nixos/postgresql: implement auto-restart & rework dependencies of postgresql.target (#424625)

nixos/modules/services/databases/postgresql.nix

@@ -769,7 +769,7 @@ in
     systemd.targets.postgresql = {
       description = "PostgreSQL";
       wantedBy = [ "multi-user.target" ];
-      bindsTo = [
+      requires = [
         "postgresql.service"
         "postgresql-setup.service"
       ];
@@ -780,8 +780,13 @@ in
 
       after = [ "network.target" ];
 
-      # To trigger the .target also on "systemctl start postgresql".
-      bindsTo = [ "postgresql.target" ];
+      # To trigger the .target also on "systemctl start postgresql" as well as on
+      # restarts & stops.
+      # Please note that postgresql.service & postgresql.target binding to
+      # each other makes the Restart=always rule racy and results
+      # in sometimes the service not being restarted.
+      wants = [ "postgresql.target" ];
+      partOf = [ "postgresql.target" ];
 
       environment.PGDATA = cfg.dataDir;
 
@@ -821,6 +826,8 @@ in
 
           ExecStart = "${cfg.finalPackage}/bin/postgres";
 
+          Restart = "always";
+
           # Hardening
           CapabilityBoundingSet = [ "" ];
           DevicePolicy = "closed";
@@ -872,7 +879,20 @@ in
         })
       ];
 
-      unitConfig.RequiresMountsFor = "${cfg.dataDir}";
+      unitConfig =
+        let
+          inherit (config.systemd.services.postgresql.serviceConfig) TimeoutSec;
+          maxTries = 5;
+          bufferSec = 5;
+        in
+        {
+          RequiresMountsFor = "${cfg.dataDir}";
+
+          # The max. time needed to perform `maxTries` start attempts of systemd
+          # plus a bit of buffer time (bufferSec) on top.
+          StartLimitIntervalSec = TimeoutSec * maxTries + bufferSec;
+          StartLimitBurst = maxTries;
+        };
     };
 
     systemd.services.postgresql-setup = {

nixos/tests/postgresql/postgresql.nix

@@ -101,6 +101,12 @@ let
           machine.fail(check_count("SELECT * FROM sth;", 4))
           machine.succeed(check_count("SELECT xpath('/test/text()', doc) FROM xmltest;", 1))
 
+          with subtest("killing postgres process should trigger an automatic restart"):
+              machine.succeed("systemctl kill -s KILL postgresql")
+
+              machine.wait_until_succeeds("systemctl is-active postgresql.service")
+              machine.wait_until_succeeds("systemctl is-active postgresql.target")
+
           with subtest("Backup service works"):
               machine.succeed(
                   "systemctl start ${backupService}.service",