mixin: Make alert threshold values parametric

examples/alerts/alerts.md

@@ -225,7 +225,7 @@ rules:
       {{ $value }} seconds for the bucket operations.
   expr: |
     (
-      histogram_quantile(0.9, sum by (job, le) (rate(thanos_objstore_bucket_operation_duration_seconds_bucket{job=~"thanos-store.*"}[5m]))) > 15
+      histogram_quantile(0.9, sum by (job, le) (rate(thanos_objstore_bucket_operation_duration_seconds_bucket{job=~"thanos-store.*"}[5m]))) > 2
     and
       sum by (job) (rate(thanos_objstore_bucket_operation_duration_seconds_count{job=~"thanos-store.*"}[5m])) > 0
     )
@@ -336,7 +336,7 @@ rules:
       }} seconds for instant queries.
   expr: |
     (
-      histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{job=~"thanos-query.*", handler="query"}[5m]))) > 90
+      histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{job=~"thanos-query.*", handler="query"}[5m]))) > 40
     and
       sum by (job) (rate(http_request_duration_seconds_bucket{job=~"thanos-query.*", handler="query"}[5m])) > 0
     )
@@ -461,7 +461,7 @@ rules:
       $value }} seconds for the replicate operations.
   expr: |
     (
-      histogram_quantile(0.9, sum by (job, le) (rate(thanos_replicate_replication_run_duration_seconds_bucket{job=~"thanos-bucket-replicate.*"}[5m]))) > 120
+      histogram_quantile(0.9, sum by (job, le) (rate(thanos_replicate_replication_run_duration_seconds_bucket{job=~"thanos-bucket-replicate.*"}[5m]))) > 20
     and
       sum by (job) (rate(thanos_replicate_replication_run_duration_seconds_bucket{job=~"thanos-bucket-replicate.*"}[5m])) > 0
     )

examples/alerts/alerts.yaml

@@ -125,7 +125,7 @@ groups:
         }} seconds for instant queries.
     expr: |
       (
-        histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{job=~"thanos-query.*", handler="query"}[5m]))) > 90
+        histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{job=~"thanos-query.*", handler="query"}[5m]))) > 40
       and
         sum by (job) (rate(http_request_duration_seconds_bucket{job=~"thanos-query.*", handler="query"}[5m])) > 0
       )
@@ -277,7 +277,7 @@ groups:
         {{ $value }} seconds for the bucket operations.
     expr: |
       (
-        histogram_quantile(0.9, sum by (job, le) (rate(thanos_objstore_bucket_operation_duration_seconds_bucket{job=~"thanos-store.*"}[5m]))) > 15
+        histogram_quantile(0.9, sum by (job, le) (rate(thanos_objstore_bucket_operation_duration_seconds_bucket{job=~"thanos-store.*"}[5m]))) > 2
       and
         sum by (job) (rate(thanos_objstore_bucket_operation_duration_seconds_count{job=~"thanos-store.*"}[5m])) > 0
       )
@@ -468,7 +468,7 @@ groups:
         $value }} seconds for the replicate operations.
     expr: |
       (
-        histogram_quantile(0.9, sum by (job, le) (rate(thanos_replicate_replication_run_duration_seconds_bucket{job=~"thanos-bucket-replicate.*"}[5m]))) > 120
+        histogram_quantile(0.9, sum by (job, le) (rate(thanos_replicate_replication_run_duration_seconds_bucket{job=~"thanos-bucket-replicate.*"}[5m]))) > 20
       and
         sum by (job) (rate(thanos_replicate_replication_run_duration_seconds_bucket{job=~"thanos-bucket-replicate.*"}[5m])) > 0
       )

