Cold starts for a cold wallet
Deploying Stanza collides with the balance sheet: a 24/7 cloud instance would be lovely and unaffordable, so the service lives on Google Cloud's free tier with cold starts, and the first bug is exactly that, requests dying because the models take five to ten minutes to download while the app is already asking. The fixes are working-poor engineering: a retry queue so no user vocabulary is ever lost (try, fail, requeue, every five minutes until it lands), and a Docker prebake that cuts the wake-up to about sixty seconds for roughly fifty cents a month of storage. The context that explains the constraint arrives at the end: the investor's Wise transfer, promised in two days of verification, has been stuck for a week, and 'estamos sin plata'.
// trace: where this idea came from
- ↳ video diary @ 6:26 (los free tiers y el cold start)
- ↳ video diary @ 8:19 (la cola de reintentos, que no se pierda nada)
- ↳ video diary @ 26:10 (estamos sin plata, la inversión trabada en Wise)
- ↳ Entry 229-1: The lemma is the unit (el servicio de lemas que aquí se despliega)
The lemma service meets the balance sheet. The lovely version, Stanza available 24/7, means renting cloud that never sleeps, and that cobra bastante ▸ 6:18; the affordable version is Google Cloud’s free tier with cold starts, ask for a slice of machine only when needed ▸ 6:26. The models make it worse: each language is its own neural net weighing hundreds of megabytes, like shipping a dictionary per tongue ▸ 9:20.
The first production bug is the architecture describing itself: requests failing because the service takes five to ten minutes to wake and download its packages, and the app was already asking ▸ 7:59. The fix protects the only thing that can’t be regenerated: the user’s vocabulary must never be lost ▸ 8:13, so a queue holds every update and retries, fail, requeue, again every five minutes, until the analysis finally lands ▸ 8:19. Next step, prebaking the models into Docker so the wake-up falls from ten minutes to about sixty seconds ▸ 12:18, at a storage cost near fifty cents a month, cheap enough to be a yes on sight ▸ 12:49.
gratis, con cola de reintentos →
The constraint has a face, and it arrives as the day’s last update: the investor’s transfer is riding Wise, which promised two days of account verification and has now held it a week, sending messages that insist there’s nothing to do ▸ 25:48. The state of the treasury is one clause, “realmente pues estamos sin plata,” and the task list grows a new item: start calling Wise ▸ 26:10. Until the phone call works, the architecture stays what the wallet can carry: free tiers, cold starts, and a queue that refuses to forget…