Cost governance and unit costs

01The problem

A one-off clean-up saves money once. Six months later, without a process, the waste is back: new test servers, new snapshots, new untagged resources. And a rising bill on its own says nothing about whether Tallybook is wasting money or simply growing.

FinOps (cloud financial operations) is the practice of making cloud spending visible, owned and tied to business value, continuously.

02The concept

Ownership and showback

Every resource has a team tag, and each team sees its own monthly cost (showback). Untagged resources are reported until they're claimed.

Budgets and alerts

A monthly budget per team and environment, with alerts at, for example, 80% and 100%, and an alert on unusual daily spikes (the same control-limit idea as monitoring).

Unit costs

Divide cost by a measure of business activity: cost per thousand requests, per invoice sent, per active customer. If the bill grows but unit cost stays flat or falls, the growth is the business growing. If unit cost rises, look for waste or a design problem.

The review cycle

Monthly: review showback, unit costs and the waste report. Quarterly: review commitments, rightsizing and the architecture.

03Example

Unit cost of the production web service: production cost per thousand web requests, by day, for August:

import pandas as pd

base = "https://academy.cloudtechanalytics.com/datasets/cloud/"
billing = pd.read_csv(base + "billing.csv", parse_dates=["date"])
traffic = pd.read_csv(base + "web_traffic.csv", parse_dates=["hour"])

prod = billing[(billing["environment"] == "production") & (billing["date"].dt.month == 8)].groupby("date")["cost_usd"].sum()
requests = traffic.groupby(traffic["hour"].dt.normalize())["requests"].sum()
unit = pd.DataFrame({"prod_cost_usd": prod, "requests": requests})
unit["usd_per_1000_requests"] = unit["prod_cost_usd"] / unit["requests"] * 1000
print("August average: 
quot;, round(unit["prod_cost_usd"].sum() / unit["requests"].sum() * 1000, 4), "per 1,000 requests")
unit["weekday"] = unit.index.day_name()
unit.groupby("weekday")["usd_per_1000_requests"].mean().round(4).sort_values()

August average: $ 0.3586 per 1,000 requests
weekday
Friday       0.2804
Monday       0.2905
Tuesday      0.3033
Wednesday    0.3134
Thursday     0.3145
Saturday     0.6761
Sunday       0.7038
Name: usd_per_1000_requests, dtype: float64

Weekends cost far more per request: the fixed fleet runs at full price for half the traffic. That's the autoscaling saving from lesson 6, seen from the business side. Now showback by team for August, including the untagged spend:

aug = billing[billing["date"].dt.month == 8]
showback = aug.groupby(aug["team"].fillna("UNTAGGED"))["cost_usd"].sum().round(2).sort_values(ascending=False)
showback

team
platform     3725.31
UNTAGGED      901.79
invoicing     670.02
payments      348.56
marketing       0.94
Name: cost_usd, dtype: float64

04Walkthrough

1. Run the cells. Which day in August had the lowest cost per thousand requests, and why?
2. Set a monthly budget for each team, 10% above its August spend, and an alert at 80%.
3. Choose a better unit for Tallybook's business than requests (invoices sent? paying customers?) and say what data you'd need.
4. Write the governance policy (the task below).

05Practice

06Check your understanding