Cockroach Labs Estate economics What idle nodes, provisioned bytes and peak-sized bills cost, worked out from your estate.
  1. Nodes 52 baseline
  2. Avg CPU utilization 15.0% idle hardware
  3. Storage provisioned 600 GB 100 GB of data
  4. Est. monthly saving 0%

Step 7 of 8 - Pricing

You pay for the area, not the rectangle

Provisioned capacity is a flat ceiling sized to the busiest fifteen minutes of the day, plus headroom for the peak you didn't forecast. The gap under that ceiling is hardware you rented and didn't use.

Run a day, 40% headroom Steady Business hours Batch spike

Provisioned capacity left unused

not started

Pick a day above. The ceiling is already drawn: it's sized to a peak of 139 vCPU plus 40%, and you pay for all of it whether the workload turns up or not. That's under the 416 vCPU the estate runs on today, which was bought a cluster at a time and works at 15.0%, not sized to this curve.

09719500:0006:0012:0018:0024:00vCPUProvisioned195 vCPUConsumed62 vCPU avgProvisioned capacity versus consumption over 24 hoursOver 24 hours, a ceiling of 195 vCPU, set above a peak of 139 vCPU, costs 4671 vCPU-hours, while the workload actually consumes 1498 vCPU-hours: 68% of the provisioned capacity goes unused.
Provisioned
0 vCPU-hours
Consumed
0 vCPU-hours
Paid for, not used
0 vCPU-hours
Read this chart as a table
Mean demand per hour against provisioned capacity, in vCPU
HourDemandProvisionedUnused
00:00 28 195 166
01:00 28 195 166
02:00 28 195 166
03:00 28 195 167
04:00 28 195 166
05:00 29 195 166
06:00 28 195 167
07:00 39 195 156
08:00 65 195 129
09:00 90 195 104
10:00 111 195 84
11:00 125 195 70
12:00 135 195 60
13:00 135 195 60
14:00 129 195 66
15:00 116 195 79
16:00 97 195 98
17:00 72 195 123
18:00 46 195 149
19:00 28 195 166
20:00 29 195 166
21:00 28 195 166
22:00 28 195 166
23:00 28 195 166
How this is worked out
  • Demand is this estate's 62 vCPU mean, shaped by the preset you picked, which also drives step 1.
  • Capacity is sized at the day's peak of 139 vCPU plus the 40% headroom you picked, giving a ceiling of 195 vCPU. The headroom is an assumption rather than a published figure, so it's yours to argue with.
  • That peak is the 10 clusters' own peaks added up, because a provisioned estate buys each of them separately. Step 1's aggregate peak of 87 vCPU is the later, smaller question: what one pool needs once the workloads share it.
  • A ceiling that high averages 32.1% CPU across the day, which is the whole day's 67.9% of unused capacity counted the other way round. The pooled cluster on step 1 runs at 35.5% against a 37.5% setpoint. Neither number sets the other: this one is how far above the peak you buy, that one is what a controller is told to hold.
Where the storage number comes from

3 Raft copies at 50% utilization (3/0.5 = 6), on top of the block store's own internal replication.

100 GB of logical data becomes 600 GB provisioned , a factor of 6× before the cost model sees it.