The SLA Builder.
Until you've defined good, every complaint is a crisis. Set your service classes and the four indicators, and take away the agreement itself: one shared benchmark, tiers that differ by price and nines, and a document your tenants can hold you to.
Service classes
The guide's model: same threshold for every class ('2.5% will serve all classes - you just need to adjust the nines'); classes differ by price, overcommit and the promised percentage of time.
Failure windows are deliberately spaced: Silver tolerates roughly twice Gold's window because it costs half as much. The guide explicitly rejects different thresholds per class - Silver could end up performing better than Gold, and the gaps are hard to explain.
The four indicators
One benchmark to rule them all: every class is measured against the same four thresholds, per VM, as 5-minute averages.
The agreement
Live preview. Copy it as Markdown and take it to the meeting.
| Class | Price | CPU | Availability | Performance |
|---|---|---|---|---|
| Gold | 1.00x | 1:1 | 99.975% | 99.9% |
| Silver | 0.50x | 1:2 | 99.95% | 99.8% |
| Bronze | 0.25x | 1:4 | 99.9% | 99.6% |
Every VM is scored each 5 minutes against cpu ready ≤ 2.5%, memory contention ≤ 1%, disk latency ≤ 10ms and network transmit dropped packets ≤ 0%. The monthly figure is the percentage of intervals each VM passes. Crossing any one indicator fails the interval; failure is binary.
Measured per VM by Guest OS response; a running but isolated VM fails. Gold 99.975% (10.8 min/month), Silver 99.95% (21.6 min/month) and Bronze 99.9% (43.2 min/month). Customer-initiated reboots, guest-caused downtime and in-window maintenance are excluded. Performance has no maintenance window: there is no scheduled slow time.
Compliance: 99.95% for every class - securing everyone is in the provider's interest. A breached month is remedied as a credit on the next billing cycle.