[richiemcilroy]

Customer Loom imports now get their own lane on the Loom importer, so they no longer wait behind operator redrives and CSV backfills.

flowchart LR
  subgraph web
    D[Dashboard / agent API / mobile / Retry] -->|importPriority: customer| W[import-loom-video workflow]
    R[Internal redrive + CSV routes] -->|importPriority: bulk| W
    C[Recovery crons] -->|lane from the run receipt| W
  end
  W -->|POST /video/import| I
  subgraph I[Loom importer]
    Q1[customer queue] --> S[slots]
    Q2[bulk queue] --> S
  end

Loading

• Bulk keeps exactly LOOM_IMPORTER_MAX_JOBS slots. Customer imports get LOOM_IMPORTER_CUSTOMER_JOBS extra slots (default 16) on top, and are started first when a general slot frees (apps/media-server/src/loom-importer/app.ts).
• Bulk never takes a customer slot:

const canStart = (lane: ImportLane) =>
	running.size <
	config.maxActiveJobs +
		(lane === "customer"
			? config.maxCustomerJobs
			: Math.min(runningCustomer, config.maxCustomerJobs));

• Customer HLS segment fetches jump the shared connection pool (Semaphore.run(work, priority) in fetch.ts).
• Each lane has its own LOOM_IMPORTER_MAX_QUEUED limit, so a full bulk queue no longer returns SERVER_BUSY to customers.
• The web app labels operator entry points bulk (loom-import-redrive.ts, api/internal/loom-imports/csv). Everything else, including the agent API and Retry, is customer. The label is stored on metadata.loomImportRun.importPriority so recovery restarts keep it.
• Only the importer receives importPriority. The media fleet request body is unchanged.
• Either deploy order is safe: an older importer strips the field, and an importer that gets no label treats the import as bulk, which is today's single FIFO queue.

Benchmark

Same box (c7gd.4xlarge importer host), test importer with 2 general slots, 1 customer slot, 48 shared connections. Each run imports a real 53-minute Loom webinar (HLS, 779 MB) 6 times as bulk; "mixed" adds one customer import of the same video 5 s in. Runs alternated base, new, new, base.

Scenario	Metric	main	This PR
Bulk only	all 6 bulk done	13.4 s, 13.5 s	13.6 s, 13.2 s
Mixed	customer import, post to done	12.1 s, 12.0 s	3.7 s, 3.7 s
Mixed	all 6 bulk done	13.5 s, 13.4 s	14.1 s, 14.3 s
Mixed	everything done	17.1 s, 17.0 s	14.1 s, 14.3 s

• With no customer imports, bulk throughput is unchanged.
• In the mixed run, bulk finishes about 0.7 s later because the customer import shares the box's bandwidth instead of waiting for the bulk queue to empty.
• On main the customer wait grows with the size of the bulk queue. Today a customer agent import waited 3 to 7 minutes behind a 1,216-video redrive.

Confidence Score: 5/5

The PR appears safe to merge; no blocking issues were found.

Summary

Gives customer Loom imports a separate queue and extra running slots.

• Customer Loom imports start ahead of bulk work.
• Loom imports keep their customer or bulk lane when they start and restart.

No actionable issues were found. No previous review threads were supplied.

_{Reviews (2) · Last reviewed commit: "feat(web): label operator Loom imports a..." · Reviewed by Greptile}

[richiemcilroy]

hey @greptileai, please re-review the PR

[capy-ai]

Capy found 1 potential issue (1 medium).

View 1 other finding in Capy.

[capy-ai]

Concierge migrations enter the customer lane

Medium Severity

The unchanged importConciergeCsvRows caller in apps/web/actions/loom.ts invokes processLoomCsvRows without importPriority. The admin migration UI dispatches successive CSV batches through this operator-only path, so the undefined value forwarded here becomes customer in processVideoOnMediaServer. A sufficiently large or slow concierge migration can therefore occupy the reserved customer slots and queue ahead of interactive imports, defeating the new isolation. Pass importPriority: "bulk" from the concierge caller, or derive that lane from its migrationRequestId, while retaining the customer default for user imports.