fix(survey): make the worker module work on a fresh environment (#703)

apps/survey.immich.app/backend/src/durable-objects/index.ts

@@ -1,11 +1,5 @@
 export { SurveyDO } from './survey-do';
 
-// Legacy alias for the original KV-backed class (see wrangler-do.jsonc v1
-// migration: `new_classes: ["SurveySession"]`). Cloudflare rejects a deploy
-// that removes a class still declared in the migration chain, so the alias
-// stays exported until a future `deleted_classes` migration retires v1.
-export { SurveyDO as SurveySession } from './survey-do';
-
 export default {
   async fetch(): Promise<Response> {
     return new Response('This worker only hosts Durable Objects', { status: 404 });

apps/survey.immich.app/backend/src/durable-objects/survey-do.ts

@@ -451,6 +451,13 @@ export class SurveyDO extends DurableObject {
       answers: Array<{ questionId: string; value: string; otherText?: string; answerMs?: number }>;
     };
     await this.respondentService.submitBatch(survey.slug!, respondentId, answers, survey);
+
+    // The unload beacon submits over HTTP even in WebSocket mode, so without this
+    // the cached choice answers miss whatever was flushed on the way out and the
+    // live tallies under-count once the respondent completes.
+    for (const a of answers) {
+      this.cache.setAnswer(respondentId, a.questionId, a.value, a.otherText ?? null);
+    }
     return new Response(null, { status: 204 });
   }
 

apps/survey.immich.app/backend/src/durable-objects/ws/ws-handler.ts

@@ -262,10 +262,6 @@ async function handleOp(
         if (error) throw new ServiceError(error, 400);
       }
 
-      for (const a of answers) {
-        cache.setAnswer(respondentId, a.questionId, a.value, a.otherText ?? null);
-      }
-
       // answer_ms is clamped via the shared helper so it stays in lockstep with the
       // HTTP path in respondent.service.
       const now = new Date().toISOString();
@@ -278,6 +274,13 @@ async function handleOp(
         `INSERT OR REPLACE INTO answers (respondent_id, question_id, answer, other_text, answered_at, answer_ms) VALUES ${placeholders}`,
         ...values,
       );
+
+      // Cache only after the write lands: a throw here leaves answers in
+      // choiceAnswers that were never persisted, and completion would fold them
+      // into the live tallies.
+      for (const a of answers) {
+        cache.setAnswer(respondentId, a.questionId, a.value, a.otherText ?? null);
+      }
       return {};
     }
 

apps/survey.immich.app/backend/wrangler-dev.jsonc

@@ -18,9 +18,6 @@
       },
     ],
   },
-  // SurveyDO is the only Durable Object class and is SQLite-backed from the
-  // start. (The earlier v1 "SurveySession" class never existed — it only
-  // produced a "no such Durable Object class is exported" warning on boot.)
   "migrations": [
     {
       "tag": "v1",

apps/survey.immich.app/backend/wrangler-do.jsonc

@@ -11,19 +11,12 @@
       },
     ],
   },
-  // The deployed sessions worker reached tag "v2" through this chain, so it must
-  // be preserved as-is: v1 created the original "SurveySession" class and v2
-  // introduced the SQLite-backed "SurveyDO". The Terraform deploy (worker.tf)
-  // asserts old_tag == new_tag == "v2" against this state. `durable-objects/
-  // index.ts` still exports `SurveyDO as SurveySession` so the v1 class name
-  // remains resolvable.
+  // This config only drives `wrangler build`; Terraform deploys the worker and
+  // applies the migration itself, so this must stay in sync with the block in
+  // deployment/modules/cloudflare/workers/survey/worker.tf.
   "migrations": [
     {
       "tag": "v1",
-      "new_classes": ["SurveySession"],
-    },
-    {
-      "tag": "v2",
       "new_sqlite_classes": ["SurveyDO"],
     },
   ],

deployment/modules/cloudflare/workers/survey/d1.tf

@@ -13,6 +13,10 @@ resource "null_resource" "d1_migrations" {
   }
 
   provisioner "local-exec" {
+    # local-exec defaults to /bin/sh, which is dash on the CI runners and has no
+    # pipefail.
+    interpreter = ["/bin/bash", "-c"]
+
     # Re-running an already-applied migration is benign ("already exists" /
     # "duplicate column"); every other error must exit non-zero so the failure is
     # visible in the workflow instead of being masked.