mixin/thanos/alerts/bucket_replicate.libsonnet

@@ -3,6 +3,8 @@
   bucket_replicate+:: {
     jobPrefix: error 'must provide job prefix for Thanos Bucket Replicate dashboard',
     selector: error 'must provide selector for Thanos Bucket Replicate dashboard',
+    errorThreshold: 10,
+    p99LatencyThreshold: 20,
   },
   prometheusAlerts+:: {
     groups+: [
@@ -32,7 +34,7 @@
                 sum(rate(thanos_replicate_replication_runs_total{result="error", %(selector)s}[5m]))
               / on (namespace) group_left
                 sum(rate(thanos_replicate_replication_runs_total{%(selector)s}[5m]))
-              ) * 100 >= 10
+              ) * 100 >= %(errorThreshold)s
             ||| % thanos.bucket_replicate,
             'for': '5m',
             labels: {
@@ -46,7 +48,7 @@
             },
             expr: |||
               (
-                histogram_quantile(0.9, sum by (job, le) (rate(thanos_replicate_replication_run_duration_seconds_bucket{%(selector)s}[5m]))) > 120
+                histogram_quantile(0.9, sum by (job, le) (rate(thanos_replicate_replication_run_duration_seconds_bucket{%(selector)s}[5m]))) > %(p99LatencyThreshold)s
               and
                 sum by (job) (rate(thanos_replicate_replication_run_duration_seconds_bucket{%(selector)s}[5m])) > 0
               )

mixin/thanos/alerts/compact.libsonnet

@@ -3,6 +3,8 @@
   compact+:: {
     jobPrefix: error 'must provide job prefix for Thanos Compact alerts',
     selector: error 'must provide selector for Thanos Compact alerts',
+    compactionErrorThreshold: 5,
+    bucketOpsErrorThreshold: 5,
   },
   prometheusAlerts+:: {
     groups+: [
@@ -41,7 +43,7 @@
                 sum by (job) (rate(thanos_compact_group_compactions_failures_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_compact_group_compactions_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(compactionErrorThreshold)s
               )
             ||| % thanos.compact,
             'for': '15m',
@@ -59,7 +61,7 @@
                 sum by (job) (rate(thanos_objstore_bucket_operation_failures_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_objstore_bucket_operations_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(bucketOpsErrorThreshold)s
               )
             ||| % thanos.compact,
             'for': '15m',

mixin/thanos/alerts/query.libsonnet

@@ -3,6 +3,11 @@
   query+:: {
     jobPrefix: error 'must provide job prefix for Thanos Query alerts',
     selector: error 'must provide selector for Thanos Query alerts',
+    httpErrorThreshold: 5,
+    grpcErrorThreshold: 5,
+    dnsErrorThreshold: 1,
+    p99QueryLatencyThreshold: 40,
+    p99QueryRangeLatencyThreshold: 90,
   },
   prometheusAlerts+:: {
     groups+: [
@@ -19,7 +24,7 @@
                 sum(rate(http_requests_total{code=~"5..", %(selector)s, handler="query"}[5m]))
               /
                 sum(rate(http_requests_total{%(selector)s, handler="query"}[5m]))
-              ) * 100 > 5
+              ) * 100 > %(httpErrorThreshold)s
             ||| % thanos.query,
             'for': '5m',
             labels: {
@@ -36,7 +41,7 @@
                 sum(rate(http_requests_total{code=~"5..", %(selector)s, handler="query_range"}[5m]))
               /
                 sum(rate(http_requests_total{%(selector)s, handler="query_range"}[5m]))
-              ) * 100 > 5
+              ) * 100 > %(httpErrorThreshold)s
             ||| % thanos.query,
             'for': '5m',
             labels: {
@@ -53,7 +58,7 @@
                 sum by (job) (rate(grpc_server_handled_total{grpc_code=~"Unknown|ResourceExhausted|Internal|Unavailable|DataLoss|DeadlineExceeded", %(selector)s}[5m]))
               /
                 sum by (job) (rate(grpc_server_started_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(grpcErrorThreshold)s
               )
             ||| % thanos.query,
             'for': '5m',
@@ -71,7 +76,7 @@
                 sum by (job) (rate(grpc_client_handled_total{grpc_code!="OK", %(selector)s}[5m]))
               /
                 sum by (job) (rate(grpc_client_started_total{%(selector)s}[5m]))
-              ) * 100 > 5
+              ) * 100 > %(grpcErrorThreshold)s
             ||| % thanos.query,
             'for': '5m',
             labels: {
@@ -88,7 +93,7 @@
                 sum by (job) (rate(thanos_querier_store_apis_dns_failures_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_querier_store_apis_dns_lookups_total{%(selector)s}[5m]))
-              ) * 100 > 1
+              ) * 100 > %(dnsErrorThreshold)s
             ||| % thanos.query,
             'for': '15m',
             labels: {
@@ -102,7 +107,7 @@
             },
             expr: |||
               (
-                histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{%(selector)s, handler="query"}[5m]))) > 90
+                histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{%(selector)s, handler="query"}[5m]))) > %(p99QueryLatencyThreshold)s
               and
                 sum by (job) (rate(http_request_duration_seconds_bucket{%(selector)s, handler="query"}[5m])) > 0
               )
@@ -119,7 +124,7 @@
             },
             expr: |||
               (
-                histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{%(selector)s, handler="query_range"}[5m]))) > 90
+                histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{%(selector)s, handler="query_range"}[5m]))) > %(p99QueryRangeLatencyThreshold)s
               and
                 sum by (job) (rate(http_request_duration_seconds_count{%(selector)s, handler="query_range"}[5m])) > 0
               )

