ADR-0006: Delete dead EC2 background processors

Status: accepted
Date: 2026-04-17

Context

PropFlow's Next.js server (EC2 / PM2) starts a handful of setInterval background processors at boot time via src/instrumentation.ts. Over the course of 2025-2026 several of them have become redundant:

Processor	File	Why it exists	Why it's dead
Email Replay	`src/lib/email/replay-stalled.ts`	Resurrect email ingestion records stuck in `processing`/`received` after a server crash.	SQS FIFO redrive + `propflow-inbound.fifo` DLQ now handle retries transparently. The processor only runs when `SQS_INBOUND_QUEUE_URL` is unset — which prod hasn't done since Phase 2.
Dead Letter Retry	`src/lib/email/retry-processor.ts`	Periodically retry outbound emails that SendGrid/SES rejected with a transient error.	SendGrid's native retry + SES bounce handler cover every real-world case. The DL table has not had a successful resend in weeks.
Sync Digest	`src/lib/rent-roll/sync-digest-processor.ts`	Fire a daily 8 AM CT digest email summarizing rent-roll sync activity.	Replaced by an ad-hoc admin trigger `POST /api/admin/sync-digest` + the daily cron described in `scripts/`. Nobody reads the auto-sent digest.
Voice Trace	`src/lib/voice/voice-trace-processor.ts`	Backfill `AgentTrace` rows from ElevenLabs 20 s after each call, because ElevenLabs finalises analysis after the webhook fires.	The voice call-ended webhook now builds the trace inline from the same payload (analysis arrives in the webhook body; the 20 s race is gone).

In addition, two processors ARE still producing value, but the timer-on-EC2 shape is the wrong primitive for them:

Processor	File	Better shape
Application Link	`src/lib/leasing/application-link-processor.ts`	EventBridge Scheduler one-shot: tour confirmation schedules an invocation for `now + postTourFollowUpDelayMinutes`.
Listings Sync	`src/lib/leasing/listings-sync-processor.ts`	EventBridge rate(15 m) rule targeting a tiny Lambda.

Decision

Delete the four dead processors (file + tests + comments). Their DynamoDB artefacts (e.g. dead-letter table entries, voice trace status flags) are left in place — no prod data is touched by this ADR.
Migrate application-link to a one-shot EventBridge Scheduler + Lambda (propflow-application-link). Call site: applyTourIntent (action=confirm and auto-confirmed propose).
Migrate listings-sync to an EventBridge rate(15m) rule + Lambda (propflow-listings-sync).
Until the new Lambdas have been live for one week, the existing setInterval versions of application-link and listings-sync stay running as a belt-and-braces safety net. After the cutover week, delete the *-processor.ts files and the startX() calls in src/instrumentation.ts.
All prod AWS resource creation (EventBridge rules, Scheduler schedules, IAM policies) is performed out-of-band via the deploy scripts checked in under lambda/<name>/deploy.sh. This ADR does not create any cloud resources on its own.

Consequences

src/instrumentation.ts shrinks from seven setInterval processors to two (both listed as "safety net — delete after cutover") plus the SQS inbound consumer.
EmailIngestionRecord.processingStartedAt, replayCount, and lastReplayAt are retained in the schema for back-compat with records written before this change; new records only populate processingStartedAt (set by src/lib/email/process-inbound-email.ts).
Voice traces are built synchronously at call-ended time and visible on the admin trace page with no ~30 s lag.
Tour-confirmation code paths now have an outbound side effect (an EventBridge Scheduler CreateSchedule call). Failure to schedule is logged + swallowed — the setInterval safety-net still catches it.

Implementation notes

Phase 5 worktree: fede/phase5-processors.
Follow-up: when the cutover completes, a refactor: remove legacy application-link + listings-sync setIntervals PR will delete src/lib/leasing/application-link-processor.ts + src/lib/leasing/listings-sync-processor.ts and the corresponding imports from src/instrumentation.ts. See docs/data-model-migration/architecture/phase5-cleanup-notes.md.