| 1 | # you can disable an alarm notification by setting the 'to' line to: silent |
| 2 | |
| 3 | template: consul_license_expiration_time |
| 4 | on: consul.license_expiration_time |
| 5 | class: Errors |
| 6 | type: ServiceMesh |
| 7 | component: Consul |
| 8 | calc: $license_expiration |
| 9 | every: 60m |
| 10 | units: seconds |
| 11 | warn: $this < 14*24*60*60 |
| 12 | crit: $this < 7*24*60*60 |
| 13 | summary: Consul license expiration on ${label:node_name} |
| 14 | info: Consul Enterprise license expiration time on node ${label:node_name} datacenter ${label:datacenter} |
| 15 | to: sysadmin |
| 16 | |
| 17 | template: consul_autopilot_health_status |
| 18 | on: consul.autopilot_health_status |
| 19 | class: Errors |
| 20 | type: ServiceMesh |
| 21 | component: Consul |
| 22 | calc: $unhealthy |
| 23 | every: 10s |
| 24 | units: status |
| 25 | warn: $this == 1 |
| 26 | delay: down 5m multiplier 1.5 max 1h |
| 27 | summary: Consul datacenter ${label:datacenter} health |
| 28 | info: Datacenter ${label:datacenter} cluster is unhealthy as reported by server ${label:node_name} |
| 29 | to: sysadmin |
| 30 | |
| 31 | template: consul_autopilot_server_health_status |
| 32 | on: consul.autopilot_server_health_status |
| 33 | class: Errors |
| 34 | type: ServiceMesh |
| 35 | component: Consul |
| 36 | calc: $unhealthy |
| 37 | every: 10s |
| 38 | units: status |
| 39 | warn: $this == 1 |
| 40 | delay: down 5m multiplier 1.5 max 1h |
| 41 | summary: Consul server ${label:node_name} health |
| 42 | info: Server ${label:node_name} from datacenter ${label:datacenter} is unhealthy |
| 43 | to: sysadmin |
| 44 | |
| 45 | template: consul_raft_leader_last_contact_time |
| 46 | on: consul.raft_leader_last_contact_time |
| 47 | class: Errors |
| 48 | type: ServiceMesh |
| 49 | component: Consul |
| 50 | lookup: average -1m unaligned of quantile_0.5 |
| 51 | every: 10s |
| 52 | units: milliseconds |
| 53 | warn: $this > (($status >= $WARNING) ? (150) : (200)) |
| 54 | crit: $this > (($status == $CRITICAL) ? (200) : (500)) |
| 55 | delay: down 5m multiplier 1.5 max 1h |
| 56 | summary: Consul leader server ${label:node_name} last contact time |
| 57 | info: Median time elapsed since leader server ${label:node_name} datacenter ${label:datacenter} was last able to contact the follower nodes |
| 58 | to: sysadmin |
| 59 | |
| 60 | template: consul_raft_leadership_transitions |
| 61 | on: consul.raft_leadership_transitions_rate |
| 62 | class: Errors |
| 63 | type: ServiceMesh |
| 64 | component: Consul |
| 65 | lookup: sum -1m unaligned |
| 66 | every: 10s |
| 67 | units: transitions |
| 68 | warn: $this > 0 |
| 69 | delay: down 5m multiplier 1.5 max 1h |
| 70 | summary: Consul server ${label:node_name} leadership transitions |
| 71 | info: There has been a leadership change and server ${label:node_name} datacenter ${label:datacenter} has become the leader |
| 72 | to: sysadmin |
| 73 | |
| 74 | template: consul_raft_thread_main_saturation |
| 75 | on: consul.raft_thread_main_saturation_perc |
| 76 | class: Utilization |
| 77 | type: ServiceMesh |
| 78 | component: Consul |
| 79 | lookup: average -1m unaligned of quantile_0.9 |
| 80 | every: 10s |
| 81 | units: percentage |
| 82 | warn: $this > (($status >= $WARNING) ? (40) : (50)) |
| 83 | delay: down 5m multiplier 1.5 max 1h |
| 84 | summary: Consul server ${label:node_name} main Raft saturation |
| 85 | info: Average saturation of the main Raft goroutine on server ${label:node_name} datacenter ${label:datacenter} |
| 86 | to: sysadmin |
| 87 | |
| 88 | template: consul_raft_thread_fsm_saturation |