mixin/thanos/alerts/receive.libsonnet

@@ -3,6 +3,10 @@
   receive+:: {
     jobPrefix: error 'must provide job prefix for Thanos Receive alerts',
     selector: error 'must provide selector for Thanos Receive alerts',
+    httpErrorThreshold: 5,
+    forwardErrorThreshold: 5,
+    refreshErrorThreshold: 0,
+    p99LatencyThreshold: 10,
   },
   prometheusAlerts+:: {
     groups+: [
@@ -19,7 +23,7 @@
                 sum(rate(http_requests_total{code=~"5..", %(selector)s, handler="receive"}[5m]))
               /
                 sum(rate(http_requests_total{%(selector)s, handler="receive"}[5m]))
-              ) * 100 > 5
+              ) * 100 > %(httpErrorThreshold)s
             ||| % thanos.receive,
             'for': '5m',
             labels: {
@@ -33,7 +37,7 @@
             },
             expr: |||
               (
-                histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{%(selector)s, handler="receive"}[5m]))) > 10
+                histogram_quantile(0.99, sum by (job, le) (rate(http_request_duration_seconds_bucket{%(selector)s, handler="receive"}[5m]))) > %(p99LatencyThreshold)s
               and
                 sum by (job) (rate(http_request_duration_seconds_count{%(selector)s, handler="receive"}[5m])) > 0
               )
@@ -53,7 +57,7 @@
                 sum by (job) (rate(thanos_receive_forward_requests_total{result="error", %(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_receive_forward_requests_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(forwardErrorThreshold)s
               )
             ||| % thanos.receive,
             'for': '5m',
@@ -71,7 +75,7 @@
                 sum by (job) (rate(thanos_receive_hashrings_file_errors_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_receive_hashrings_file_refreshes_total{%(selector)s}[5m]))
-              > 0
+              > %(refreshErrorThreshold)s
               )
             ||| % thanos.receive,
             'for': '15m',

mixin/thanos/alerts/rule.libsonnet

@@ -3,6 +3,10 @@
   rule+:: {
     jobPrefix: error 'must provide job prefix for Thanos Rule alerts',
     selector: error 'must provide selector for Thanos Rule alerts',
+    grpcErrorThreshold: 5,
+    rulerDnsErrorThreshold: 1,
+    alertManagerDnsErrorThreshold: 1,
+    evalErrorThreshold: 5,
   },
   prometheusAlerts+:: {
     groups+: [
@@ -45,7 +49,7 @@
                 sum by (job) (rate(prometheus_rule_evaluation_failures_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(prometheus_rule_evaluations_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(evalErrorThreshold)s
               )
             ||| % thanos.rule,
 
