Text recognition: three jobs at once at night

The transcription service now runs up to three jobs at the same time in the night window (19:00–06:00, Friday evening to Monday morning without a break). By day it stays at one job. The values come from two measurement runs on the production host.

What changes on the server

Setting	Before	Now	Why
Jobs at once at night	1	3	Three jobs finish 2.75× what one does; the app's typical response time did not move.
CPU cap	4 cores	12 cores	Four threads per job. Lowest priority stays: every app request goes first.
Memory cap	10 GB	24 GB	Three jobs of 15 min peaked at 13 GB. A 3-hour recording fits beside two others.
Hold back new night jobs at CPU pressure	off	above 15 %	During an AI run beside three jobs the app slowed noticeably (pressure up to 28 %).
Stop a job at memory pressure	off	above 10 % for 1 min	Normal nights stay below 3 %; the stopped job is queued again automatically.
Time estimate per second of audio	1.5 s	1.7 s	Each job is about 10 % slower when three run together.

Unchanged: 4 GB of memory always stay free for the host; at most one job per museum by day. The same values are set for the self-hosted (ADK) installation.

Hours of recording finished per hour at night

1 job (before)

1.6 h

3 jobs (now)

4.3 h

3 jobs during an AI run

2.2 h

Measured on 11 and 12 November 2026 with a 15-minute conversation at quality "best".

What the owner does after it lands

Update the stack on the host as usual (sudo server/update_stack.sh). It recreates the service; a job running at that moment is queued again by its museum, nothing is lost.
Run sudo sh server/transcribe/check-night.sh (works from fish). It shows the limits the kernel really applies, which can differ from the configuration, plus what the service logged. Expected: CPU weight 1, cap 1200000 100000, memory 25769803776, oom score 1000.
On Monday, run it again: it then counts the jobs stopped for lack of memory and the nights where CPU pressure held jobs back over the weekend.

A follow-up card on Monday morning takes both outputs, notes the kernel values in the service documentation and the host maintenance notes, and reviews the weekend.