Etcd panels (#137)

* Add some etcd duration histogram graphs * Add etcd_pod picker * Add etcd_pod variable and picker --------- Signed-off-by: Andrew Collins <[email protected]>
cloud-bulldozer · Sep 27, 2024 · f374918 · f374918
1 parent 1b03ca2
commit f374918
Show file tree

Hide file tree

Showing 4 changed files with 106 additions and 28 deletions.
diff --git a/assets/etcd-on-cluster-dashboard/panels.libsonnet b/assets/etcd-on-cluster-dashboard/panels.libsonnet
@@ -5,12 +5,13 @@ local g = import 'github.com/grafana/grafonnet/gen/grafonnet-latest/main.libsonn
     local timeSeries = g.panel.timeSeries,
     local custom = timeSeries.fieldConfig.defaults.custom,
     local options = timeSeries.options,
+    local standardOptions = timeSeries.standardOptions,
+    local byRegexp = timeSeries.standardOptions.override.byRegexp,
 
     base(title, unit, targets, gridPos):
       timeSeries.new(title)
       + timeSeries.queryOptions.withTargets(targets)
       + timeSeries.datasource.withType('prometheus')
-      + timeSeries.datasource.withUid('$Datasource')
       + timeSeries.standardOptions.withUnit(unit)
       + timeSeries.gridPos.withX(gridPos.x)
       + timeSeries.gridPos.withY(gridPos.y)
@@ -38,7 +39,30 @@ local g = import 'github.com/grafana/grafonnet/gen/grafonnet-latest/main.libsonn
         'mean',
         'max',
       ])
-      + options.legend.withDisplayMode('table'),
+      + options.legend.withDisplayMode('table')
+      + options.legend.withSortBy('Max')
+      + options.legend.withSortDesc(true),
+
+    generalCounter(title, unit, targets, gridPos):
+      self.base(title, unit, targets, gridPos)
+      + options.legend.withCalcs([
+        'first',
+        'min',
+        'max',
+        'last',
+      ]),
+
+    histogramStatsRightHand(title, unit, targets, gridPos, leftAxis):
+      self.generalCounter(title, unit, targets, gridPos)
+      + custom.withAxisLabel(leftAxis)
+      + options.legend.withDisplayMode('table')
+      + options.legend.withSortBy('Max')
+      + standardOptions.withOverrides([
+        byRegexp.new('.*irate.*')
+        + byRegexp.withProperty('custom.axisPlacement', 'right')
+        + byRegexp.withProperty('custom.axisLabel', 'rate')
+        + byRegexp.withProperty('unit', 'none'),
+      ]),
 
     withoutCalcsAgg(title, unit, targets, gridPos):
       self.base(title, unit, targets, gridPos)

diff --git a/assets/etcd-on-cluster-dashboard/queries.libsonnet b/assets/etcd-on-cluster-dashboard/queries.libsonnet
@@ -12,22 +12,46 @@ local generateTimeSeriesQuery(query, legend) = [
 {
   CPUUsage: {
     query():
-      generateTimeSeriesQuery('sum(irate(container_cpu_usage_seconds_total{namespace="openshift-etcd", container="etcd"}[2m])) by (pod) * 100', '{{ pod }}'),
+      generateTimeSeriesQuery('sum(irate(container_cpu_usage_seconds_total{namespace="openshift-etcd", container="etcd",pod=~"$etcd_pod"}[2m])) by (pod) * 100', '{{ pod }}'),
   },
 
   memoryUsage: {
     query():
-      generateTimeSeriesQuery('sum(avg_over_time(container_memory_working_set_bytes{container="",pod!="", namespace=~"openshift-etcd.*"}[2m])) BY (pod, namespace)', '{{ pod }}'),
+      generateTimeSeriesQuery('sum(avg_over_time(container_memory_working_set_bytes{container="",pod!="", namespace=~"openshift-etcd.*",pod=~"$etcd_pod"}[2m])) BY (pod, namespace)', '{{ pod }}'),
   },
 
   diskWalSyncDuration: {
     query():
-      generateTimeSeriesQuery('histogram_quantile(0.99, sum(irate(etcd_disk_wal_fsync_duration_seconds_bucket{namespace="openshift-etcd"}[2m])) by (pod, le))', '{{pod}} WAL fsync'),
+      generateTimeSeriesQuery('histogram_quantile(0.99, sum(rate(etcd_disk_wal_fsync_duration_seconds_bucket{namespace="openshift-etcd",pod=~"$etcd_pod"}[5m])) by (pod, le))', '{{pod}} WAL fsync'),
   },
 
