Availability and SLOs

01The problem

Tallybook promises customers that the app is available "99.9% of the time". Over June, July and August, there were six outages. Nobody had added them up against the promise, and nobody could say which change would prevent the most downtime: more web servers, a better deployment process, or a different database setup.

02The concept

SLOs and error budgets

A service level objective (SLO) is an availability target, such as 99.9% a month. The allowed downtime is the error budget: 0.1% of a 30-day month is 43.2 minutes. When the budget is spent, reliability work takes priority over new features.

Components in series

If the app needs the load balancer and the web tier and the API and the database, its availability is the product of theirs. Every component in the chain lowers the total.

Components in parallel

If any one of several redundant copies is enough, the chance they're all down at once is the product of their unavailabilities. Two copies at 99.5% each give 1 − 0.005² = 99.9975%.

Single points of failure

A component with no redundancy (one database in one zone) often dominates downtime, no matter how many web servers you add.

03Example

Downtime by month against a 99.9% SLO:

import pandas as pd

base = "https://academy.cloudtechanalytics.com/datasets/cloud/"
outages = pd.read_csv(base + "outages.csv", parse_dates=["start"])
outages["month"] = outages["start"].dt.strftime("%Y-%m")

minutes_in_month = {"2026-06": 30 * 1440, "2026-07": 31 * 1440, "2026-08": 31 * 1440}
monthly = outages.groupby("month")["minutes"].sum().to_frame("downtime_min")
monthly["budget_min"] = [round(minutes_in_month[m] * 0.001, 1) for m in monthly.index]
monthly["availability_pct"] = [round(100 * (1 - d / minutes_in_month[m]), 3) for m, d in zip(monthly.index, monthly["downtime_min"])]
print(monthly)
outages.groupby("component")["minutes"].sum().sort_values(ascending=False)

downtime_min  budget_min  availability_pct
month
2026-06            69        43.2            99.840
2026-07            86        44.6            99.807
2026-08           102        44.6            99.772
component
Database         102
Web              102
Load balancer     31
API               22
Name: minutes, dtype: int64

Tallybook missed its target every month. The database and the web tier at month-end account for most of the downtime. Now the design arithmetic, with illustrative availabilities for each component:

def series(*parts):
    total = 1.0
    for p in parts:
        total *= p
    return total

def parallel(p, copies):
    return 1 - (1 - p) ** copies

LB, SERVER, DB_SINGLE_ZONE, DB_MULTI_ZONE = 0.9999, 0.995, 0.995, 0.9995   # illustrative

designs = {
    "now: 6 web, 4 API, single-zone DB": series(LB, parallel(SERVER, 6), parallel(SERVER, 4), DB_SINGLE_ZONE),
    "multi-zone DB": series(LB, parallel(SERVER, 6), parallel(SERVER, 4), DB_MULTI_ZONE),
    "multi-zone DB, 2 web, 2 API": series(LB, parallel(SERVER, 2), parallel(SERVER, 2), DB_MULTI_ZONE),
}
for name, a in designs.items():
    print(f"{name}: {a:.5%}  (about {(1 - a) * 30 * 1440:.0f} minutes down in a 30-day month)")

now: 6 web, 4 API, single-zone DB: 99.49005%  (about 220 minutes down in a 30-day month)
multi-zone DB: 99.94000%  (about 26 minutes down in a 30-day month)
multi-zone DB, 2 web, 2 API: 99.93501%  (about 28 minutes down in a 30-day month)

With a single-zone database, the design can't reach 99.9% however many servers it has: the database alone allows more downtime than the whole budget. A multi-zone database (a standby copy in a second zone that takes over automatically) moves the design comfortably past the target. Notice too that the month-end outages weren't a design-availability problem at all: every server was up, there just weren't enough of them (lesson 6).

04Walkthrough

1. Run the cells. Which outages would a multi-zone database have prevented?
2. Which outage would a gradual deployment (one server at a time) have prevented? Which would certificate expiry alerts have prevented?
3. Calculate the availability with 3 API servers instead of 4. Does it matter?
4. Write the reliability plan (the task below).

05Practice

06Check your understanding