Mercurial > code > home > repos > victoriametrics
annotate alert_rules.py @ 44:ccb6a2889e73
typo?
| author | drewp@bigasterisk.com |
|---|---|
| date | Wed, 14 Feb 2024 19:11:13 -0800 |
| parents | f05be84926e2 |
| children | e1db51416e73 |
| rev | line source |
|---|---|
| 23 | 1 """ |
| 2 pdm run invoke push-config | |
| 3 | |
| 4 docs: https://prometheus.io/docs/prometheus/latest/configuration/alerting_rules/ | |
| 5 "Whenever the alert expression results in one or more vector | |
| 6 elements at a given point in time, the alert counts as active for | |
| 7 these elements' label sets." | |
| 8 also https://www.metricfire.com/blog/top-5-prometheus-alertmanager-gotchas/#Missing-metrics | |
| 9 | |
| 10 """ | |
| 11 | |
| 12 import json | |
| 13 | |
| 14 | |
| 15 def k8sRules(): | |
| 16 # from https://awesome-prometheus-alerts.grep.to/rules.html | |
| 17 return [ | |
| 18 { | |
| 34 | 19 "alert": "metricsTargetMissing", |
| 20 "expr": 'up{job!~"cm-acme-.*"} == 0', | |
| 36 | 21 'for': '10m', |
| 31 | 22 "labels": { |
| 23 "severity": "critical" | |
| 24 }, | |
| 23 | 25 "annotations": { |
| 34 | 26 "summary": "metrics target missing (instance {{ $labels.instance }})", |
| 27 "description": "A metrics target has disappeared. An exporter might be crashed.\n VALUE = {{ $value }}", | |
| 23 | 28 }, |
| 29 }, | |
| 30 { | |
| 31 "alert": "KubernetesMemoryPressure", | |
| 32 "expr": 'kube_node_status_condition{condition="MemoryPressure",status="true"} == 1', | |
| 33 "for": "2m", | |
| 31 | 34 "labels": { |
| 35 "severity": "critical" | |
| 36 }, | |
| 23 | 37 "annotations": { |
| 38 "summary": "Kubernetes memory pressure (instance {{ $labels.instance }})", | |
| 39 "description": "{{ $labels.node }} has MemoryPressure condition\n VALUE = {{ $value }}", | |
| 40 }, | |
| 41 }, | |
| 42 { | |
| 43 "alert": "KubernetesDiskPressure", | |
| 44 "expr": 'kube_node_status_condition{condition="DiskPressure",status="true"} == 1', | |
| 45 "for": "2m", | |
| 31 | 46 "labels": { |
| 47 "severity": "critical" | |
| 48 }, | |
| 23 | 49 "annotations": { |
| 50 "summary": "Kubernetes disk pressure (instance {{ $labels.instance }})", | |
| 51 "description": "{{ $labels.node }} has DiskPressure condition\n VALUE = {{ $value }}", | |
| 52 }, | |
| 53 }, | |
| 54 { | |
| 55 "alert": "KubernetesOutOfDisk", | |
| 56 "expr": 'kube_node_status_condition{condition="OutOfDisk",status="true"} == 1', | |
| 57 "for": "2m", | |
| 31 | 58 "labels": { |
| 59 "severity": "critical" | |
| 60 }, | |
| 23 | 61 "annotations": { |
| 62 "summary": "Kubernetes out of disk (instance {{ $labels.instance }})", | |
| 63 "description": "{{ $labels.node }} has OutOfDisk condition\n VALUE = {{ $value }}", | |
| 64 }, | |
| 65 }, | |
| 66 { | |
| 67 "alert": "KubernetesJobFailed", | |
| 68 "expr": "kube_job_status_failed > 0", | |
| 31 | 69 "labels": { |
| 70 "severity": "warning" | |
| 71 }, | |
| 23 | 72 "annotations": { |
| 73 "summary": "Kubernetes Job failed (instance {{ $labels.instance }})", | |
| 74 "description": "Job {{$labels.namespace}}/{{$labels.exported_job}} failed to complete\n VALUE = {{ $value }}", | |
| 75 }, | |
| 76 }, | |
| 77 { | |
| 78 "alert": "KubernetesPodCrashLooping", | |
| 79 "expr": "increase(kube_pod_container_status_restarts_total[1m]) > 3", | |
| 80 "for": "2m", | |
| 31 | 81 "labels": { |
| 82 "severity": "warning" | |
| 83 }, | |
| 23 | 84 "annotations": { |
| 85 "summary": "Kubernetes pod crash looping (instance {{ $labels.instance }})", | |
| 86 "description": "Pod {{ $labels.pod }} is crash looping\n VALUE = {{ $value }}", | |
| 87 }, | |
| 88 }, | |
| 89 { | |
| 31 | 90 "alert": |