-  diskBackendSyncDuration: {
+  diskWalSyncDurationSum: {
     query():
-      generateTimeSeriesQuery('histogram_quantile(0.99, sum(irate(etcd_disk_backend_commit_duration_seconds_bucket{namespace="openshift-etcd"}[2m])) by (pod, le))', '{{pod}} DB fsync'),
+      generateTimeSeriesQuery('irate(etcd_disk_wal_fsync_duration_seconds_sum{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '2m irate WAL sum {{instance}} ')
+      + generateTimeSeriesQuery('etcd_disk_wal_fsync_duration_seconds_sum{namespace="openshift-etcd",pod=~"$etcd_pod"}', 'WAL sum {{instance}} '),
+  },
+
+  diskWalSyncDurationCount: {
+    query():
+      generateTimeSeriesQuery('irate(etcd_disk_wal_fsync_duration_seconds_count{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '2m irate WAL count {{instance}} ')
+      + generateTimeSeriesQuery('etcd_disk_wal_fsync_duration_seconds_count{namespace="openshift-etcd",pod=~"$etcd_pod"}', 'WAL count {{instance}} '),
+  },
+
+  diskBackendCommitDuration: {
+    query():
+      generateTimeSeriesQuery('histogram_quantile(0.99, sum(rate(etcd_disk_backend_commit_duration_seconds_bucket{namespace="openshift-etcd",pod=~"$etcd_pod"}[5m])) by (pod, le))', '{{pod}} DB fsync'),
+  },
+
+  diskBackendCommitDurationSum: {
+    query():
+      generateTimeSeriesQuery('irate(etcd_disk_backend_commit_duration_seconds_sum{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '2m irate WAL sum {{instance}} ')
+      + generateTimeSeriesQuery('etcd_disk_backend_commit_duration_seconds_sum{namespace="openshift-etcd",pod=~"$etcd_pod"}', 'WAL sum {{instance}} '),
+  },
+
+  diskBackendCommitDurationCount: {
+    query():
+      generateTimeSeriesQuery('irate(etcd_disk_backend_commit_duration_seconds_count{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '2m irate WAL count {{instance}} ')
+      + generateTimeSeriesQuery('etcd_disk_backend_commit_duration_seconds_count{namespace="openshift-etcd",pod=~"$etcd_pod"}', 'WAL count {{instance}} '),
   },
 
   etcdContainerDiskWrites: {
@@ -37,8 +61,8 @@ local generateTimeSeriesQuery(query, legend) = [
 
   dbSize: {
     query():
-      generateTimeSeriesQuery('etcd_mvcc_db_total_size_in_bytes{namespace="openshift-etcd"}', '{{pod}} DB physical size')
-      + generateTimeSeriesQuery('etcd_mvcc_db_total_size_in_use_in_bytes{namespace="openshift-etcd"}', '{{pod}} DB logical size'),
+      generateTimeSeriesQuery('etcd_mvcc_db_total_size_in_bytes{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{pod}} DB physical size')
+      + generateTimeSeriesQuery('etcd_mvcc_db_total_size_in_use_in_bytes{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{pod}} DB logical size'),
   },
 
   containerNetworkTraffic: {
@@ -49,19 +73,19 @@ local generateTimeSeriesQuery(query, legend) = [
 
   p99PeerToPeerLatency: {
     query():
-      generateTimeSeriesQuery('histogram_quantile(0.99, rate(etcd_network_peer_round_trip_time_seconds_bucket{namespace="openshift-etcd"}[2m]))', '{{pod}}'),
+      generateTimeSeriesQuery('histogram_quantile(0.99, rate(etcd_network_peer_round_trip_time_seconds_bucket{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m]))', '{{pod}}'),
   },
 
