alerting for gateways and relays

backend/src/ee/services/gateway-v2/gateway-v2-dal.ts

@@ -10,11 +10,16 @@ export type TGatewayV2DALFactory = ReturnType<typeof gatewayV2DalFactory>;
 export const gatewayV2DalFactory = (db: TDbClient) => {
   const orm = ormify(db, TableName.GatewayV2);
 
-  const find = async (filter: TFindFilter<TGatewaysV2>, { offset, limit, sort, tx }: TFindOpt<TGatewaysV2> = {}) => {
+  const find = async (
+    filter: TFindFilter<TGatewaysV2> & { isHeartbeatStale?: boolean },
+    { offset, limit, sort, tx }: TFindOpt<TGatewaysV2> = {}
+  ) => {
     try {
+      const { isHeartbeatStale, ...regularFilter } = filter;
+
       const query = (tx || db.replicaNode())(TableName.GatewayV2)
         // eslint-disable-next-line @typescript-eslint/no-misused-promises
-        .where(buildFindFilter(filter, TableName.GatewayV2))
+        .where(buildFindFilter(regularFilter, TableName.GatewayV2))
         .join(TableName.Identity, `${TableName.Identity}.id`, `${TableName.GatewayV2}.identityId`)
         .join(
           TableName.IdentityOrgMembership,
@@ -24,6 +29,21 @@ export const gatewayV2DalFactory = (db: TDbClient) => {
         .select(selectAllTableCols(TableName.GatewayV2))
         .select(db.ref("name").withSchema(TableName.Identity).as("identityName"));
 
+      if (isHeartbeatStale) {
+        const oneHourAgo = new Date();
+        oneHourAgo.setHours(oneHourAgo.getHours() - 1);
+        void query.where(`${TableName.GatewayV2}.heartbeat`, "<", oneHourAgo);
+        void query.where((v) => {
+          void v
+            .whereNull(`${TableName.GatewayV2}.healthAlertedAt`)
+            .orWhere(
+              `${TableName.GatewayV2}.healthAlertedAt`,
+              "<",
+              db.ref("heartbeat").withSchema(TableName.GatewayV2)
+            );
+        });
+      }
+
       if (limit) void query.limit(limit);
       if (offset) void query.offset(offset);
       if (sort) {

backend/src/ee/services/gateway-v2/gateway-v2-service.ts

@@ -2,14 +2,16 @@ import net from "node:net";
 
 import { ForbiddenError } from "@casl/ability";
 import * as x509 from "@peculiar/x509";
+import { CronJob } from "cron";
 
-import { TRelays } from "@app/db/schemas";
+import { OrgMembershipRole, TRelays } from "@app/db/schemas";
 import { PgSqlLock } from "@app/keystore/keystore";
 import { crypto } from "@app/lib/crypto";
 import { DatabaseErrorCode } from "@app/lib/error-codes";
 import { BadRequestError, DatabaseError, NotFoundError } from "@app/lib/errors";
 import { GatewayProxyProtocol } from "@app/lib/gateway/types";
 import { withGatewayV2Proxy } from "@app/lib/gateway-v2/gateway-v2";
+import { logger } from "@app/lib/logger";
 import { OrgServiceActor } from "@app/lib/types";
 import { ActorAuthMethod, ActorType } from "@app/services/auth/auth-type";
 import { constructPemChainFromCerts } from "@app/services/certificate/certificate-fns";
@@ -20,6 +22,10 @@ import {
 } from "@app/services/certificate-authority/certificate-authority-fns";
 import { TKmsServiceFactory } from "@app/services/kms/kms-service";
 import { KmsDataKey } from "@app/services/kms/kms-types";
+import { TNotificationServiceFactory } from "@app/services/notification/notification-service";
+import { NotificationType } from "@app/services/notification/notification-types";
+import { TOrgDALFactory } from "@app/services/org/org-dal";
+import { SmtpTemplates, TSmtpService } from "@app/services/smtp/smtp-service";
 
