Product Update
Cold-start cut from 141s to 66s on Starter
azmat••2 min read
When a new Seaboss tenant signed up, the wait from "click checkout" to "agent answering Telegram" used to be 141 seconds on our Starter tier. As of RASF-276 Phase 3, that's 66 seconds. Pro tier dropped from 141s to 88s.
Where the time was going
Most of those 141 seconds were not the Hetzner VPS booting — that's the ~40-50s floor we can't move from the API side. The rest was Seaboss's own provisioner doing work that, on reflection, was happening at the wrong time:
- Pulling the OpenClaw agent container.
- Pulling the
llm-proxysidecar. - Pulling the
n8nworkflow engine. - Starting them, waiting for n8n's owner-setup, waiting for health probes.
Every tenant got the same containers. Every tenant waited for the same pulls. That's a textbook case for moving work into the snapshot.
Phase 3 baking
The fix is structural: when we cut a new Gold release, we provision a one-off VPS, install the containers, perform any expected first-boot configuration (n8n owner setup, llm-proxy env injection), then take a Hetzner snapshot of the running disk. Future tenants spawn from that snapshot, not from a clean Ubuntu image.
Per-tenant boot becomes: rsync the tenant's specific environment file into
/opt/seaboss/.env, then docker start llm-proxy. The containers are
already there. n8n is already initialized.
What it looks like now
| Tier | Hetzner spawn | Phase 3 internal | Total wall-clock | Improvement |
|---|---|---|---|---|
| Starter (cpx21) | 50s | 16s | 66s | 53% faster |
| Pro (cpx31) | 40s | 48s | 88s | 38% faster |
Pro is slower than Starter on the internal-work side because the n8n owner
setup and the optional browser-use-local install both run there but skip
on Starter.
What it cost
Phase 3 baking added a one-time ~5 minutes of build time when we cut a new Gold release, plus snapshot storage on Hetzner (~15GB per tier × 3 tiers). At beta scale that's a rounding error against the wall-clock minutes we're saving per tenant.
What's next
The next big cut would be pre-acquiring a VPS during the customer's onboarding interview, so the ~40-50s Hetzner boot overlaps with conversation time rather than blocking it. Filed as RASF-749 and sitting in the backlog.