ADD: Obol Alerts (#2018)

* ADD: Alert Rules * CHANGE: name of alert files * ADD: Function for Obol Alerts in Stereum * FIX: Correct alarm filtering logic in alert boxes (#2017) * FIX: update alarm filter in node page * FIX: update alarm filter in control page * ADD: Alert Rules * CHANGE: name of alert files * ADD: Function for Obol Alerts in Stereum * Fix: the control alert with obol alerts * fix: tooltip of yellow obol alert * FIX: now filter algo for the alarm controls * IMPROVE: error handling * FIX: new alarm filter algo in the node alert box * FIX: wire obol warning in the node alert box --------- Co-authored-by: mabasian <[email protected]>
stereum-dev · Sep 3, 2024 · f17d1a9 · f17d1a9
1 parent ff4a058
commit f17d1a9
Show file tree

Hide file tree

Showing 10 changed files with 1,080 additions and 91 deletions.
diff --git a/controls/roles/manage-service/files/grafana/CharonService/alerting/CharonAlerts.yaml b/controls/roles/manage-service/files/grafana/CharonService/alerting/CharonAlerts.yaml
diff --git a/...sNodeExporterService/alerting/alerts.yaml → ...orterService/alerting/HardwareAlerts.yaml b/...sNodeExporterService/alerting/alerts.yaml → ...orterService/alerting/HardwareAlerts.yaml
diff --git a/launcher/public/output.css b/launcher/public/output.css
@@ -1800,10 +1800,6 @@ video {
   height: 50px;
 }
 
-.h-\[53px\]{
-  height: 53px;
-}
-
 .h-\[554px\]{
   height: 554px;
 }
@@ -5017,11 +5013,6 @@ video {
   --tw-shadow: var(--tw-shadow-colored);
 }
 
-.shadow-zinc-700{
-  --tw-shadow-color: #3f3f46;
-  --tw-shadow: var(--tw-shadow-colored);
-}
-
 .shadow-zinc-800{
   --tw-shadow-color: #27272a;
   --tw-shadow: var(--tw-shadow-colored);
@@ -5626,11 +5617,6 @@ html body {
   background-color: rgb(43 48 52 / var(--tw-bg-opacity));
 }
 
-.hover\:bg-\[\#2f5a50\]:hover{
-  --tw-bg-opacity: 1;
-  background-color: rgb(47 90 80 / var(--tw-bg-opacity));
-}
-
 .hover\:bg-\[\#325e5a\]:hover{
   --tw-bg-opacity: 1;
   background-color: rgb(50 94 90 / var(--tw-bg-opacity));

diff --git a/launcher/src/backend/Monitoring.js b/launcher/src/backend/Monitoring.js
@@ -436,11 +436,11 @@ export class Monitoring {
     var query =
       rpc_method.trim().indexOf("{") < 0
         ? JSON.stringify({
-            jsonrpc: "2.0",
-            method: rpc_method.trim(),
-            params: rpc_params,
-            id: 1,
-          })
+          jsonrpc: "2.0",
+          method: rpc_method.trim(),
+          params: rpc_params,
+          id: 1,
+        })
         : rpc_method;
 
     // Define default response
@@ -2623,8 +2623,8 @@ export class Monitoring {
     const addr_type = Array.isArray(addr)
       ? "arr"
       : typeof addr === "string" && ["public", "local"].includes(addr)
-      ? "str"
-      : "invalid";
+        ? "str"
+        : "invalid";
     addr = addr_type == "str" ? addr.toLowerCase().trim() : addr;
     if (addr_type == "invalid") {
       return {
@@ -2712,7 +2712,7 @@ export class Monitoring {
     for (let i = 0; i < serviceInfos.length; i++) {
       const hashDependencies =
         serviceInfos[i].config.dependencies.consensusClients.length ||
-        serviceInfos[i].config.dependencies.executionClients.length
+          serviceInfos[i].config.dependencies.executionClients.length
           ? "yes"
           : "no";
       easyInfos.push({
@@ -3249,9 +3249,8 @@ rm -rf diskoutput
         const parsedJson = JSON.parse(stdoutJson);
 
         let message =
-          `${parsedJson?.message || ""}${parsedJson?.message && parsedJson?.stacktraces ? "\n" : ""}${
-            parsedJson?.stacktraces || ""
-          }`.trim() || output;
+          `${parsedJson?.message || ""}${parsedJson?.message && parsedJson?.stacktraces ? "\n" : ""}${parsedJson?.stacktraces || ""
+            }`.trim() || output;
 