 import { TLicenseServiceFactory } from "../license/license-service";
 import { PamResource } from "../pam-resource/pam-resource-enums";
@@ -39,6 +45,9 @@ type TGatewayV2ServiceFactoryDep = {
   gatewayV2DAL: TGatewayV2DALFactory;
   relayDAL: TRelayDALFactory;
   permissionService: TPermissionServiceFactory;
+  orgDAL: Pick<TOrgDALFactory, "findOrgMembersByRole">;
+  notificationService: Pick<TNotificationServiceFactory, "createUserNotifications">;
+  smtpService: Pick<TSmtpService, "sendMail">;
 };
 
 export type TGatewayV2ServiceFactory = ReturnType<typeof gatewayV2ServiceFactory>;
@@ -50,7 +59,10 @@ export const gatewayV2ServiceFactory = ({
   relayService,
   gatewayV2DAL,
   relayDAL,
-  permissionService
+  permissionService,
+  orgDAL,
+  notificationService,
+  smtpService
 }: TGatewayV2ServiceFactoryDep) => {
   const $validateIdentityAccessToGateway = async (orgId: string, actorId: string, actorAuthMethod: ActorAuthMethod) => {
     const orgLicensePlan = await licenseService.getPlan(orgId);
@@ -878,6 +890,78 @@ export const gatewayV2ServiceFactory = ({
     });
   };
 
+  const $healthcheckNotify = async () => {
+    const unhealthyGateways = await gatewayV2DAL.find({
+      isHeartbeatStale: true
+    });
+
+    if (unhealthyGateways.length === 0) return;
+
+    logger.warn(
+      { gatewayIds: unhealthyGateways.map((g) => g.id) },
+      "Found gateways with last heartbeat over an hour ago. Sending notifications."
+    );
+
+    await Promise.all(unhealthyGateways.map((gw) => gatewayV2DAL.updateById(gw.id, { healthAlertedAt: new Date() })));
+
+    const gatewaysByOrg = unhealthyGateways.reduce<Record<string, (typeof unhealthyGateways)[number][]>>((acc, gw) => {
+      if (!acc[gw.orgId]) {
+        acc[gw.orgId] = [];
+      }
+      acc[gw.orgId].push(gw);
+      return acc;
+    }, {});
+
+    for await (const [orgId, gateways] of Object.entries(gatewaysByOrg)) {
+      try {
+        const admins = await orgDAL.findOrgMembersByRole(orgId, OrgMembershipRole.Admin);
+        if (admins.length === 0) {
+          logger.warn({ orgId }, "Organization has no admins to notify about unhealthy gateway.");
+          // eslint-disable-next-line no-continue
+          continue;
+        }
+
+        const gatewayNames = gateways.map((g) => `"${g.name}"`).join(", ");
+        const body = `The following gateway(s) in your organization may be offline as they haven't reported a heartbeat in over an hour: ${gatewayNames}. Please check their status.`;
+
+        await notificationService.createUserNotifications(
+          admins.map((admin) => ({
+            userId: admin.user.id,
+            orgId,
+            type: NotificationType.GATEWAY_HEALTH_ALERT,
+            title: "Gateway Health Alert",
+            body,
+            link: "/organization/networking"
+          }))
+        );
+
+        await smtpService.sendMail({
+          recipients: admins.map((admin) => admin.user.email).filter((v): v is string => !!v),
+          subjectLine: "Gateway Health Alert",
+          substitutions: {
+            type: "gateway",
+            names: gatewayNames
+          },
+          template: SmtpTemplates.HealthAlert
+        });
+      } catch (error) {
+        logger.error(error, `Failed to send gateway health notifications for organization [orgId=${orgId}]`);
+      }
+    }
+  };
+
+  const initializeHealthcheckNotify = async () => {
+    logger.info("Setting up background notification process for gateway v2 health-checks");
+
+    await $healthcheckNotify();
+
+    // run every 5 minutes
+    const job = new CronJob("*/5 * * * *", $healthcheckNotify);
+    job.start();
+
+    return job;
+  };
+
   return {
     listGateways,
     registerGateway,
@@ -885,6 +969,7 @@ export const gatewayV2ServiceFactory = ({
     getPAMConnectionDetails,
     deleteGatewayById,
     heartbeat,
-    getPamSessionKey
+    getPamSessionKey,
+    initializeHealthcheckNotify
   };
 };

backend/src/ee/services/relay/relay-dal.ts

@@ -1,11 +1,47 @@
 import { TDbClient } from "@app/db";
-import { TableName } from "@app/db/schemas";
-import { ormify } from "@app/lib/knex";
+import { TableName, TRelays } from "@app/db/schemas";
+import { DatabaseError } from "@app/lib/errors";
+import { buildFindFilter, ormify, TFindFilter, TFindOpt } from "@app/lib/knex";
 
 export type TRelayDALFactory = ReturnType<typeof relayDalFactory>;
 
 export const relayDalFactory = (db: TDbClient) => {
   const orm = ormify(db, TableName.Relay);
 