   peerNetworkTraffic: {
     query():
-      generateTimeSeriesQuery('rate(etcd_network_peer_received_bytes_total{namespace="openshift-etcd"}[2m])', 'rx {{pod}} Peer Traffic')
-      + generateTimeSeriesQuery('rate(etcd_network_peer_sent_bytes_total{namespace="openshift-etcd"}[2m])', 'tx {{pod}} Peer Traffic'),
+      generateTimeSeriesQuery('rate(etcd_network_peer_received_bytes_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', 'rx {{pod}} Peer Traffic')
+      + generateTimeSeriesQuery('rate(etcd_network_peer_sent_bytes_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', 'tx {{pod}} Peer Traffic'),
   },
 
   gRPCNetworkTraffic: {
     query():
-      generateTimeSeriesQuery('rate(etcd_network_client_grpc_received_bytes_total{namespace="openshift-etcd"}[2m])', 'rx {{pod}}')
-      + generateTimeSeriesQuery('rate(etcd_network_client_grpc_sent_bytes_total{namespace="openshift-etcd"}[2m])', 'tx {{pod}}'),
+      generateTimeSeriesQuery('rate(etcd_network_client_grpc_received_bytes_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', 'rx {{pod}}')
+      + generateTimeSeriesQuery('rate(etcd_network_client_grpc_sent_bytes_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', 'tx {{pod}}'),
   },
 
   activeStreams: {
@@ -77,17 +101,17 @@ local generateTimeSeriesQuery(query, legend) = [
 
   dbSpaceUsed: {
     query():
-      generateTimeSeriesQuery('(etcd_mvcc_db_total_size_in_bytes{namespace="openshift-etcd"} / etcd_server_quota_backend_bytes{namespace="openshift-etcd"})*100', '{{pod}}'),
+      generateTimeSeriesQuery('(etcd_mvcc_db_total_size_in_bytes{namespace="openshift-etcd",pod=~"$etcd_pod"} / etcd_server_quota_backend_bytes{namespace="openshift-etcd",pod=~"$etcd_pod"})*100', '{{pod}}'),
   },
 
   dbLeftCapacity: {
     query():
-      generateTimeSeriesQuery('etcd_server_quota_backend_bytes{namespace="openshift-etcd"} - etcd_mvcc_db_total_size_in_bytes{namespace="openshift-etcd"}', '{{pod}}'),
+      generateTimeSeriesQuery('etcd_server_quota_backend_bytes{namespace="openshift-etcd",pod=~"$etcd_pod"} - etcd_mvcc_db_total_size_in_bytes{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{pod}}'),
   },
 
   dbSizeLimit: {
     query():
-      generateTimeSeriesQuery('etcd_server_quota_backend_bytes{namespace="openshift-etcd"}', '{{ pod }} Quota Bytes'),
+      generateTimeSeriesQuery('etcd_server_quota_backend_bytes{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{ pod }} Quota Bytes'),
   },
 
   raftProposals: {
@@ -115,34 +139,34 @@ local generateTimeSeriesQuery(query, legend) = [
 
   keys: {
     query():
-      generateTimeSeriesQuery('etcd_debugging_mvcc_keys_total{namespace="openshift-etcd"}', '{{ pod }} Num keys'),
+      generateTimeSeriesQuery('etcd_debugging_mvcc_keys_total{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{ pod }} Num keys'),
   },
 
   leaderElectionsPerDay: {
     query():
-      generateTimeSeriesQuery('changes(etcd_server_leader_changes_seen_total{namespace="openshift-etcd"}[1d])', '{{instance}} Total Leader Elections Per Day'),
+      generateTimeSeriesQuery('changes(etcd_server_leader_changes_seen_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[1d])', '{{instance}} Total Leader Elections Per Day'),
   },
 
