The read-only 14-day Observe trial
Connect a read-only token. Hetscale reads your history, writes a report, then records for 14 days what it would have done live.
What it does
- Reads 30 days of metric history for the servers behind your load balancer.
- Produces a day-0 report, minutes after connecting, with ranged results: capacity needed, peaks, an estimated cost range, periods a server was unhealthy, days with CPU pinned at 100%, and a recommended policy labelled "default".
- For 14 days, records every decision it would have taken live, with the reason, prefixed "would", using the same rules as real healing and scaling.
- Shows the signals it used (CPU, requests per second, load balancer target health) and the ones it cannot see (RAM, latency). The report can only use what Hetzner exposes.
What it does not do
It does not create, change or delete anything. It cannot: the token is read-only, and Hetzner enforces that.
- No server is created or deleted.
- No load balancer target is added or removed.
- Nothing on your servers is installed or changed.
Sample report
group web4 × cx33fsn130 days of history + 14 observed daysHetzner list price, 2026-09-17illustrative sample
Capacity needed2–3 nodesmost hours; 4 at peakconfidence: medium
Times over capacity1.6–1.9×at peak; weekdays, about 6 h/day
Compute today€34/month4 × cx33, fixed
Estimated with autoscaling€22–23/montha range, not a promise
Nodes needed per hour, last 30 days
fleet today, fixed · 4
day 1day 15day 30
The space above the bars is what you pay for and do not use.A typical weekday
fleet today, fixed · 4
00:0012:0023:00
Peak shape decides how many burst node-hours you would buy.Signals
Used
- CPU per node
- Requests per second at the load balancer
- Load balancer target health
Not visible
- RAM
needs the optional agent, later - Latency
needs the optional agent, later - Queue depth
needs the optional agent, later
Unhealthy periods
web-1
web-2
web-3
web-4
CPU
day 1day 30
web-2 was unhealthy twice; both times it would have been replaced automatically. CPU at 100%: 2 days.
Recommended policydefault · Balanced preset
- Scale out
- CPU above 75% for 3 minutes
- Scale in
- CPU below 35% for 10 minutes
- Cooldown
- after every action
- Warm-up
- after a node becomes active
- Minimum
- 2 nodes
- Maximum
- 4 nodes
What it would have done
- wouldcreate web-5CPU 81% for 3 min · range 2–4, now 3 → 4
- woulddrain web-5CPU 31% for 10 min · 4 → 3
- wouldreplace web-2unhealthy 9 min · health failed
- wouldcreate web-5CPU 79% for 3 min · 3 → 4
- woulddrain web-4CPU 33% for 10 min · 3 → 2
- wouldreplace web-2unhealthy 41 min · health failed
22 scale-outs, 22 scale-ins, 2 replacements in 14 days.
Every number in a real report is a range or a count from your own data. Nothing in it is a promise. Sharing a report is opt-in and strips names and IPs.
Connect read-onlyTrial mechanics
- 14 days, one read-only token, no card.
- Polling stops when the trial ends. "Refresh my report" is allowed once every 30 days after that, without continuous polling.
- The read token is deleted 7–30 days after the trial ends; the report is kept.
- Trial identity is keyed by your Hetzner server and load balancer ids, not only by email.
- Per-group Observe mode is also a permanent feature of every plan: groups drop to Observe on a failed payment or a downgrade, and nodes are never destroyed.
After the trial
You choose. Stay in Observe, run one Rehearsal to see a node created and removed while you watch, or move a group to Propose or Live. If you do nothing, nothing happens.