trails/apps/journal/app/jobs/deliver-activity.ts
Ullrich Schäfer 855244747c
journal: typed job seam + Komoot credentials through the manager
Two related fixes from the architecture review:

Typed job seam. Job payloads were `unknown` end-to-end: every handler
opened with `job.data as SomePayload`, every enqueue site passed a bare
string queue name and an unchecked object, and a typo meant a runtime
failure in a background worker. packages/jobs gains defineJob(), which
keeps the payload typed inside the handler and performs the
contravariance cast once inside the package. The journal declares all
14 queues and their payload shapes in app/jobs/payloads.ts; enqueue()/
enqueueOptional() and defineJournalJob() key off that map, so enqueue
sites and handlers cannot drift and queue names are compile-checked.

Komoot credential bypass. The bulk-import route enqueued the raw
credentials JSONB in the pg-boss payload, skipping withFreshCredentials
entirely: credentials sat at rest in the job table, were never
refreshed if stale, and markNeedsRelink never fired. The payload now
carries only the serviceId; the handler resolves fresh credentials
through the ConnectedServiceManager at execution time, and marks the
import batch failed (instead of leaving it pending forever) when
credential resolution itself fails.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-06-10 02:05:24 +02:00

130 lines
4.5 KiB
TypeScript

import { defineJournalJob } from "./payloads.ts";
import { and, eq } from "drizzle-orm";
import { activities, users } from "@trails-cool/db/schema/journal";
import { getDb } from "../lib/db.ts";
import { getOrigin } from "../lib/config.server.ts";
import { getFederation } from "../lib/federation.server.ts";
import {
activityToCreate,
activityToDelete,
type FederatableActivity,
} from "../lib/federation-objects.server.ts";
import {
getCachedRemoteActor,
upsertRemoteActor,
type DeliveryPayload,
} from "../lib/federation-delivery.server.ts";
import { logger } from "../lib/logger.server.ts";
/**
* Outbound pacing (spec 5.5): never exceed 1 request/second per remote
* host. In-process map is sufficient — pg-boss works this queue
* sequentially in the single journal process.
*/
const lastSendPerHost = new Map<string, number>();
const MIN_INTERVAL_MS = 1000;
async function paceHost(host: string): Promise<void> {
const last = lastSendPerHost.get(host) ?? 0;
const wait = last + MIN_INTERVAL_MS - Date.now();
if (wait > 0) await new Promise((r) => setTimeout(r, wait));
lastSendPerHost.set(host, Date.now());
}
/**
* Deliver one activity to one remote follower's inbox (spec 5.3/5.4).
* One job per (activity, recipient) so each delivery retries with
* exponential backoff independently (configured at enqueue time —
* see enqueueActivityDeliveries). A thrown error marks the attempt
* failed and pg-boss retries; exhausting the budget is the permanent
* failure, logged by the final catch.
*/
export const deliverActivityJob = defineJournalJob({
name: "deliver-activity",
expireInSeconds: 60,
async handler(jobs) {
for (const job of jobs) {
const p = job.data;
try {
await deliverOne(p);
} catch (err) {
logger.warn(
{ err, action: p.action, objectIri: p.objectIri, recipient: p.recipientActorIri },
"deliver-activity attempt failed (pg-boss will retry until budget exhausted)",
);
throw err;
}
}
},
});
async function deliverOne(p: DeliveryPayload): Promise<void> {
const federation = getFederation();
const ctx = federation.createContext(new URL(getOrigin()), undefined);
// Build the activity to send. For `create`, re-read the row at
// delivery time: if it was deleted or un-publicized since enqueue,
// skip rather than leak.
let activity;
if (p.action === "create") {
const db = getDb();
const [row] = await db
.select()
.from(activities)
.where(and(eq(activities.id, p.activityId!), eq(activities.visibility, "public")))
.limit(1);
if (!row) {
logger.info({ objectIri: p.objectIri }, "deliver-activity: activity gone or non-public; skipping");
return;
}
// Spec 9.3: flipping the profile to private stops federation — also
// for deliveries already enqueued when the flip happened.
const [owner] = await db
.select({ profileVisibility: users.profileVisibility })
.from(users)
.where(eq(users.username, p.ownerUsername))
.limit(1);
if (!owner || owner.profileVisibility !== "public") {
logger.info({ objectIri: p.objectIri }, "deliver-activity: owner no longer public; skipping");
return;
}
activity = activityToCreate(row as FederatableActivity, p.ownerUsername);
} else {
activity = activityToDelete(p.objectIri, p.ownerUsername);
}
// Resolve the recipient's inbox: cached remote_actors row first,
// actor-document fetch as fallback (which also primes the cache).
const recipientIri = new URL(p.recipientActorIri);
let inboxUrl: URL;
const cached = await getCachedRemoteActor(p.recipientActorIri);
if (cached?.inboxUrl) {
inboxUrl = new URL(cached.inboxUrl);
} else {
await paceHost(recipientIri.host);
const actor = await ctx.lookupObject(recipientIri);
const fetchedInbox =
actor != null && "inboxId" in actor ? (actor.inboxId as URL | null) : null;
if (!fetchedInbox) {
throw new Error(`deliver-activity: cannot resolve inbox for ${p.recipientActorIri}`);
}
inboxUrl = fetchedInbox;
await upsertRemoteActor({
actorIri: p.recipientActorIri,
inboxUrl: inboxUrl.href,
domain: recipientIri.host,
});
}
await paceHost(inboxUrl.host);
await ctx.sendActivity(
// Identifier and username are the same thing in our actor model.
{ identifier: p.ownerUsername },
{ id: recipientIri, inboxId: inboxUrl },
activity,
);
logger.info(
{ action: p.action, objectIri: p.objectIri, recipient: p.recipientActorIri },
"deliver-activity: delivered",
);
}