-  return orm;
+  const find = async (
+    filter: TFindFilter<TRelays> & { isHeartbeatStale?: boolean },
+    { offset, limit, sort, tx }: TFindOpt<TRelays> = {}
+  ) => {
+    try {
+      const { isHeartbeatStale, ...regularFilter } = filter;
+
+      const query = (tx || db.replicaNode())(TableName.Relay)
+        // eslint-disable-next-line @typescript-eslint/no-misused-promises
+        .where(buildFindFilter(regularFilter, TableName.Relay));
+
+      if (isHeartbeatStale) {
+        const oneHourAgo = new Date();
+        oneHourAgo.setHours(oneHourAgo.getHours() - 1);
+        void query.where(`${TableName.Relay}.heartbeat`, "<", oneHourAgo);
+        void query.where((v) => {
+          void v
+            .whereNull(`${TableName.Relay}.healthAlertedAt`)
+            .orWhere(`${TableName.Relay}.healthAlertedAt`, "<", db.ref("heartbeat").withSchema(TableName.Relay));
+        });
+      }
+
+      if (limit) void query.limit(limit);
+      if (offset) void query.offset(offset);
+      if (sort) {
+        void query.orderBy(sort.map(([column, order, nulls]) => ({ column: column as string, order, nulls })));
+      }
+
+      const docs = await query;
+      return docs;
+    } catch (error) {
+      throw new DatabaseError({ error, name: `${TableName.Relay}: Find` });
+    }
+  };
+
+  return { ...orm, find };
 };

backend/src/ee/services/relay/relay-service.ts

@@ -2,12 +2,14 @@ import { isIP } from "node:net";
 
 import { ForbiddenError } from "@casl/ability";
 import * as x509 from "@peculiar/x509";
+import { CronJob } from "cron";
 
-import { TRelays } from "@app/db/schemas";
+import { OrgMembershipRole, TRelays } from "@app/db/schemas";
 import { PgSqlLock } from "@app/keystore/keystore";
 import { crypto } from "@app/lib/crypto";
 import { BadRequestError, ForbiddenRequestError, NotFoundError } from "@app/lib/errors";
 import { createRelayConnection } from "@app/lib/gateway-v2/gateway-v2";
+import { logger } from "@app/lib/logger";
 import { ActorAuthMethod, ActorType } from "@app/services/auth/auth-type";
 import { constructPemChainFromCerts, prependCertToPemChain } from "@app/services/certificate/certificate-fns";
 import { CertExtendedKeyUsage, CertKeyAlgorithm, CertKeyUsage } from "@app/services/certificate/certificate-types";
@@ -17,6 +19,11 @@ import {
 } from "@app/services/certificate-authority/certificate-authority-fns";
 import { TKmsServiceFactory } from "@app/services/kms/kms-service";
 import { KmsDataKey } from "@app/services/kms/kms-types";
+import { TNotificationServiceFactory } from "@app/services/notification/notification-service";
+import { NotificationType } from "@app/services/notification/notification-types";
+import { TOrgDALFactory } from "@app/services/org/org-dal";
+import { SmtpTemplates, TSmtpService } from "@app/services/smtp/smtp-service";
+import { TUserDALFactory } from "@app/services/user/user-dal";
 
