Server stalls from sharing: what is left since 5 November

From ~/.obeya/stalls.log, 30 October to 9 November 2026. “Sharing builds the shared cards once per pass” landed on main on 5 November, 08:36.

6stalls with share.ts in the stack since 5 November (4 days); 14 in the 3 days before
1.2 sthe longest of them; 4.8 s before
5 of 6over 0.5 s, but the time went into database reads that take 3 ms normally

Every server stall, by day

share.ts in the stackother code

Before the fix, the long share.ts stalls were the server building each shared card on its own (archiveOf ← card ← find ← checkSites). Since then, sharing has no stall of its own: its stalls fall into the same few windows as stalls in unrelated code.

The six since the fix

When (local)StallWhere most of the time wentJust before (app.log)
Sat 7 Nov, 18:400.6 sall ← items ← countAgents (25 %); share.ts only 8 %Restart with new code, Parakeet voice loading
Sun 8 Nov, 20:120.7 sall ← shares ← sharedCards ← checkSites (81 %)Qwen3 voice loaded (16.3 s), after a restart
Sun 8 Nov, 20:180.9 sall ← shares (37 %), all ← archiveOf (34 %)Qwen3 voice reloaded twice for language switches
Sun 8 Nov, 20:321.2 sall ← shares ← sharedCards ← checkSites (66 %)Restart, then Qwen3 voice loaded (21.3 s)
Sun 8 Nov, 20:341.1 scards, JSON parse, URL in others ← publishTwo minutes after that load
Sun 8 Nov, 20:430.5 sall ← shares (39 %), all ← archiveOf (32 %)Landings in a row

Why it is not the sharing code

Database readOn a copy of the database, warmInside a stall
All shared cards of the Obeya canvas (446 rows, shares)3 ms (38 ms the first time)0.3–0.8 s
Workstreams of the archived projects (archiveOf)1 ms (18 ms the first time)0.1–0.3 s
Conversation waiting on a card (waitingTalk, unrelated to sharing)0.1 ms0.5–0.9 s

Reads 100 to several thousand times slower than normal, in sharing and elsewhere alike, mean the server process was held up, not that the code got slower. Every server stall since 5 November falls in a window right after an Obeya restart or a load of the narration voice: Qwen3-TTS, 4.2 GB, which is loaded again on each switch between German and English (six times in 30 minutes on 8 November). The machine has 48 GB of memory and had 11.6 GB in swap: the idle server's memory gets paged out, and its next reads wait for it to come back.

Proposed cards

Qwen3 narration voice: switching language reloads the 4 GB model. Keeps one voice process for both languages, which removes most of the loads these stalls follow.

stalls.log names page-ins, so a stall from memory pressure reads as one. Each entry also says how many pages the server had to read back during the stall, so the next review sees this at a glance.