
-  fluentd: {}
+  fluentd:
+    metrics: {}
   controlNamespace: infra
 ---
 apiVersion: logging.banzaicloud.io/v1beta1
@@ -49,8 +50,12 @@ metadata:
   name: infra
 spec:
   loggingRef: infra
+  # this is required to reload even if there are pending tasks in one of the queues
+  # requires grace to be set, which is 5 by default
+  forceHotReloadAfterGrace: true
   inputTail:
     storage.type: filesystem
+    storage.pause_on_chunks_overlimit: "off"
   positiondb:
     hostPath:
       path: ""
@@ -59,7 +64,15 @@ spec:
       path: ""
   network:
     connectTimeout: 2
+    keepaliveMaxRecycle: 20
   metrics: {}
+  bufferStorage:
+    storage.max_chunks_up: 10
+  forwardOptions:
+    storage.total_limit_size: 50MB
+  image:
+    tag: 3.1.10-debug
+  configHotReload: {}
 ---
 apiVersion: logging.banzaicloud.io/v1beta1
 kind: LoggingRoute
diff --git a/docs/configuration/crds/v1beta1/fluentbit_types.md b/docs/configuration/crds/v1beta1/fluentbit_types.md
@@ -128,6 +128,11 @@ Set the flush time in seconds.nanoseconds. The engine loop uses a Flush timeout
 
 Default: 1
 
+### forceHotReloadAfterGrace (bool, optional) {#fluentbitspec-forcehotreloadaftergrace}
+
+HotReload pauses all inputs and waits until they finish. In certain situations this is unacceptable, for example if an output is down for a longer time. An undocumented option called "Hot_Reload.Ensure_Thread_Safety Off" can be used at the [SERVICE] config to force hotreload after the grace period. Please note that it might result in a SIGSEGV, but worst case kubelet will restart the container. See https://github.com/fluent/fluent-bit/pull/7509 
+
+
 ### forwardOptions (*ForwardOptions, optional) {#fluentbitspec-forwardoptions}
 
 
diff --git a/pkg/resources/fluentbit/config.go b/pkg/resources/fluentbit/config.go
@@ -26,6 +26,9 @@ var fluentBitConfigTemplate = `
 [SERVICE]
     Flush        {{ .Flush }}
     Grace        {{ .Grace }}
+    {{- if .ForceHotReloadAfterGrace }}
+    Hot_Reload.Ensure_Thread_Safety off
+    {{- end }}
     Daemon       Off
     Log_Level    {{ .LogLevel }}
     Parsers_File {{ .DefaultParsers }}
diff --git a/pkg/resources/fluentbit/configsecret.go b/pkg/resources/fluentbit/configsecret.go
@@ -64,24 +64,25 @@ type fluentBitConfig struct {
 		Port    int32
 		Path    string
 	}
-	Flush                   int32
-	Grace                   int32
-	LogLevel                string
-	EnabledIPv6             bool
-	CoroStackSize           int32
-	Output                  map[string]string
-	Input                   fluentbitInputConfig
-	Inputs                  []fluentbitInputConfigWithTenant
-	DisableKubernetesFilter bool
-	KubernetesFilter        map[string]string
-	AwsFilter               map[string]string
-	BufferStorage           map[string]string
-	FilterModify            []v1beta1.FilterModify
-	FluentForwardOutput     *fluentForwardOutputConfig
-	SyslogNGOutput          *syslogNGOutputConfig
-	DefaultParsers          string
-	CustomParsers           string
-	HealthCheck             *v1beta1.HealthCheck
+	Flush                    int32
+	Grace                    int32
+	LogLevel                 string
+	EnabledIPv6              bool
+	CoroStackSize            int32
+	Output                   map[string]string
+	ForceHotReloadAfterGrace bool
+	Input                    fluentbitInputConfig
+	Inputs                   []fluentbitInputConfigWithTenant
+	DisableKubernetesFilter  bool
+	KubernetesFilter         map[string]string
+	AwsFilter                map[string]string
+	BufferStorage            map[string]string
+	FilterModify             []v1beta1.FilterModify
+	FluentForwardOutput      *fluentForwardOutputConfig
+	SyslogNGOutput           *syslogNGOutputConfig
+	DefaultParsers           string
+	CustomParsers            string
+	HealthCheck              *v1beta1.HealthCheck
 }
 
 type fluentForwardOutputConfig struct {
@@ -214,15 +215,16 @@ func (r *Reconciler) configSecret() (runtime.Object, reconciler.DesiredState, er
 	}
 
 	input := fluentBitConfig{
-		Flush:                   r.fluentbitSpec.Flush,
-		Grace:                   r.fluentbitSpec.Grace,
-		LogLevel:                r.fluentbitSpec.LogLevel,
-		EnabledIPv6:             r.fluentbitSpec.EnabledIPv6,
-		CoroStackSize:           r.fluentbitSpec.CoroStackSize,
-		Namespace:               r.Logging.Spec.ControlNamespace,
-		DisableKubernetesFilter: disableKubernetesFilter,
-		FilterModify:            r.fluentbitSpec.FilterModify,
-		HealthCheck:             r.fluentbitSpec.HealthCheck,
+		Flush:                    r.fluentbitSpec.Flush,
+		Grace:                    r.fluentbitSpec.Grace,
+		ForceHotReloadAfterGrace: r.fluentbitSpec.ForceHotReloadAfterGrace,
+		LogLevel:                 r.fluentbitSpec.LogLevel,
+		EnabledIPv6:              r.fluentbitSpec.EnabledIPv6,
+		CoroStackSize:            r.fluentbitSpec.CoroStackSize,
+		Namespace:                r.Logging.Spec.ControlNamespace,
+		DisableKubernetesFilter:  disableKubernetesFilter,
+		FilterModify:             r.fluentbitSpec.FilterModify,
+		HealthCheck:              r.fluentbitSpec.HealthCheck,
 	}
 
 	input.DefaultParsers = fmt.Sprintf("%s/%s", StockConfigPath, "parsers.conf")
diff --git a/pkg/sdk/logging/api/v1beta1/fluentbit_types.go b/pkg/sdk/logging/api/v1beta1/fluentbit_types.go
@@ -77,6 +77,11 @@ type FluentbitSpec struct {
 	Flush int32 `json:"flush,omitempty"  plugin:"default:1"`
 	// Set the grace time in seconds as Integer value. The engine loop uses a Grace timeout to define wait time on exit.
 	Grace int32 `json:"grace,omitempty" plugin:"default:5"`
+	// HotReload pauses all inputs and waits until they finish. In certain situations this is unacceptable, for example if an output is down for a longer time.
+	// An undocumented option called "Hot_Reload.Ensure_Thread_Safety Off" can be used at the [SERVICE] config to force hotreload after the grace period.
+	// Please note that it might result in a SIGSEGV, but worst case kubelet will restart the container.
+	// See https://github.com/fluent/fluent-bit/pull/7509
+	ForceHotReloadAfterGrace bool `json:"forceHotReloadAfterGrace,omitempty"`
 	// Set the logging verbosity level. Allowed values are: error, warn, info, debug and trace. Values are accumulative, e.g: if 'debug' is set, it will include error, warning, info and debug.  Note that trace mode is only available if Fluent Bit was built with the WITH_TRACE option enabled.
 	LogLevel string `json:"logLevel,omitempty" plugin:"default:info"`
 	// Set the coroutines stack size in bytes. The value must be greater than the page size of the running system. Don't set too small value (say 4096), or coroutine threads can overrun the stack buffer.