| 91 "KubernetesClientCertificateExpiresNextWeek", | |
| 92 "expr": | |
| 93 'apiserver_client_certificate_expiration_seconds_count{job="apiserver"} > 0 and histogram_quantile(0.01, sum by (job, le) (rate(apiserver_client_certificate_expiration_seconds_bucket{job="apiserver"}[5m]))) < 7*24*60*60', | |
| 94 "labels": { | |
| 95 "severity": "warning" | |
| 96 }, | |
| 23 | 97 "annotations": { |
| 98 "summary": "Kubernetes client certificate expires next week (instance {{ $labels.instance }})", | |
| 99 "description": "A client certificate used to authenticate to the apiserver is expiring next week.\n VALUE = {{ $value }}", | |
| 100 }, | |
| 101 }, | |
| 102 { | |
| 103 "alert": "container_waiting", | |
| 34 | 104 "expr": "sum by (namespace, pod, container)(kube_pod_container_status_waiting!=0)", |
| 105 "annotations": { | |
| 106 "description": '', | |
| 107 "dashboard": "https://bigasterisk.com/k/clusters/local/namespaces/{{ $labels.namespace }}/pods/{{ $labels.pod }}", | |
| 108 }, | |
| 23 | 109 "for": "2m", |
| 110 }, | |
| 111 ] | |
| 112 | |
| 113 | |
|
32
eb1de82c93aa
refactor the merging of all the groups
drewp@bigasterisk.com
parents:
31
diff
changeset
|
114 def allRules(ctx): |
| 23 | 115 return { |
| 116 "groups": [ | |
| 117 { | |
| 118 "name": "k8s", | |
|
28
e114edff93dc
more explicit intervals. try to get a single day of notification out of a disk err increase
drewp@bigasterisk.com
parents:
27
diff
changeset
|
119 "interval": "1m", |
| 23 | 120 "rules": k8sRules(), |
| 121 }, | |
| 122 # | |
| 123 # any presence of starlette_request_duration_seconds_created{app_name="starlette",method="GET",path="/",status_code="200"} 1.6460176156784086e+09 means someone forgot to set app name | |
| 124 { | |
| 31 | 125 "name": |
| 126 "Outages", | |
| 127 "interval": | |
| 128 "1m", | |
| 23 | 129 "rules": [ |
| 130 { | |
| 131 "alert": "powereagleStalled", | |
| 132 "expr": "rate(house_power_w[100m]) == 0", | |
| 133 "for": "0m", | |
| 31 | 134 "labels": { |
| 135 "severity": "losingData" | |
| 136 }, | |
| 23 | 137 "annotations": { |
| 138 "summary": "power eagle data stalled", | |
| 139 "description": "logs at https://bigasterisk.com/k/clusters/local/namespaces/default/deployments/power-eagle/logs", | |
| 140 }, | |
| 141 }, | |
| 142 { | |
| 143 "alert": "powereagleAbsent", | |
| 144 "expr": "absent_over_time(house_power_w[5m])", | |
| 145 "for": "2m", | |
| 31 | 146 "labels": { |
| 147 "severity": "losingData" | |
| 148 }, | |
| 23 | 149 "annotations": { |
| 150 "summary": "power eagle data missing", | |
| 151 "description": "logs at https://bigasterisk.com/k/clusters/local/namespaces/default/deployments/power-eagle/logs", | |
| 152 }, | |
| 153 }, | |
| 154 { | |
| 155 "alert": "absent_zigbee", | |
| 156 "expr": 'absent(container_last_seen{container="zigbee2mqtt"})', | |
| 157 }, | |
| 158 { | |
| 159 "alert": "net_routes_sync", | |
| 160 "expr": 'rate(starlette_request_duration_seconds_count{app_name="net_routes",path="/routes"}[5m]) < 1/70', | |
| 161 "for": "10m", | |
| 31 | 162 "labels": { |
| 163 "severity": "houseUsersAffected" | |
| 164 }, | |
| 23 | 165 "annotations": { |
| 166 "summary": "net_routes is not getting regular updates" | |
| 167 }, | |
| 168 }, | |
| 169 ], | |
| 170 }, | |
| 171 { | |
| 36 | 172 "name": |
| 173 "disk_errs", | |
| 174 "interval": | |
| 175 "2d", | |
| 31 | 176 "rules": [{ |
| 36 | 177 "alert": "zpool_device_error_increase", |
| 178 "labels": { | |
| 179 "severity": "warning" | |
| 180 }, | |
| 181 "expr": 'increase(zpool_device_error_count[3d]) > 0', | |
| 182 }, { | |
| 31 | 183 "alert": "zpool_device_error_count", |