   slowOperations: {
     query():
-      generateTimeSeriesQuery('delta(etcd_server_slow_apply_total{namespace="openshift-etcd"}[2m])', '{{ pod }} slow applies')
-      + generateTimeSeriesQuery('delta(etcd_server_slow_read_indexes_total{namespace="openshift-etcd"}[2m])', '{{ pod }} slow read indexes'),
+      generateTimeSeriesQuery('delta(etcd_server_slow_apply_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '{{ pod }} slow applies')
+      + generateTimeSeriesQuery('delta(etcd_server_slow_read_indexes_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '{{ pod }} slow read indexes'),
   },
 
   keyOperations: {
     query():
-      generateTimeSeriesQuery('rate(etcd_mvcc_put_total{namespace="openshift-etcd"}[2m])', '{{ pod }} puts/s')
-      + generateTimeSeriesQuery('rate(etcd_mvcc_delete_total{namespace="openshift-etcd"}[2m])', '{{ pod }} deletes/s'),
+      generateTimeSeriesQuery('rate(etcd_mvcc_put_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '{{ pod }} puts/s')
+      + generateTimeSeriesQuery('rate(etcd_mvcc_delete_total{namespace="openshift-etcd",pod=~"$etcd_pod"}[2m])', '{{ pod }} deletes/s'),
   },
 
   heartBeatFailure: {
     query():
-      generateTimeSeriesQuery('etcd_server_heartbeat_send_failures_total{namespace="openshift-etcd"}', '{{ pod }} heartbeat failures')
-      + generateTimeSeriesQuery('etcd_server_health_failures{namespace="openshift-etcd"}', '{{ pod }} health failures'),
+      generateTimeSeriesQuery('etcd_server_heartbeat_send_failures_total{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{ pod }} heartbeat failures')
+      + generateTimeSeriesQuery('etcd_server_health_failures{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{ pod }} health failures'),
   },
 
   compactedKeys: {
     query():
-      generateTimeSeriesQuery('etcd_debugging_mvcc_db_compaction_keys_total{namespace="openshift-etcd"}', '{{ pod  }} keys compacted'),
+      generateTimeSeriesQuery('etcd_debugging_mvcc_db_compaction_keys_total{namespace="openshift-etcd",pod=~"$etcd_pod"}', '{{ pod  }} keys compacted'),
   },
 }
diff --git a/assets/etcd-on-cluster-dashboard/variables.libsonnet b/assets/etcd-on-cluster-dashboard/variables.libsonnet
@@ -9,4 +9,15 @@ local var = g.dashboard.variable;
     + var.query.withRefresh(1)
     + var.query.selectionOptions.withMulti(false)
     + var.query.selectionOptions.withIncludeAll(false),
+
+  etcd_pod:
+    var.query.new('etcd_pod')
+    + var.query.withDatasourceFromVariable(self.Datasource)
+    + var.query.queryTypes.withLabelValues(
+      'pod',
+      'etcd_cluster_version',
+    )
+    + var.query.withRefresh(2)
+    + var.query.selectionOptions.withMulti()
+    + var.query.selectionOptions.withIncludeAll(true),
 }
diff --git a/templates/General/etcd-on-cluster-dashboard.jsonnet b/templates/General/etcd-on-cluster-dashboard.jsonnet
@@ -14,6 +14,7 @@ g.dashboard.new('etcd-cluster-info dashoard')
 + g.dashboard.graphTooltip.withSharedCrosshair()
 + g.dashboard.withVariables([
   variables.Datasource,
+  variables.etcd_pod,
 ])
 
