fix(firestore-send-email): match the extension's cpu, concurrency and ingress

kits/firestore-send-email/CHANGELOG.md

@@ -1,3 +1,5 @@
+- fix: deploy `processQueue` with the extension's runtime settings: `concurrency: 1`, `cpu: "gcf_gen1"` (0.1666 vCPU at 256MiB) and `ingressSettings: "ALLOW_INTERNAL_ONLY"`. `setGlobalOptions` can't change these. Before, the kit took the 2nd gen defaults of concurrency `80`, 1 vCPU and `ALLOW_ALL`. A burst of mail with attachments landed on one 256MiB instance, which ran out of memory and dropped every event in flight on it: the documents were left in `PENDING` or `PROCESSING`, or with no `delivery` field, and were never retried or marked as failed, so the mail was never sent. Existing kit deployments switch on their next deploy.
+
 - fix: an Eventarc publish failure no longer fails the function. A channel that was deleted or is unreachable answers `PERMISSION_DENIED`, and the publish was awaited before the work the function exists to do, so the mail was never sent. The publish failure is now logged as a warning and the invocation continues. Deploys also declare `eventarcpublishing.googleapis.com`, so the API is enabled on your project whether or not you use events.
 
 - docs: the no-region fallback needs an explicit empty `DATABASE_REGION=` line in `.env`. Omitting the key entirely fails a non-interactive deploy, because the CLI resolves parameters before it reads their defaults. Corrects the note on the `DATABASE_REGION` mapping fix below; no behavior change.

kits/firestore-send-email/README.md

@@ -280,6 +280,11 @@ a `TypeError` about reading `attachments` into `delivery.error`. It now writes
 - SendGrid is still selected by an `smtp.sendgrid.net` connection URI with the
   API key in `SMTP_PASSWORD`, and Outlook hosts still get their explicit
   transport configuration.
+- The function still runs like the extension's: 0.1666 vCPU
+  (`cpu: "gcf_gen1"`), one event per instance (`concurrency: 1`) and
+  internal-only ingress, rather than the 2nd gen defaults of 1 vCPU,
+  concurrency `80` and `ALLOW_ALL`, so a burst of mail scales out across
+  instances instead of sharing one. `setGlobalOptions` can't change these.
 - The function still runs with a 120 second timeout, and `TLS_OPTIONS`,
   `DATABASE`, `USERS_COLLECTION` and `TEMPLATES_COLLECTION` behave as before.
 

kits/firestore-send-email/src/index.ts

@@ -108,6 +108,15 @@ function ensureInitialized(): Promise<HandlerContext> {
 export const processQueue = onDocumentWritten(
   {
     ...(deploy.region ? { region: deploy.region } : {}),
+    // The extension ran this function at 0.1666 vCPU, one event per instance,
+    // with internal-only ingress. Left unset, the CLI deploys the 2nd gen
+    // defaults: 1 vCPU, concurrency 80 and `ALLOW_ALL`. At concurrency 80, a
+    // burst of mail with attachments lands on one 256MiB instance, which runs
+    // out of memory and drops every event in flight on it. The trigger has no
+    // retry policy, so that mail is never sent.
+    concurrency: 1,
+    cpu: "gcf_gen1",
+    ingressSettings: "ALLOW_INTERNAL_ONLY",
     document: deploy.document,
     database: deploy.database,
     timeoutSeconds: 120,

kits/firestore-send-email/tests/index.test.ts

@@ -89,4 +89,13 @@ describe("processQueue options", () => {
     const document = options.document as { toCEL(): string };
     expect(document.toCEL()).toContain("params.MAIL_COLLECTION");
   });
+
+  // Concurrency 80 let a burst of mail with attachments land on one 256MiB
+  // instance, which ran out of memory and dropped it.
+  test("the function runs one event per gen1 CPU instance with internal-only ingress", async () => {
+    const options = await loadTriggerOptions();
+    expect(options.concurrency).toBe(1);
+    expect(options.cpu).toBe("gcf_gen1");
+    expect(options.ingressSettings).toBe("ALLOW_INTERNAL_ONLY");
+  });
 });