| 184 "labels": { | |
| 185 "severity": "warning" | |
|
28
e114edff93dc
more explicit intervals. try to get a single day of notification out of a disk err increase
drewp@bigasterisk.com
parents:
27
diff
changeset
|
186 }, |
| 36 | 187 "expr": 'zpool_device_error_count > 0', |
| 31 | 188 }], |
|
28
e114edff93dc
more explicit intervals. try to get a single day of notification out of a disk err increase
drewp@bigasterisk.com
parents:
27
diff
changeset
|
189 }, |
|
e114edff93dc
more explicit intervals. try to get a single day of notification out of a disk err increase
drewp@bigasterisk.com
parents:
27
diff
changeset
|
190 { |
| 31 | 191 "name": |
| 34 | 192 "front_door", |
| 193 "interval": | |
| 194 "5m", | |
| 195 "rules": [ | |
| 196 { | |
| 36 | 197 "alert": "front_door_reader_esp32_no_mqtt", |
| 198 'expr': 'hw_connected{job="fingerprint"} < 1', | |
| 199 "annotations": { | |
| 200 "summary": "see https://bigasterisk.com/front-door-lock/" | |
| 201 }, | |
| 202 }, | |
| 203 { | |
| 204 "alert": "fronr_door_reader_svc_down", | |
| 205 'expr': 'up{job="fingerprint"} < 1', | |
| 206 "annotations": { | |
| 207 "summary": "see https://bigasterisk.com/front-door-lock/" | |
| 208 }, | |
| 209 }, | |
| 210 { | |
| 211 "alert": "fronr_door_reader_svc_reader_no_mqtt", | |
| 212 'expr': 'mqtt_connected{job="fingerprint"} < 1', | |
| 213 "annotations": { | |
| 214 "summary": "see https://bigasterisk.com/front-door-lock/" | |
| 215 }, | |
| 34 | 216 }, |
| 217 { | |
| 36 | 218 "alert": "front_door_lock_svc_down", |
| 219 'expr': 'up{job="front-door-lock"} < 1', | |
| 220 "annotations": { | |
| 221 "summary": "see https://bigasterisk.com/front-door-lock/" | |
| 222 }, | |
| 34 | 223 }, |
| 36 | 224 { |
| 225 "alert": "front_door_lock_svc_no_mqtt", | |
| 226 'expr': 'mqtt_connected{job="front-door-lock"} < 1', | |
| 227 "annotations": { | |
| 228 "summary": "see https://bigasterisk.com/front-door-lock/" | |
| 229 }, | |
| 230 }, | |
| 231 { | |
| 232 "alert": "front_door_lock_esp32_no_mqtt", | |
| 233 'expr': 'hw_connected{job="front-door-lock"} < 1', | |
| 234 "annotations": { | |
| 235 "summary": "see https://bigasterisk.com/front-door-lock/" | |
| 236 }, | |
| 237 }, | |
| 238 ], | |
| 34 | 239 }, |
| 240 { | |
|
41
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
241 "name": "http", |
|
37
6e27d280b598
watch https cert ages (testing with 45d)
drewp@bigasterisk.com
parents:
36
diff
changeset
|
242 "interval": "1h", |
|
41
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
243 'rules': [ |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
244 { |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
245 'alert': 'old_https_certs', |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
246 'expr': 'min by (source) (x509_cert_enddate - now())/86400 < 15', |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
247 }, { |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
248 'alert': 'high_500_response_rate', |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
249 'expr': 'avg_over_time(rate(sum by (envoy_cluster_name) (envoy_cluster_internal_upstream_rq_xx{envoy_response_code_class="5"})[20m])) > 0.02', |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
250 }, |
|
407ee7fbda13
rm double metrics; add alert for too-many-500s
drewp@bigasterisk.com
parents:
40
diff
changeset
|
251 ], |
|
37
6e27d280b598
watch https cert ages (testing with 45d)
drewp@bigasterisk.com
parents:
36
diff
changeset
|
252 }, |
|
6e27d280b598
watch https cert ages (testing with 45d)
drewp@bigasterisk.com
parents:
36
diff
changeset
|
253 { |
| 42 | 254 "name": "ping", |
| 255 "interval": "1m", | |
