Client-only server update leaves the app empty for 96 s on a large database [📡#734]

🤖 claude/anthropic claude-opus-5-5 (medium)

With a 22 GB database, a client-only server update leaves the app empty for about 96 s.
From the server Update click to threads showing takes about 110 s.

Origin: user request, 2026-10-11, to profile the backend/frontend split update flow in the QA VM.

update-run-labelled.mp4

Setup

Item	Value
VM	`qa-desktop`, 4 vCPU, 8 GiB
Backend	`t3code.service` (userdata server)
Frontend	desktop AppImage, client-only
From	`v0.0.46-hyprws-nightly.20261011.1006` (`d3367ee34cff`)
To	`v0.0.46-hyprws-nightly.20261011.1007` (`68d9fe1e315f`)
Database	22.4 GB `statev2.sqlite`, copy of a real install
Driving	UI only

Timings

Step	Real	Clip
1. Download desktop update	5.5 s	2x
2. Restart & install click to confirm dialog	15.5 s (harness, see below)	5x
3. Confirm to frontend back	3.9 s	1x
4. Server Update: download and preflight	10 s	4x
5. Server down, app empty	91.5 s	10x
6. Server up to threads showing	6 s	1x
Idle spans between steps		cut

A run with a 562 MB database had about 2.5 s of downtime, so step 5 grows with database size.

Cause

backupDatabaseOnce in apps/server/src/serviceLauncher.ts copies and fsyncs statev2.sqlite with its -wal and -shm files.
It runs in #startTrial after the old server stops, so the whole copy is downtime.

Evidence	Value
Guest `cp` + `sync` of the DB	94.9 s
Measured downtime	91.5 s
Migrations `.1006` to `.1007`	none
Backup after commit	discarded

The backup bought nothing for this update.

Proposed fixes

Fix	Change	Gain
a	Skip the backup when preflight reports no pending migrations	~90 s less downtime here
b	Keep last known projects, threads, and route during reconnect	No empty "Connect to a computer" state; no lost draft
UI	One server-update state through the restart: stages, elapsed time, done state	User sees progress instead of an empty app
c (option)	Online SQLite snapshot before the old server stops	Backup off the downtime path; writes after the snapshot are lost on rollback

Today the route is lost on reconnect: an open draft became a new thread.

Other findings

Dialog delay: test harness, not the app
Renderer click to dialog: 222 ms, no stalls
Harness: RSI-Software/donjor-skills#1100
Closed: Find why the desktop install confirm dialog takes 15 s [🐛#1705] #1709
Frontend restart: brief empty-project flash

Verification

Nightly 0.0.46-hyprws-nightly.20261011.1014 (75aad8f), QA VM, 10.9 GB copy of a real DB.
Children: #1706, #1707, #1708.

Installed launcher	Backup	Downtime	Runs
`.1009`	taken	41 to 76 s	4
`.1013`	skipped	1.5 to 2.5 s	4

Draft, route, and threads survived all 11 trials.

happy-update.mp4

Open gap

Moved to #1865: the service unit keeps its install-time launcher.
The Backup skip ships (#1711) and was verified above with launcher .1013.

Stop during the backup (2026-10-24)

A stop during a migration backup hangs until systemd kills the launcher at 90 s.
Approved plan: cancel the copy on stop and keep the update pending; name the backup in the client's restart label.
Cheaper-backup options stay out of scope; re-measure after upstream #22630.