Fix DB connection timeouts in k8s with aggressive TCP keepalive (#92)

SSL connections between pods and PostgreSQL are silently dying,
causing 15s checkout timeouts and cascading Oban failures.

- Set Linux TCP keepalive: 15s idle, 5s interval, 3 probes (30s detection)
- Add PostgreSQL statement_timeout (30s) to kill hung queries server-side
- Add disconnect_on_error_codes for fast recovery from PG shutdowns
- Bump default pool_size 20 → 30 to reduce contention under load

Reviewed-on: graham/towerops-web#92

config/runtime.exs

@@ -148,6 +148,26 @@ if config_env() == :prod do
   # OBAN_QUEUE_SCALE divides all queue sizes (e.g. "5" = 1/5th capacity for staging)
   oban_scale = String.to_integer(System.get_env("OBAN_QUEUE_SCALE") || "1")
 
+  # Aggressive TCP keepalive for Kubernetes environments.
+  # Default OS keepalive is ~2 hours on Linux, far too slow for k8s where
+  # load balancers and proxies silently drop idle connections within minutes.
+  # Linux raw socket options (SOL_TCP=6):
+  #   TCP_KEEPIDLE  (4) = 15s  - seconds idle before first keepalive probe
+  #   TCP_KEEPINTVL (5) = 5s   - seconds between subsequent probes
+  #   TCP_KEEPCNT   (6) = 3    - failed probes before connection is dead
+  # Total dead-connection detection: 15 + (5 * 3) = 30 seconds
+  tcp_keepalive_opts =
+    if System.get_env("KUBERNETES_SERVICE_HOST") do
+      [
+        {:keepalive, true},
+        {:raw, 6, 4, <<15::native-32>>},
+        {:raw, 6, 5, <<5::native-32>>},
+        {:raw, 6, 6, <<3::native-32>>}
+      ]
+    else
+      [keepalive: true]
+    end
+
   config :libcluster, topologies: libcluster_topologies
 
   # Configure Swoosh to use Req for API requests
@@ -306,14 +326,19 @@ if config_env() == :prod do
   config :towerops, Towerops.Repo,
     ssl: ssl_config,
     url: database_url,
-    pool_size: String.to_integer(System.get_env("POOL_SIZE") || "20"),
+    pool_size: String.to_integer(System.get_env("POOL_SIZE") || "30"),
     # Detect stale connections faster: if a query takes longer than queue_target (2s)
     # on average over queue_interval (5s), DBConnection will disconnect idle connections
     # to cycle in fresh ones. Helps recover from "ssl recv: closed" errors.
     queue_target: 2_000,
     queue_interval: 5_000,
-    # TCP keepalive prevents proxies/firewalls from dropping idle connections
-    socket_options: maybe_ipv6 ++ [keepalive: true]
+    # PostgreSQL server-side statement timeout: kill any query running longer than 30s.
+    # Prevents queries from hanging indefinitely on dead SSL connections.
+    parameters: [application_name: "towerops", statement_timeout: "30000"],
+    # Disconnect immediately on fatal PostgreSQL error codes for faster pool recovery:
+    #   57P01 = admin_shutdown, 57P02 = crash_shutdown, 57P03 = cannot_connect_now
+    disconnect_on_error_codes: [:"57P01", :"57P02", :"57P03"],
+    socket_options: maybe_ipv6 ++ tcp_keepalive_opts
 
   config :towerops, Towerops.Vault,
     ciphers: [