fix(oban): use DynamicLifeline in prod; add /health/live + split k8s probes

Prod paired the Oban Pro Smart engine + unique workers with core
Oban.Plugins.Lifeline, whose rescue has no unique_violation handling. An
orphaned executing CheckExecutorWorker job rescued back to available collides
with its already-scheduled successor on oban_jobs_unique_index (23505),
crashing the plugin every 60s and leaving ~150 orphans stuck. Switch prod to
Oban.Pro.Plugins.DynamicLifeline, which repairs the conflict via
Smart.clear_uniq_violation. Dev keeps core Lifeline (Basic engine).

Add a shallow /health/live endpoint that does not touch db/redis. k8s liveness
and startup probes now target it so a transient dependency outage can't kill or
block boot of a healthy pod; readiness keeps the deep /health to gate the load
balancer. (deployment.yaml probe/replica change pushed separately, after the
image carrying /health/live is live.)

Also drop unused POSTGRES_* keys from the secrets example.

CHANGELOG.txt

@@ -1,3 +1,26 @@
+2026-05-20
+fix(oban): use Oban Pro DynamicLifeline in prod to stop unique_violation crash loop
+  CheckExecutorWorker's `unique` constraint covers available/scheduled/retryable
+  (executing excluded so perform/1 can self-schedule). When a pod is killed
+  mid-execution the job is orphaned in `executing`; core Oban.Plugins.Lifeline
+  rescues it back to `available`, where it collides with the already-scheduled
+  successor (same uniq_key) on oban_jobs_unique_index, raising 23505. Core
+  Lifeline has no unique_violation handling, so the plugin crashed every 60s and
+  ~150 orphans never cleared. Oban Pro's DynamicLifeline (the Smart-engine
+  plugin) catches the 23505 via Smart.clear_uniq_violation and repairs it.
+  Prod only (runtime.exs, Smart engine); dev keeps core Lifeline (Basic engine).
+  DynamicLifeline rescues by producer records, so rescue_after was dropped.
+  Files: config/runtime.exs, config/dev.exs
+
+feat(health): add shallow /health/live liveness endpoint and split k8s probes
+  /health/live returns 200 without querying db/redis. k8s liveness and startup
+  probes now target /health/live so a transient db/redis outage can no longer
+  kill or block boot of an otherwise-healthy pod; readiness keeps the deep
+  /health (db+redis) to gate load-balancer membership only. Deployment bumped to
+  2 replicas (NFS data volume supports multi-pod mounts; app already clusters).
+  Files: lib/towerops_web/router.ex, lib/towerops_web/controllers/health_controller.ex,
+  test/towerops_web/controllers/health_controller_test.exs, k8s/deployment.yaml
+
 2026-05-14
 fix(monitoring): add Oban unique constraint to CheckExecutorWorker to stop duplicate job buildup
   CheckExecutorWorker had no `unique` constraint. Every insert path —

config/dev.exs

@@ -119,7 +119,10 @@ config :towerops, Oban,
     {Oban.Plugins.Pruner, max_age: 60},
     # Rescue orphaned jobs (when node crashes)
     {Oban.Plugins.Reindexer, schedule: "0 2 * * *"},
-    # Rescue jobs stuck in executing state (mark as cancelled after 10 minutes)
+    # Dev uses the core Basic engine (the Pro Smart engine is prod-only, see
+    # runtime.exs), so core Lifeline is correct here. Prod uses
+    # Oban.Pro.Plugins.DynamicLifeline because core Lifeline crashes on the
+    # unique_violation that the Smart engine's unique index can raise on rescue.
     {Oban.Plugins.Lifeline, rescue_after: to_timeout(minute: 10)}
   ]
 

config/runtime.exs

@@ -336,8 +336,13 @@ if config_env() == :prod do
          "oban_jobs_args_index",
          "oban_jobs_meta_index"
        ]},
-      # Rescue jobs stuck in executing state (mark as cancelled after 10 minutes)
+      # Rescue jobs orphaned in the executing state. DynamicLifeline (not core
-      {Oban.Plugins.Lifeline, rescue_after: to_timeout(minute: 10)}
+      # Lifeline) is required with the Pro Smart engine + unique workers: it
+      # catches the unique_violation that occurs when a rescued orphan collides
+      # with an already-scheduled successor (same uniq_key) and repairs it,
+      # whereas core Lifeline crashes on that 23505. It rescues by producer
+      # records rather than a time threshold, so rescue_after does not apply.
+      Oban.Pro.Plugins.DynamicLifeline
     ]
 