@@ -95,7 +99,7 @@
                 sum by (job) (rate(grpc_server_handled_total{grpc_code=~"Unknown|ResourceExhausted|Internal|Unavailable|DataLoss|DeadlineExceeded", %(selector)s}[5m]))
               /
                 sum by (job) (rate(grpc_server_started_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(grpcErrorThreshold)s
               )
             ||| % thanos.rule,
             'for': '5m',
@@ -124,7 +128,7 @@
                 sum by (job) (rate(thanos_ruler_query_apis_dns_failures_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_ruler_query_apis_dns_lookups_total{%(selector)s}[5m]))
-              * 100 > 1
+              * 100 > %(rulerDnsErrorThreshold)s
               )
             ||| % thanos.rule,
             'for': '15m',
@@ -142,7 +146,7 @@
                 sum by (job) (rate(thanos_ruler_alertmanagers_dns_failures_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_ruler_alertmanagers_dns_lookups_total{%(selector)s}[5m]))
-              * 100 > 1
+              * 100 > %(alertManagerDnsErrorThreshold)s
               )
             ||| % thanos.rule,
             'for': '15m',

mixin/thanos/alerts/store.libsonnet

@@ -3,6 +3,11 @@
   store+:: {
     jobPrefix: error 'must provide job prefix for Thanos Store alerts',
     selector: error 'must provide selector for Thanos Store alerts',
+    grpcErrorThreshold: 5,
+    compactionErrorThreshold: 5,
+    seriesGateErrorThreshold: 2,
+    bucketOpsErrorThreshold: 5,
+    bucketOpsP99LatencyThreshold: 2,
   },
   prometheusAlerts+:: {
     groups+: [
@@ -19,7 +24,7 @@
                 sum by (job) (rate(grpc_server_handled_total{grpc_code=~"Unknown|ResourceExhausted|Internal|Unavailable|DataLoss|DeadlineExceeded", %(selector)s}[5m]))
               /
                 sum by (job) (rate(grpc_server_started_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(grpcErrorThreshold)s
               )
             ||| % thanos.store,
             'for': '5m',
@@ -34,7 +39,7 @@
             },
             expr: |||
               (
-                histogram_quantile(0.9, sum by (job, le) (rate(thanos_bucket_store_series_gate_duration_seconds_bucket{%(selector)s}[5m]))) > 2
+                histogram_quantile(0.9, sum by (job, le) (rate(thanos_bucket_store_series_gate_duration_seconds_bucket{%(selector)s}[5m]))) > %(seriesGateErrorThreshold)s
               and
                 sum by (job) (rate(thanos_bucket_store_series_gate_duration_seconds_count{%(selector)s}[5m])) > 0
               )
@@ -54,7 +59,7 @@
                 sum by (job) (rate(thanos_objstore_bucket_operation_failures_total{%(selector)s}[5m]))
               /
                 sum by (job) (rate(thanos_objstore_bucket_operations_total{%(selector)s}[5m]))
-              * 100 > 5
+              * 100 > %(bucketOpsErrorThreshold)s
               )
             ||| % thanos.store,
             'for': '15m',
@@ -69,7 +74,7 @@
             },
             expr: |||
               (
-                histogram_quantile(0.9, sum by (job, le) (rate(thanos_objstore_bucket_operation_duration_seconds_bucket{%(selector)s}[5m]))) > 15
+                histogram_quantile(0.9, sum by (job, le) (rate(thanos_objstore_bucket_operation_duration_seconds_bucket{%(selector)s}[5m]))) > %(bucketOpsP99LatencyThreshold)s
               and
                 sum by (job) (rate(thanos_objstore_bucket_operation_duration_seconds_count{%(selector)s}[5m])) > 0
               )