 import { verifyHostInputValidity } from "../dynamic-secret/dynamic-secret-fns";
 import { TLicenseServiceFactory } from "../license/license-service";
@@ -40,7 +47,11 @@ export const relayServiceFactory = ({
   relayDAL,
   kmsService,
   licenseService,
-  permissionService
+  permissionService,
+  orgDAL,
+  notificationService,
+  smtpService,
+  userDAL
 }: {
   instanceRelayConfigDAL: TInstanceRelayConfigDALFactory;
   orgRelayConfigDAL: TOrgRelayConfigDALFactory;
@@ -48,6 +59,10 @@ export const relayServiceFactory = ({
   kmsService: TKmsServiceFactory;
   licenseService: TLicenseServiceFactory;
   permissionService: TPermissionServiceFactory;
+  orgDAL: Pick<TOrgDALFactory, "findOrgMembersByRole">;
+  notificationService: Pick<TNotificationServiceFactory, "createUserNotifications">;
+  smtpService: Pick<TSmtpService, "sendMail">;
+  userDAL: Pick<TUserDALFactory, "find">;
 }) => {
   const $getInstanceCAs = async () => {
     const instanceConfig = await instanceRelayConfigDAL.transaction(async (tx) => {
@@ -1193,12 +1208,108 @@ export const relayServiceFactory = ({
     return deletedRelay;
   };
 
+  const $healthcheckNotify = async () => {
+    const oneHourAgo = new Date();
+    oneHourAgo.setHours(oneHourAgo.getHours() - 1);
+    const unhealthyRelays = await relayDAL.find({
+      isHeartbeatStale: true
+    });
+
+    if (unhealthyRelays.length === 0) return;
+
+    logger.warn(
+      { relayIds: unhealthyRelays.map((g) => g.id) },
+      "Found relays with last heartbeat over an hour ago. Sending notifications."
+    );
+
+    await Promise.all(unhealthyRelays.map((r) => relayDAL.updateById(r.id, { healthAlertedAt: new Date() })));
+
+    const relaysByOrg = unhealthyRelays.reduce<Record<string, TRelays[]>>((acc, r) => {
+      const key = r.orgId ?? "instance";
+      if (!acc[key]) {
+        acc[key] = [];
+      }
+      acc[key].push(r);
+      return acc;
+    }, {});
+
+    for await (const [orgId, relays] of Object.entries(relaysByOrg)) {
+      try {
+        if (orgId === "instance") {
+          const superAdmins = await userDAL.find({
+            superAdmin: true
+          });
+
+          const recipients = superAdmins.map((admin) => admin.email).filter((v): v is string => !!v);
+
+          if (recipients.length > 0) {
+            const relayNames = relays.map((r) => `"${r.name}"`).join(", ");
+            await smtpService.sendMail({
+              recipients,
+              subjectLine: "Relay Health Alert",
+              substitutions: {
+                type: "relay",
+                names: relayNames
+              },
+              template: SmtpTemplates.HealthAlert
+            });
+          }
+        } else {
+          const admins = await orgDAL.findOrgMembersByRole(orgId, OrgMembershipRole.Admin);
+          if (admins.length === 0) {
+            // eslint-disable-next-line no-continue
+            continue;
+          }
+
+          const relayNames = relays.map((r) => `"${r.name}"`).join(", ");
+          const body = `The following relay(s) in your organization may be offline as they haven't reported a heartbeat in over an hour: ${relayNames}. Please check their status.`;
+
+          await notificationService.createUserNotifications(
+            admins.map((admin) => ({
+              userId: admin.user.id,
+              orgId,
+              type: NotificationType.RELAY_HEALTH_ALERT,
+              title: "Relay Health Alert",
+              body,
+              link: "/organization/networking"
+            }))
+          );
+
+          await smtpService.sendMail({
+            recipients: admins.map((admin) => admin.user.email).filter((v): v is string => !!v),
+            subjectLine: "Relay Health Alert",
+            substitutions: {
+              type: "relay",
+              names: relayNames
+            },
+            template: SmtpTemplates.HealthAlert
+          });
+        }
+      } catch (error) {
+        logger.error(error, `Failed to send relay health notifications for organization [orgId=${orgId}]`);
+      }
+    }
+  };
+
+  const initializeHealthcheckNotify = async () => {
+    logger.info("Setting up background notification process for relay health-checks");
+
+    await $healthcheckNotify();
+
+    // run every 5 minutes
+    const job = new CronJob("*/5 * * * *", $healthcheckNotify);
+    job.start();
+
+    return job;
+  };
+
   return {
     registerRelay,
     getCredentialsForGateway,
     getCredentialsForClient,
     getRelays,
     deleteRelay,
-    heartbeat
+    heartbeat,
+    initializeHealthcheckNotify
   };
 };