| 89 | on: consul.raft_thread_fsm_saturation_perc |
| 90 | class: Utilization |
| 91 | type: ServiceMesh |
| 92 | component: Consul |
| 93 | lookup: average -1m unaligned of quantile_0.9 |
| 94 | every: 10s |
| 95 | units: milliseconds |
| 96 | warn: $this > (($status >= $WARNING) ? (40) : (50)) |
| 97 | delay: down 5m multiplier 1.5 max 1h |
| 98 | summary: Consul server ${label:node_name} FSM Raft saturation |
| 99 | info: Average saturation of the FSM Raft goroutine on server ${label:node_name} datacenter ${label:datacenter} |
| 100 | to: sysadmin |
| 101 | |
| 102 | template: consul_client_rpc_requests_exceeded |
| 103 | on: consul.client_rpc_requests_exceeded_rate |
| 104 | class: Errors |
| 105 | type: ServiceMesh |
| 106 | component: Consul |
| 107 | lookup: sum -1m unaligned |
| 108 | every: 10s |
| 109 | units: requests |
| 110 | warn: $this > (($status >= $WARNING) ? (0) : (5)) |
| 111 | delay: down 5m multiplier 1.5 max 1h |
| 112 | summary: Consul server ${label:node_name} RPC requests rate |
| 113 | info: Number of rate-limited RPC requests made by server ${label:node_name} datacenter ${label:datacenter} |
| 114 | to: sysadmin |
| 115 | |
| 116 | template: consul_client_rpc_requests_failed |
| 117 | on: consul.client_rpc_requests_failed_rate |
| 118 | class: Errors |
| 119 | type: ServiceMesh |
| 120 | component: Consul |
| 121 | lookup: sum -1m unaligned |
| 122 | every: 10s |
| 123 | units: requests |
| 124 | warn: $this > (($status >= $WARNING) ? (0) : (5)) |
| 125 | delay: down 5m multiplier 1.5 max 1h |
| 126 | summary: Consul server ${label:node_name} failed RPC requests |
| 127 | info: number of failed RPC requests made by server ${label:node_name} datacenter ${label:datacenter} |
| 128 | to: sysadmin |
| 129 | |
| 130 | template: consul_node_health_check_status |
| 131 | on: consul.node_health_check_status |
| 132 | class: Errors |
| 133 | type: ServiceMesh |
| 134 | component: Consul |
| 135 | calc: $warning + $critical |
| 136 | every: 10s |
| 137 | units: status |
| 138 | warn: $this != nan AND $this != 0 |
| 139 | delay: down 5m multiplier 1.5 max 1h |
| 140 | summary: Consul node health check ${label:check_name} on ${label:node_name} |
| 141 | info: Node health check ${label:check_name} has failed on server ${label:node_name} datacenter ${label:datacenter} |
| 142 | to: sysadmin |
| 143 | |
| 144 | template: consul_service_health_check_status |
| 145 | on: consul.service_health_check_status |
| 146 | class: Errors |
| 147 | type: ServiceMesh |
| 148 | component: Consul |
| 149 | calc: $warning + $critical |
| 150 | every: 10s |
| 151 | units: status |
| 152 | warn: $this == 1 |
| 153 | delay: down 5m multiplier 1.5 max 1h |
| 154 | summary: Consul service health check ${label:check_name} service ${label:service_name} node ${label:node_name} |
| 155 | info: Service health check ${label:check_name} for service ${label:service_name} has failed on server ${label:node_name} datacenter ${label:datacenter} |
| 156 | to: sysadmin |
| 157 | |
| 158 | template: consul_gc_pause_time |
| 159 | on: consul.gc_pause_time |
| 160 | class: Errors |
| 161 | type: ServiceMesh |
| 162 | component: Consul |
| 163 | lookup: sum -1m unaligned |
| 164 | every: 10s |
| 165 | units: seconds |
| 166 | warn: $this > (($status >= $WARNING) ? (1) : (2)) |
| 167 | crit: $this > (($status >= $WARNING) ? (2) : (5)) |
| 168 | delay: down 5m multiplier 1.5 max 1h |
| 169 | summary: Consul server ${label:node_name} garbage collection pauses |
| 170 | info: Time spent in stop-the-world garbage collection pauses on server ${label:node_name} datacenter ${label:datacenter} |
| 171 | to: sysadmin |