| 256 "rules": [{ | |
| 257 "alert": "ping_failed", | |
| 258 "expr": 'max_over_time(probe_success{job="ping"}[1m]) < 1', | |
| 259 }] | |
| 260 }, | |
| 261 { | |
| 34 | 262 "name": |
| 31 | 263 "alerts", |
| 23 | 264 "rules": [ |
| 265 { | |
| 266 "alert": "kube_node_status_bad_condition", | |
| 267 "for": "2h", | |
| 31 | 268 "labels": { |
| 269 "severity": "warning" | |
| 270 }, | |
| 23 | 271 "expr": 'kube_node_status_condition{condition=~".*Pressure",status="true"} > 0', |
| 272 }, | |
| 273 { | |
| 274 "alert": "housePower", | |
|
28
e114edff93dc
more explicit intervals. try to get a single day of notification out of a disk err increase
drewp@bigasterisk.com
parents:
27
diff
changeset
|
275 "for": "1h", |
| 31 | 276 "labels": { |
| 277 "severity": "waste" | |
| 278 }, | |
| 23 | 279 "expr": "house_power_w > 4000", |
| 31 | 280 "annotations": { |
| 281 "summary": "house power usage over 4KW" | |
| 282 }, | |
| 23 | 283 }, |
| 284 { | |
| 285 "alert": "host_root_fs_space_low", | |
| 286 "for": "20m", | |
| 31 | 287 "labels": { |
| 288 "severity": "warning" | |
| 289 }, | |
| 34 | 290 "expr": 'disk_free{host!="garage",path="/"} < 20G', |
| 23 | 291 }, |
| 292 { | |
| 293 "alert": "zpool_space_low", | |
| 294 "for": "20m", | |
| 31 | 295 "labels": { |
| 296 "severity": "warning" | |
| 297 }, | |
| 23 | 298 "expr": 'last_over_time(zfs_pool_free_bytes{pool="stor7"}[1h]) < 100G', |
| 299 }, | |
| 300 { | |
| 301 "alert": "disk_week_incr", | |
| 302 "for": "20m", | |
| 31 | 303 "labels": { |
| 304 "severity": "warning" | |
| 305 }, | |
| 23 | 306 "expr": 'round(increase(disk_used{path=~"/my/.*"}[1d])/1M) > 5000', |
| 31 | 307 "annotations": { |
| 308 "summary": "high mb/week on zfs dir" | |
| 309 }, | |
| 23 | 310 }, |
| 311 { | |
| 312 "alert": "high_logging", | |
| 31 | 313 "for": "3h", |
| 314 "labels": { | |
| 315 "severity": "waste" | |
| 316 }, | |
| 23 | 317 "expr": "sum by (container) (rate(kubelet_container_log_filesystem_used_bytes[3h])) > 4k", |
| 31 | 318 "annotations": { |
| 319 "summary": "high log output rate" | |
| 320 }, | |
| 23 | 321 }, |
| 322 { | |
| 323 "alert": "stale_process", | |
| 324 "for": "1d", | |
| 31 | 325 "labels": { |
| 326 "severity": "dataRisk" | |
| 327 }, | |
| 23 | 328 "expr": "round((time() - filestat_modification_time/1e9) / 86400) > 14", |
| 31 | 329 "annotations": { |
| 330 "summary": "process time is old" | |
| 331 }, | |
| 23 | 332 }, |
| 333 { | |
| 334 "alert": "starlette", | |
| 335 "for": "1m", | |
| 31 | 336 "labels": { |
| 337 "severity": "fix" | |
| 338 }, | |
| 23 | 339 "expr": 'starlette_request_duration_seconds_created{app_name="starlette"}', |
| 31 | 340 "annotations": { |
| 341 "summary": "set starlette app name" | |
| 342 }, | |
| 23 | 343 }, |
| 344 { | |
| 345 "alert": "ssl_certs_expiring_soon", | |
| 346 "expr": "min((min_over_time(probe_ssl_earliest_cert_expiry[1d])-time())/86400) < 10", | |
| 31 | 347 "labels": { |
| 348 "severity": "warning" | |
| 349 }, | |
| 23 | 350 "annotations": { |
| 351 "summary": "cert expiring soon. See https://bigasterisk.com/grafana/d/z1YtDa3Gz/certs?orgId=1\nVALUE = {{ $value }}" | |
| 352 }, | |
| 353 }, | |
| 354 ], | |
| 355 }, | |
|
32
eb1de82c93aa
refactor the merging of all the groups
drewp@bigasterisk.com
parents:
31
diff
changeset
|
356 ] + hostsExpectedOnline(ctx)['groups'] |
| 23 | 357 } |
| 358 | |
| 359 | |
| 360 def _runJson(ctx, cmd): | |
| 361 return json.loads(ctx.run(cmd, hide="stdout").stdout) | |
| 362 | |
| 363 | |
| 364 def hostsExpectedOnline(ctx): | |
| 365 return _runJson(ctx, "cd /my/serv/lanscape; pdm run python hosts_expected_online.py") | |
| 366 | |
| 367 |