prod garage: add health check using admin api's '/health'

1 files changed, 64 insertions, 35 deletions

diff --git a/cluster/prod/app/garage/deploy/garage.hcl b/cluster/prod/app/garage/deploy/garage.hcl
index 26f76de..7ed963c 100644
--- a/cluster/prod/app/garage/deploy/garage.hcl
+++ b/cluster/prod/app/garage/deploy/garage.hcl
@@ -14,7 +14,7 @@ job "garage" {
       port "rpc" { static = 3901 }
       port "web" { static = 3902 }
       port "admin" { static = 3903 }
-      port "k2v" { static = 3904 } 
+      port "k2v" { static = 3904 }
     }
 
     update {
@@ -26,7 +26,6 @@ job "garage" {
     task "server" {
       driver = "docker"
       config {
-        advertise_ipv6_address = true
         image = "dxflrs/garage:v0.8.2"
         command = "/garage"
         args = [ "server" ]
@@ -70,6 +69,34 @@ job "garage" {
 
       kill_timeout = "20s"
 
+      restart {
+        interval = "30m"
+        attempts = 10
+        delay    = "15s"
+        mode     = "delay"
+      }
+
+      #### Configuration for service ports: admin port (internal use only)
+
+      service {
+        port = "admin"
+        address_mode = "host"
+        name = "garage-admin"
+        # Check that Garage is alive and answering TCP connections
+        check {
+          type = "tcp"
+          interval = "60s"
+          timeout = "5s"
+          check_restart {
+            limit = 3
+            grace = "90s"
+            ignore_warnings = false
+          }
+        }
+      }
+
+      #### Configuration for service ports: externally available ports (API, web)
+
       service {
         tags = [
           "garage_api",
@@ -77,13 +104,13 @@ job "garage" {
           "tricot *.garage.deuxfleurs.fr",
           "tricot-site-lb",
         ]
-        port = 3900
-        address_mode = "driver"
+        port = "s3"
+        address_mode = "host"
         name = "garage-api"
+        # Check 1: Garage is alive and answering TCP connections
         check {
+          name = "garage-api-live"
           type = "tcp"
-          port = 3900
-          address_mode = "driver"
           interval = "60s"
           timeout = "5s"
           check_restart {
@@ -92,6 +119,15 @@ job "garage" {
             ignore_warnings = false
           }
         }
+        # Check 2: Garage is in a healthy state and requests should be routed here
+        check {
+          name = "garage-api-healthy"
+          port = "admin"
+          type = "http"
+          path = "/health"
+          interval = "60s"
+          timeout = "5s"
+        }
       }
 
       service {
@@ -105,13 +141,13 @@ job "garage" {
             "tricot-add-header X-Content-Type-Options nosniff",
             "tricot-site-lb",
         ]
-        port = 3902
-        address_mode = "driver"
+        port = "web"
+        address_mode = "host"
         name = "garage-web"
+        # Check 1: Garage is alive and answering TCP connections
         check {
+          name = "garage-web-live"
           type = "tcp"
-          port = 3902
-          address_mode = "driver"
           interval = "60s"
           timeout = "5s"
           check_restart {
@@ -120,23 +156,14 @@ job "garage" {
             ignore_warnings = false
           }
         }
-      }
-
-      service {
-        port = 3903
-        address_mode = "driver"
-        name = "garage-admin"
+        # Check 2: Garage is in a healthy state and requests should be routed here
         check {
-          type = "tcp"
-          port = 3903
-          address_mode = "driver"
+          name = "garage-web-healthy"
+          port = "admin"
+          type = "http"
+          path = "/health"
           interval = "60s"
           timeout = "5s"
-          check_restart {
-            limit = 3
-            grace = "90s"
-            ignore_warnings = false
-          }
         }
       }
 
@@ -146,13 +173,13 @@ job "garage" {
           "tricot k2v.deuxfleurs.fr",
           "tricot-site-lb",
         ]
-        port = 3904
-        address_mode = "driver"
+        port = "k2v"
+        address_mode = "host"
         name = "garage-k2v"
+        # Check 1: Garage is alive and answering TCP connections
         check {
+          name = "garage-k2v-live"
           type = "tcp"
-          port = 3904
-          address_mode = "driver"
           interval = "60s"
           timeout = "5s"
           check_restart {
@@ -161,13 +188,15 @@ job "garage" {
             ignore_warnings = false
           }
         }
-      }
-
-      restart { 
-        interval = "30m"  
-        attempts = 10  
-        delay    = "15s"  
-        mode     = "delay"
+        # Check 2: Garage is in a healthy state and requests should be routed here
+        check {
+          name = "garage-k2v-healthy"
+          port = "admin"
+          type = "http"
+          path = "/health"
+          interval = "60s"
+          timeout = "5s"
+        }
       }
     }
   }