         return {
           pubkey: pubkey,
@@ -3341,4 +3340,113 @@ rm -rf diskoutput
       ];
     }
   }
+  /**
+   * Will gather metrics from Prometheus and evaluate.
+   * If thresholds are exceeded, an alert will be generated and added to the retuned array.
+   * @returns {Object[]} Array of alerts e.g. [{name: "Cluster in Unknown Status", level: "warning"}, {name: "Beacon Node Down", level: "critical"}]
+   */
+  async fetchObolCharonAlerts() {
+    try {
+      const serviceInfos = await this.getServiceInfos("CharonService");
+      if (serviceInfos.length < 1) {
+        return [];
+      }
+      const queries = {
+        app_monitoring_readyz: "max((app_monitoring_readyz)) by (cluster_name, cluster_hash, cluster_peer)",
+        cluster_missed_attestations: "max(increase(core_tracker_failed_duties_total[10m])) by (cluster_hash, cluster_name)",
+        cluster_failure_rate: "floor(100 * (max(increase(core_tracker_success_duties_total[15m])) by (cluster_hash, cluster_name) / max(increase(core_tracker_expect_duties_total[15m])) by (cluster_hash, cluster_name)))",
+        percentage_failed_sync_message_duty: "(\n    sum(increase(core_tracker_failed_duties_total[1h])) by (cluster_name,cluster_hash,cluster_peer)\n) \n/ \n(\n    sum(increase(core_tracker_failed_duties_total[1h])) by (cluster_name,cluster_hash,cluster_peer) \n    + \n    sum(increase(core_bcast_broadcast_total[1h])) by (cluster_name,cluster_hash,cluster_peer) \n)",
+        connected_relays: "group (p2p_relay_connections) by (cluster_peer)",
+        peer_ping_latency: "histogram_quantile(0.90, sum(rate(p2p_ping_latency_secs_bucket[2m])) by (le,peer))",
+      }
+
+      const queryPromises = Object.entries(queries).map(([key, query]) => {
+        return this.queryPrometheus(encodeURIComponent(query)).then(result => ({ key, result }));
+      });
+
+      const results = await Promise.all(queryPromises);
+
+      let alerts = results.map((metric) => {
+        if (metric.result.status != "success") {
+          return;
+        }
+        if (metric.key === "peer_ping_latency") {
+          let value = Math.max(...metric.result.data.result.map((r) => r.value[1]));
+          return this.parseObolCharonAlerts(metric.key, value);
+        }
+        let value = metric.result.data.result[0].value[1];
+        return this.parseObolCharonAlerts(metric.key, value);
+      }).filter((alert) => alert);
+
+      return alerts;
+
+    } catch (error) {
+      log.error("Fetching Obol Charon Alerts Failed:\n" + error);
+      return []
+    }
+  }
+
+  parseObolCharonAlerts(key, value) {
+    value = 0
+    //app_monitoring_readyz
+    if (key === "app_monitoring_readyz") {
+      switch (value) {
+        case 0:
+          return {
+            name: "Cluster in Unknown Status",
+            level: "warning"
+          };
+        case 2:
+          return {
+            name: "Beacon Node Down",
+            level: "critical"
+          };
+        case 4:
+          return {
+            name: "Cluster Insufficient Peers",
+            level: "warning"
+          };
+        case 6:
+          return {
+            name: "Cluster Missing Validators",
+            level: "critical"
+          };
+        case 7:
+          return {
+            name: "Beacon Node Zero Peers",
+            level: "critical"
+          };
+      }
+    }
+    if (key === "cluster_missed_attestations" && value > 0) {
+      return {
+        name: "Cluster Missed Attestations",
+        level: "critical"
+      };
+    }
+    if (key === "cluster_failure_rate" && value < 95) {
+      return {
+        name: "Cluster Failure Rate",
+        level: "critical"
+      };
+    }
+    if (key === "percentage_failed_sync_message_duty" && value > 0.1) {
+      return {
+        name: "Failed Sync Msg Duty",
+        level: "critical"
+      };
+    }
+    if (key === "connected_relays" && value < 1) {
+      return {
+        name: "Num. Connected Relays",
+        level: "warning"
+      };
+    }
+    if (key === "peer_ping_latency" && value > 0.4) {
+      return {
+        name: "Peer Ping Latency",
+        level: "warning"
+      };
+    }
+  }
 }
diff --git a/launcher/src/backend/SSHService.js b/launcher/src/backend/SSHService.js
@@ -571,7 +571,10 @@ export class SSHService {
    * @param {Client} [conn]
    * @returns `void`
    */
-  async uploadFileSSH(localPath, remotePath, conn = this.getConnectionFromPool()) {
+  async uploadFileSSH(localPath, remotePath, conn) {
+    if (!conn) {
+      conn = await this.getConnectionFromPool();
+    }
     return new Promise((resolve, reject) => {
       const readStream = fs.createReadStream(localPath);
       readStream.on("error", reject);
@@ -621,7 +624,7 @@ export class SSHService {
         if (item.isDirectory()) {
           await this.uploadDirectorySSH(localFilePath, remoteFilePath, conn);
         } else {
-          await this.uploadFileSSH(localFilePath, remoteFilePath);
+          await this.uploadFileSSH(localFilePath, remoteFilePath, conn);
         }
       }
       return true;

diff --git a/launcher/src/backend/ValidatorAccountManager.js b/launcher/src/backend/ValidatorAccountManager.js
@@ -1127,7 +1127,7 @@ export class ValidatorAccountManager {
       this.nodeConnection.sshService.exec(`rm -rf ${dataDir}`);
       const result = await this.nodeConnection.sshService.uploadDirectorySSH(path.normalize(localPath), dataDir);
       if (result) {
-        log.info("Obol Backup downloaded from: ", localPath);
+        log.info("Obol Backup uploaded from: ", localPath);
       }
     } catch (err) {
       log.error("Error uploading Obol Backup: ", err);

diff --git a/launcher/src/background.js b/launcher/src/background.js
@@ -741,6 +741,10 @@ ipcMain.handle("readGasConfigFile", async (event, args) => {
   return await tekuGasLimitConfig.readGasConfigFile(args);
 });
 
+ipcMain.handle("fetchObolCharonAlerts", async () => {
+  return await monitoring.fetchObolCharonAlerts();
+});
+
 // Scheme must be registered before the app is ready
 protocol.registerSchemesAsPrivileged([{ scheme: "app", privileges: { secure: true, standard: true } }]);