A Control Plane Leasing Work to a Worker

"The 512 GB Studio owns the truth; the 192 GB Studio does the work; a lease with an expiry is the only thing between them. Four thousand nine hundred jobs later there is nothing half-written."

The Pattern

The video-memory engine keeps the household's film archive searchable: it inventories a media library, transcribes every video, and summarizes the transcripts. Its architecture splits along the fleet's capacities. Office is the control plane: it holds the application, the queue and the one process allowed to write the database. Worker — the 192 GB M2 Ultra — scans the library, runs the transcoder and submits audio to the transcription provider. The handshake between them is a lease: the worker asks for a job, the control plane marks it with the worker's name, a token and an expiry, the worker does it and returns the result, and the result counts only when the control plane accepts it. If the worker dies mid-job, the lease expires and the job is offered again; if the worker returns something the control plane rejects, nothing is recorded. Read live and read-only on 2026-09-15: 4,927 transcription jobs, 4,921 completed, 6 failed, zero leases open, 2,546 videos, 3,011,560 transcript segments. The same pattern runs the model store, whose engine sits on office and whose fetch worker sits with the archive, and the fleet-control tool, which drives every Mac over ssh from office.

Why the Writer Is Singular

The single-writer rule is the fleet's most important invariant and the least visible one. Every household engine's database has exactly one process that writes it, on exactly one Mac, and every command-line tool on every other Mac is an API client that asks that process. It is why office's service count is six times any other machine's, why a worker can be rebooted without a thought, and why the quest workshop that produced this quest could be driven from a laptop while its engine logged every stage on office. The rule costs a network round trip per write and buys the absence of an entire class of bug — two processes disagreeing about a row — that a fleet of nine machines would otherwise meet weekly.

What the Summary Column Says

The queue's summary table carries the cloud-by-choice lesson in advance. Every one of its 2,463 summaries was requested through the hub's own daemon — the server alias column has one value — and 2,462 of them ran on cloud models through that daemon's cloud tier; one ran on a local 31B. The control plane did not choose local because the machine could not: the hub holds a 31B in bf16 and the pool could run a far larger one. It chose the cloud tier for a summarizer because the frontier models summarize better and the daemon makes the tier a setting. That is the whole fleet in one column: a local control plane, a local worker, transcripts and summaries fetched from the cloud by that worker, and a model tier chosen by quality rather than by where the silicon is.

Code

lease.py — the queue counted read-only: jobs, leases, and where the summaries ran·python

#!/usr/bin/env python3
"""A control plane leasing work to a worker: the queue counted, read-only. The control
plane on office keeps the only writer; a worker takes a lease with a token and an expiry
and the result counts only when the control plane accepts it. Path from the environment."""
import os, sqlite3
con = sqlite3.connect(f"file:{os.environ['RECALL_DB']}?mode=ro", uri=True)
con.execute("PRAGMA query_only=ON")
print("jobs by kind and status:", con.execute("select kind, status, count(*) from jobs group by kind, status order by kind, status").fetchall())
print("open leases:             ", con.execute("select lease_owner, count(*) from jobs where lease_owner is not null group by lease_owner").fetchall())
print("summaries by model:      ", con.execute("select model, count(*) from summary_jobs group by model order by 2 desc").fetchall())
print("summaries by daemon host:", con.execute("select server_alias, count(*) from summary_jobs group by server_alias").fetchall())
print("corpus:                  ", con.execute("select count(*) from videos").fetchone()[0], "videos,",
      con.execute("select count(*) from transcript_segments").fetchone()[0], "transcript segments")
# office, 2026-09-15:
# jobs by kind and status: [('transcribe', 'completed', 4921), ('transcribe', 'failed', 6)]
# open leases:              []
# summaries by model:       [('glm-5.2:cloud', 2436), ('glm-5.3:cloud', 12), ('kimi-k3:cloud', 10), ('deepseek-v4.1-flash:cloud', 4), ('gemma4:31b-it-qat', 1)]
# summaries by daemon host: [('server', 2463)]
# corpus:                   2546 videos, 3011560 transcript segments

Exercise

Find one queue in your own setup — a job table, a task list, a cron that writes files — and answer three questions on your card: which single process writes it, what a lease or claim looks like, and what happens to a job whose worker vanishes. If any answer is 'nothing', that is the row to fix.

Hint

A file that two scripts append to is two writers. A job marked 'running' with no expiry is a lease that never returns. The control plane pattern is small: an owner column, a token, a timestamp, and a rule that the owner of the database is the only one who changes it.