Schedules and pricing models

01The problem

Tallybook's twelve development servers run 24 hours a day, 7 days a week. The engineers use them roughly from 7am to 6pm on weekdays: about a third of the hours in a week. The rest is paid for and wasted.

Meanwhile, the production servers that genuinely run all the time are paid at the full on-demand price. Cloud providers offer large discounts to customers who commit to a year of use, and even bigger ones for spare capacity that can be taken back at short notice. Tallybook uses neither.

02The concept

Schedules

Stop non-production servers outside working hours and start them again in the morning, automatically. Their disks are kept, so nothing is lost. Engineers who need a server out of hours can start it themselves.

Pricing models

Commit only to the baseline

Commit to what runs every hour of the year (after rightsizing and clean-up), never to peaks. Over-commitment is paying for servers you no longer use.

03Example

The saving from scheduling the development servers to run 7am to 6pm on weekdays (August 2026 had 21 weekdays):

import pandas as pd

base = "https://academy.cloudtechanalytics.com/datasets/cloud/"
resources = pd.read_csv(base + "resources.csv")
util = pd.read_csv(base + "utilisation.csv", parse_dates=["hour"])

dev = resources[(resources["type"] == "vm") & resources["name"].str.startswith("dev-")]
HOURS_IN_AUGUST = 744
SCHEDULED_HOURS = 11 * 21          # 7am to 6pm, 21 weekdays

dev = dev.assign(current_usd=dev["hourly_usd"] * HOURS_IN_AUGUST, scheduled_usd=dev["hourly_usd"] * SCHEDULED_HOURS)
print("Development servers:", len(dev))
print("August cost now: 
quot;, round(dev["current_usd"].sum(), 2), "  with a schedule: 
quot;, round(dev["scheduled_usd"].sum(), 2))

u = util[util["resource_id"].isin(dev["resource_id"])]
in_hours = (u["hour"].dt.weekday < 5) & u["hour"].dt.hour.between(7, 17)
print("Average CPU in scheduled hours:", round(u.loc[in_hours, "cpu_pct"].mean(), 1), "  outside them:", round(u.loc[~in_hours, "cpu_pct"].mean(), 1))

Development servers: 12
August cost now: $ 595.2   with a schedule: $ 184.8
Average CPU in scheduled hours: 24.5   outside them: 2.0

The usage data confirms the servers do almost nothing outside the schedule. Now the production baseline and pricing models, with illustrative discounts:

prod = resources[(resources["type"] == "vm") & (resources["environment"] == "production")].copy()
prod["role"] = prod["name"].str.extract(r"prod-(\w+)-")[0]
monthly = prod.groupby("role")["hourly_usd"].sum() * 730

COMMITTED, SPOT = 0.35, 0.65     # illustrative discounts
options = pd.DataFrame({
    "on_demand": monthly,
    "committed_1yr": monthly * (1 - COMMITTED),
    "spot": monthly * (1 - SPOT),
}).round(2)
options

on_demand  committed_1yr    spot
role
api         584.0         379.60  204.40
web         438.0         284.70  153.30
worker      219.0         142.35   76.65

Web and API servers must always be available, so they suit a commitment (after rightsizing the API servers, as in lesson 3, so you don't commit to the old size). The invoice workers process jobs from a queue: if a spot server is reclaimed, its job goes back on the queue and another server picks it up. That makes them good spot candidates, with one on-demand server kept as a floor.

04Walkthrough

1. Run the cells. How much would a schedule save over a full year?
2. Redo the commitment calculation using the rightsized API size from lesson 3.
3. Explain why committing for the web servers' month-end peak would be a mistake.
4. Write the recommendation (the task below).

05Practice

06Check your understanding