deployment/modules/cloudflare/workers/survey/worker.tf

@@ -109,6 +109,51 @@ resource "cloudflare_worker" "sessions" {
   }
 }
 
+# Cloudflare applies a version's migration when the version is deployed, and
+# rejects it unless old_tag matches the tag the Worker is already on. A fresh
+# Worker therefore needs the class-creating migration with no old_tag, and every
+# later deploy needs a no-op migration carrying that same tag — no single block
+# is valid in both states. This version exists only to run the first migration
+# and is frozen once created, so each new environment bootstraps itself and the
+# rolling version below stays on the no-op tag forever after.
+resource "cloudflare_worker_version" "sessions_bootstrap" {
+  account_id         = var.cloudflare_account_id
+  worker_id          = cloudflare_worker.sessions.id
+  compatibility_date = "2025-06-03"
+
+  main_module = "sessions.js"
+
+  modules = [{
+    name         = "sessions.js"
+    content_file = "${var.dist_dir}/sessions.js"
+    content_type = "application/javascript+module"
+  }]
+
+  migrations = {
+    new_tag            = "v1"
+    new_sqlite_classes = ["SurveyDO"]
+  }
+
+  lifecycle {
+    ignore_changes = all
+  }
+}
+
+resource "cloudflare_workers_deployment" "sessions_bootstrap" {
+  account_id  = var.cloudflare_account_id
+  script_name = cloudflare_worker.sessions.name
+  strategy    = "percentage"
+
+  versions = [{
+    version_id = cloudflare_worker_version.sessions_bootstrap.id
+    percentage = 100
+  }]
+
+  lifecycle {
+    ignore_changes = all
+  }
+}
+
 resource "cloudflare_worker_version" "sessions" {
   account_id         = var.cloudflare_account_id
   worker_id          = cloudflare_worker.sessions.id
@@ -123,9 +168,11 @@ resource "cloudflare_worker_version" "sessions" {
   }]
 
   migrations = {
-    old_tag = "v2"
-    new_tag = "v2"
+    old_tag = "v1"
+    new_tag = "v1"
   }
+
+  depends_on = [cloudflare_workers_deployment.sessions_bootstrap]
 }
 
 resource "cloudflare_workers_deployment" "sessions" {
@@ -137,6 +184,12 @@ resource "cloudflare_workers_deployment" "sessions" {
     version_id = cloudflare_worker_version.sessions.id
     percentage = 100
   }]
+
+  lifecycle {
+    # Redeploying the frozen bootstrap would otherwise silently roll the Worker
+    # back to the code it was first created with.
+    replace_triggered_by = [cloudflare_workers_deployment.sessions_bootstrap]
+  }
 }
 
 resource "cloudflare_worker" "api" {
@@ -146,6 +199,11 @@ resource "cloudflare_worker" "api" {
   observability = {
     enabled = true
   }
+
+  # Ordering only, so that destroy runs in reverse and tears this Worker down
+  # first: Cloudflare refuses to delete the sessions Worker while this one still
+  # binds its Durable Object namespace.
+  depends_on = [cloudflare_worker.sessions]
 }
 
 resource "cloudflare_worker_version" "api" {
@@ -184,16 +242,22 @@ data "cloudflare_zone" "immich_app" {
   }
 }
 
+# A Worker with no deployed version does not exist as far as the routes API is
+# concerned, and referencing only the name would let these race the deployment.
 resource "cloudflare_workers_route" "survey_api_root" {
   zone_id = data.cloudflare_zone.immich_app.zone_id
   pattern = "${module.domain.fqdn}/api"
   script  = cloudflare_worker.api.name
+
+  depends_on = [cloudflare_workers_deployment.api]
 }
 
 resource "cloudflare_workers_route" "survey_api_wildcard" {
   zone_id = data.cloudflare_zone.immich_app.zone_id
   pattern = "${module.domain.fqdn}/api/*"
   script  = cloudflare_worker.api.name
+
+  depends_on = [cloudflare_workers_deployment.api]
 }
 
 module "domain" {