   # ## SSL Support

k8s/secrets.example.yaml

@@ -40,13 +40,8 @@ metadata:
   namespace: towerops
 type: Opaque
 stringData:
-  # Full ecto:// URL — what runtime.exs reads.
+  # Full ecto:// URL — the only DB var the app reads (runtime.exs).
   DATABASE_URL: "postgresql://towerops:CHANGE_ME@postgres-host:5432/towerops"
-  POSTGRES_HOST: "CHANGE_ME_postgres_host"
-  POSTGRES_PORT: "5432"
-  POSTGRES_DB: "towerops"
-  POSTGRES_USER: "towerops"
-  POSTGRES_PASSWORD: "CHANGE_ME"
 ---
 apiVersion: v1
 kind: Secret

lib/towerops_web/controllers/health_controller.ex

@@ -6,6 +6,17 @@ defmodule ToweropsWeb.HealthController do
   alias Ecto.Adapters.SQL
   alias Towerops.Repo
 
+  # Liveness probe: confirms the BEAM/web stack is responsive. Deliberately does
+  # NOT touch the database or redis — liveness failure causes k8s to kill the
+  # pod, so it must not react to transient dependency outages. Dependency health
+  # belongs in readiness (index/2), which only removes the pod from the load
+  # balancer.
+  def live(conn, _params) do
+    conn
+    |> put_resp_content_type("application/json")
+    |> send_resp(200, Jason.encode!(%{status: "ok"}))
+  end
+
   def index(conn, _params) do
     require Logger
 

lib/towerops_web/router.ex

@@ -66,9 +66,13 @@ defmodule ToweropsWeb.Router do
     plug ToweropsWeb.GraphQL.Context
   end
 
-  # Health check endpoint for Kubernetes probes (no authentication required)
+  # Health check endpoints for Kubernetes probes (no authentication required).
+  # /health is a deep check (db + redis) used for readiness/startup; /health/live
+  # is a shallow liveness check that must not depend on external services, so a
+  # transient db/redis blip can't make k8s kill an otherwise-healthy pod.
   scope "/", ToweropsWeb do
     get "/health", HealthController, :index, log: false
+    get "/health/live", HealthController, :live, log: false
   end
 
   # Public crawlable endpoints (no pipeline, no CSRF/session required)

priv/static/changelog.txt

@@ -1,3 +1,7 @@
+2026-05-20 — Reliability improvements
+* The app now rides through brief database hiccups instead of restarting, reducing momentary unavailability
+* Background monitoring jobs recover cleanly after a restart instead of getting stuck
+
 2026-05-14 — Monitoring check reliability fix
 * Fixed an issue where monitoring checks could build up a large backlog of duplicate scheduled runs, delaying check execution
 * Checks now stay on schedule instead of falling behind

test/towerops_web/controllers/health_controller_test.exs

@@ -1,6 +1,33 @@
 defmodule ToweropsWeb.HealthControllerTest do
   use ToweropsWeb.ConnCase, async: false
 
+  describe "GET /health/live" do
+    test "returns 200 with status ok", %{conn: conn} do
+      conn = get(conn, "/health/live")
+      result = Jason.decode!(response(conn, 200))
+      assert result["status"] == "ok"
+    end
+
+    test "stays 200 even when redis is configured but unreachable", %{conn: conn} do
+      # Liveness must be decoupled from dependencies: an unreachable redis (or
+      # db) must NOT fail liveness, otherwise k8s kills a pod that is merely
+      # waiting on a transient dependency blip.
+      previous = Application.get_env(:towerops, :redis)
+      Application.put_env(:towerops, :redis, host: "198.51.100.1", port: 1, timeout: 50)
+
+      try do
+        conn = get(conn, "/health/live")
+        assert response(conn, 200)
+      after
+        if previous do
+          Application.put_env(:towerops, :redis, previous)
+        else
+          Application.delete_env(:towerops, :redis)
+        end
+      end
+    end
+  end
+
   describe "GET /health" do
     test "returns 200 when database is connected", %{conn: conn} do
       conn = get(conn, "/health")