garage: harmonize staging and prod (checks, services)

1 files changed, 68 insertions, 34 deletions

diff --git a/cluster/staging/app/garage/deploy/garage.hcl b/cluster/staging/app/garage/deploy/garage.hcl
index 7a7b44b..b8e7227 100644
--- a/cluster/staging/app/garage/deploy/garage.hcl
+++ b/cluster/staging/app/garage/deploy/garage.hcl
@@ -73,12 +73,42 @@ job "garage-staging" {
       kill_signal = "SIGINT"
       kill_timeout = "20s"
 
+      restart {
+        interval = "5m"
+        attempts = 10
+        delay    = "1m"
+        mode     = "delay"
+      }
+
       service {
         name = "garage-staging-rpc"
         tags = ["garage-staging-rpc"]
         port = "rpc"
       }
 
+      #### Configuration for service ports: admin port (internal use only)
+
+      service {
+        name = "garage-staging-admin"
+        tags = [
+          "garage-staging-admin",
+        ]
+        port = "admin"
+        check {
+          name = "garage-tcp-liveness-check"
+          type = "tcp"
+          interval = "60s"
+          timeout = "5s"
+          check_restart {
+            limit = 3
+            grace = "90s"
+            ignore_warnings = false
+          }
+        }
+      }
+
+      #### Configuration for service ports: externally available ports (S3 API, K2V, web)
+
       service {
         name = "garage-staging-s3-api"
         tags = [
@@ -90,7 +120,21 @@ job "garage-staging" {
           "tricot-site-lb",
         ]
         port = "s3"
+        # Check 1: Garage is alive and answering TCP connections
+        check {
+          name = "garage-staging-api-live"
+          type = "tcp"
+          interval = "60s"
+          timeout = "5s"
+          check_restart {
+            limit = 3
+            grace = "90s"
+            ignore_warnings = false
+          }
+        }
+        # Check 2: Garage is in a healthy state and requests should be routed here
         check {
+          name = "garage-staging-api-healthy"
           port = "admin"
           type = "http"
           path = "/health"
@@ -108,7 +152,21 @@ job "garage-staging" {
           "tricot-site-lb",
         ]
         port = "k2v"
+        # Check 1: Garage is alive and answering TCP connections
+        check {
+          name = "garage-staging-k2v-live"
+          type = "tcp"
+          interval = "60s"
+          timeout = "5s"
+          check_restart {
+            limit = 3
+            grace = "90s"
+            ignore_warnings = false
+          }
+        }
+        # Check 2: Garage is in a healthy state and requests should be routed here
         check {
+          name = "garage-staging-k2v-healthy"
           port = "admin"
           type = "http"
           path = "/health"
@@ -134,52 +192,28 @@ job "garage-staging" {
             "tricot-site-lb",
         ]
         port = "web"
+        # Check 1: Garage is alive and answering TCP connections
         check {
-          port = "admin"
-          type = "http"
-          path = "/health"
-          interval = "60s"
-          timeout = "5s"
-        }
-      }
-
-      service {
-        name = "garage-staging-admin"
-        tags = [
-          "garage-staging-admin",
-        ]
-        port = "admin"
-        check {
-          name = "garage-admin-health-check"
-          type = "http"
-          path = "/health"
+          name = "garage-staging-web-live"
+          type = "tcp"
           interval = "60s"
           timeout = "5s"
           check_restart {
-            limit = 10
+            limit = 3
             grace = "90s"
-            ignore_warnings = true
+            ignore_warnings = false
           }
         }
+        # Check 2: Garage is in a healthy state and requests should be routed here
         check {
-          name = "garage-tcp-liveness-check"
-          type = "tcp"
+          name = "garage-staging-web-healthy"
+          port = "admin"
+          type = "http"
+          path = "/health"
           interval = "60s"
           timeout = "5s"
-          check_restart {
-            limit = 3
-            grace = "90s"
-            ignore_warnings = true
-          }
         }
       }
-
-      restart {
-        interval = "5m"
-        attempts = 10
-        delay    = "1m"
-        mode     = "delay"
-      }
     }
   }
 }