 + g.dashboard.withPanels([
@@ -24,11 +25,29 @@ g.dashboard.new('etcd-cluster-info dashoard')
     panels.timeSeries.generalUsageAgg('CPU usage', 'percent', queries.CPUUsage.query(), { x: 0, y: 1, w: 12, h: 8 }),
     panels.timeSeries.generalUsageAgg('Memory usage', 'bytes', queries.memoryUsage.query(), { x: 12, y: 1, w: 12, h: 8 }),
     panels.timeSeries.generalUsageAgg('Disk WAL Sync Duration', 's', queries.diskWalSyncDuration.query(), { x: 0, y: 8, w: 12, h: 8 }),
-    panels.timeSeries.generalUsageAgg('Disk Backend Sync Duration', 's', queries.diskBackendSyncDuration.query(), { x: 12, y: 8, w: 12, h: 8 }),
+    panels.timeSeries.generalUsageAgg('Disk Backend Sync Duration', 's', queries.diskBackendCommitDuration.query(), { x: 12, y: 8, w: 12, h: 8 }),
     panels.timeSeries.generalUsageAgg('Etcd container disk writes', 'Bps', queries.etcdContainerDiskWrites.query(), { x: 0, y: 16, w: 12, h: 8 }),
     panels.timeSeries.generalUsageAgg('DB Size', 'bytes', queries.dbSize.query(), { x: 12, y: 16, w: 12, h: 8 }),
   ]),
 
+  g.panel.row.new('WAL fsync Duration Detailed')
+  + g.panel.row.withGridPos({ x: 0, y: 14, w: 24, h: 1 })
+  + g.panel.row.withCollapsed(true)
+  + g.panel.row.withPanels([
+    panels.timeSeries.generalUsageAgg('WAL fsync Duration p99', 's', queries.diskWalSyncDuration.query(), { x: 0, y: 0, w: 8, h: 8 }),
+    panels.timeSeries.histogramStatsRightHand('WAL fsync Duration sum', 'none', queries.diskWalSyncDurationSum.query(), { x: 8, y: 0, w: 8, h: 8 }, 'sum'),
+    panels.timeSeries.histogramStatsRightHand('WAL fsync Duration count', 'none', queries.diskWalSyncDurationCount.query(), { x: 16, y: 0, w: 8, h: 8 }, 'count'),
+  ]),
+
+  g.panel.row.new('Backend Commit Duration Detailed')
+  + g.panel.row.withGridPos({ x: 0, y: 14, w: 24, h: 1 })
+  + g.panel.row.withCollapsed(true)
+  + g.panel.row.withPanels([
+    panels.timeSeries.generalUsageAgg('Backend Commit Duration', 's', queries.diskBackendCommitDuration.query(), { x: 0, y: 0, w: 8, h: 8 }),
+    panels.timeSeries.histogramStatsRightHand('Backend Commit Duration sum', 'none', queries.diskBackendCommitDurationSum.query(), { x: 8, y: 0, w: 8, h: 8 }, 'sum'),
+    panels.timeSeries.histogramStatsRightHand('Backend Commit Duration count', 'none', queries.diskBackendCommitDurationCount.query(), { x: 16, y: 0, w: 8, h: 8 }, 'count'),
+  ]),
+
   g.panel.row.new('Network Usage')
   + g.panel.row.withGridPos({ x: 0, y: 14, w: 24, h: 1 })
   + g.panel.row.withCollapsed(true)
@@ -62,7 +81,7 @@ g.dashboard.new('etcd-cluster-info dashoard')
     panels.timeSeries.GeneralInfo('Leader Elections Per Day', 'short', queries.leaderElectionsPerDay.query(), { x: 0, y: 12, w: 12, h: 6 }),
     panels.timeSeries.GeneralInfo('Slow Operations', 'ops', queries.slowOperations.query(), { x: 0, y: 20, w: 12, h: 8 }),
     panels.timeSeries.GeneralInfo('Key Operations', 'ops', queries.keyOperations.query(), { x: 12, y: 20, w: 12, h: 8 }),
-    panels.timeSeries.GeneralInfo('Heartbeat Failures', 'short', queries.heartBeatFailure.query(), { x: 0, y: 28, w: 12, h: 8 }),
+    panels.timeSeries.generalCounter('Heartbeat Failures', 'short', queries.heartBeatFailure.query(), { x: 0, y: 28, w: 12, h: 8 }),
     panels.timeSeries.GeneralInfo('Compacted Keys', 'short', queries.compactedKeys.query(), { x: 12, y: 28, w: 12, h: 8 }),
   ]),