improvement(trigger): downsize connector sync and document processing machines (#8699)

Author: waleedlatif1

apps/sim/background/knowledge-connector-sync.ts

@@ -98,13 +98,17 @@ export const knowledgeConnectorSync = task({
   id: 'knowledge-connector-sync',
   maxDuration: CONNECTOR_SYNC_MAX_DURATION_SECONDS,
   /**
-   * Sized from production telemetry: peak sampled RSS 2.6 GB and peak 1.4 vCPU,
-   * so `large-1x` holds ~3x memory and ~2.8x CPU headroom. No `outOfMemory`
-   * escalation: an OOM is a SIGKILL, so the run never reaches the terminal
-   * write that clears `syncLockToken`, and the escalated attempt would find the
-   * row still `syncing` and skip. The stale-lock reaper owns that recovery.
+   * Sized from production telemetry: p99.9 sampled RSS ~650 MB, peak 2.6 GB, and
+   * peak 1.4 vCPU, so `medium-2x` (4 GB, 2 vCPU) still clears the worst case.
+   * The run is I/O-bound — it mostly waits on source APIs and their rate-limit
+   * backoff — and Trigger.dev bills wall-clock per preset, so `large-1x` paid
+   * double for headroom no run used. Not `medium-1x`: 2 GB sits below the peak,
+   * and there is no `outOfMemory` escalation to absorb it. An OOM is a SIGKILL,
+   * so the run never reaches the terminal write that clears `syncLockToken`,
+   * and the escalated attempt would find the row still `syncing` and skip. The
+   * stale-lock reaper owns that recovery.
    */
-  machine: 'large-1x',
+  machine: 'medium-2x',
   retry: {
     maxAttempts: 3,
     factor: 2,

apps/sim/background/knowledge-processing.ts

@@ -319,12 +319,13 @@ export const processDocument = task({
   id: 'knowledge-process-document',
   maxDuration: envNumber(env.KB_CONFIG_MAX_DURATION, 600),
   /**
-   * Sized from production telemetry: peak sampled RSS 902 MB and peak 1.2 vCPU
-   * across a corpus where no document exceeded 2 GB, so `medium-2x` holds ~4x
-   * memory and ~1.7x CPU headroom over the observed worst case. The prior
-   * `large-1x` reserved 8 GB against a worst case using an eighth of it.
+   * Sized from production telemetry: p99.9 sampled RSS ~740 MB and peak just
+   * under 2 GB, so `medium-1x` (2 GB) holds the overwhelming majority and the
+   * rare document past it escalates through `outOfMemory` below. Processing waits on embedding and storage calls far
+   * more than it computes, so the single vCPU stretches a CPU-bound parse
+   * slightly but halves the wall-clock rate every run is billed at.
    */
-  machine: 'medium-2x',
+  machine: 'medium-1x',
   retry: {
     /**
      * The ceiling for thrown errors: database retries use all of it, and