# Availability in Numbers: Nines, Downtime, and Combined Parts

> Turn uptime percentages into allowed downtime, then see how parts in sequence and in parallel change the total reliability of a service.

**In short:** Availability is the share of time a system works, and each extra nine in the percentage cuts the downtime it is allowed.

From "System design", a free course by Venkat R. Murty: https://designervenkat.online/learn/system-design/availability-in-numbers

## What You Will Learn

Many people say a website is either up or down. Real teams need more detail than that. They want to know how often a service works, and how much it may be down in a year. This lesson shows how to write that as a percentage, turn the percentage into minutes and hours, and combine the numbers of several parts that work together.

You will also see why a chain of parts is always weaker than its weakest link, and why a spare copy can make the whole system far safer. A worked example uses a food delivery checkout, and a short quiz at the end lets you test the ideas on new cases.

- **Availability:** The share of time a system is working and able to do its job, usually written as a percentage.
- **Uptime:** The time a system is up and doing its job.
- **Downtime:** The time a system is down, or cannot do its job.
- **Nines:** A short name for an availability level. 99.9% is called three nines because it has three 9s in a row.
- **Sequence:** Parts that must all work at the same time, so one failed part stops the whole chain. Also called components in series.
- **Parallel:** Spare copies of a part, where the group keeps working as long as at least one copy works.

## What Is Availability as a Share of Time?

Availability is the uptime divided by the total time. Uptime is the part of the time when the system works. Downtime is the rest. Add uptime and downtime and you get the total time, so the formula can also be written with downtime in it.

$$
A = \frac{\text{uptime}}{\text{uptime} + \text{downtime}}
$$

In words: availability equals uptime divided by the sum of uptime and downtime. A service that works for 999 out of every 1000 hours has 99.9% availability, because 999 divided by 1000 is 0.999.

## How Do You Read the Nines?

Teams call the percentage the nines, because the count of 9s shows how strict the target is. Each extra nine cuts the allowed downtime by about ten times. Three common levels are shown below, with the yearly downtime each one allows.

- **99.9% (three nines):** About 8.76 hours of downtime per year. This is a common target for business web services.
- **99.99% (four nines):** About 52.6 minutes of downtime per year. Going from three nines to four cuts the yearly budget tenfold.
- **99.999% (five nines):** About 5.26 minutes of downtime per year. Meeting this level usually needs spare copies in more than one place.

To get these numbers yourself, multiply the downtime share by the hours in a year. There are 365 days times 24 hours, which is 8,760 hours. At 99.99%, the downtime share is 0.0001, and 0.0001 times 8,760 is 0.876 hours, or about 52.6 minutes.

1. **Write the downtime share:** Subtract the availability from 1. For 99.99%, that gives 0.0001.
2. **Multiply by the hours in a year:** Use 8,760 hours. The result is 0.876 hours.
3. **Convert to minutes:** Multiply hours by 60. The result is about 52.6 minutes per year.

### Downtime Per Day and Per Month

The same method works for any period. Take the downtime share and multiply it by the hours in that period. For a 30-day month, there are 720 hours. At 99.9%, the allowed downtime is 0.001 times 720, which is 0.72 hours, or about 43.2 minutes.

> **Tip:** Teams often track a monthly budget as well as a yearly one. A small monthly budget shows trouble sooner, because one bad night can use up a large share of it, even when the yearly total still looks healthy.

## How Do Components in Sequence Affect Availability?

Many requests need several parts to work in order. If a checkout needs a web server and a database, both must be up for an order to finish. That is a sequence. To find the total availability of a sequence, you multiply the availability of each part. The total is always lower than the weakest part.

$$
A_{\text{total}} = A_1 \times A_2
$$

In words: total availability equals A1 times A2. For two parts at 99.9% each, the total is 0.999 times 0.999, which equals 0.998001. That is about 99.8%, which is lower than either part alone.

### Why Do Chains Get Weaker?

Each new required step multiplies the result again, so the total keeps shrinking. Five services in a row, each at 99.9%, give about 99.5% together. That is much lower than one 99.9% service on its own, so every extra required step costs some reliability.

> **Tip:** For each step, ask what happens if that step fails. If the user loses the action, the step belongs in the sequence, and it counts against the total. Remove needless steps from the required path when you can.

## How Do Components in Parallel Affect Availability?

Parallel parts are spare copies. The group fails only if every copy fails at the same time, so it is far more reliable than any one copy. This use of spares is called redundancy. The formula first finds the chance that each copy is down, multiplies those chances together, and then subtracts the result from one.

$$
A_{\text{total}} = 1 - (1 - A_1)(1 - A_2)
$$

In words: total availability equals one minus the product of one minus A1 and one minus A2. Two copies at 99.9% each are down together only when both are down, which has a chance of 0.001 times 0.001. That is one in a million, so the pair gives 99.9999% availability.

### When the Parallel Math Holds

The formula assumes the copies fail separately. If two copies sit in one building and share one power supply, a single power cut can take both down. The real result is then much worse than the formula suggests.

- **Separate causes:** The copies should not share a single weak point, such as one power line, one network switch, or one faulty software update.
- **Working detection:** Something must notice that a copy has failed and send work to a healthy copy. A dead copy that still looks alive adds no safety.

The short program below checks the single-copy and pair numbers. It uses exact fractions, so the printed results are always the same.

```python
from fractions import Fraction

part = Fraction(999, 1000)  # one copy, 99.9%
pair = 1 - (1 - part) ** 2  # two copies side by side
print(f'one copy: {float(part * 100):.1f}%')
print(f'two copies: {float(pair * 100):.4f}%')

```

Output:

```text
one copy: 99.9%
two copies: 99.9999%

```

## Worked Example: Food Delivery Checkout

A food delivery app has a checkout screen. To take payment and save the order, the checkout needs a web server and a database. Each one is 99.9% available. We will work out the checkout's availability first with one copy of each, then with two copies of each.

1. **Map the chain:** The checkout needs both parts at the same time, so they are in sequence. If either one fails, the order fails.
2. **Multiply for one copy each:** Multiply 0.999 by 0.999. The result is 0.998001, or about 99.8%.
3. **Convert to downtime:** The downtime share is 0.001999. Multiply by 8,760 hours and the checkout allows about 17.5 hours of downtime per year.
4. **Add a second copy of each:** Each pair of copies now has 99.9999% availability. The checkout needs both pairs, so multiply them: about 99.9998%. That allows roughly 63 seconds of downtime per year.
5. **Check a single change:** If you double only the database, the web server still limits the checkout to about 99.9%. The weakest part in a sequence sets the ceiling until you fix it.

The next program repeats the sequence result and the yearly downtime for the checkout with one copy of each part.

```python
from fractions import Fraction

hours_per_year = 365 * 24

checkout = Fraction(999, 1000) * Fraction(999, 1000)
down = (1 - checkout) * hours_per_year
print(f'checkout: {float(checkout * 100):.4f}%')
print(f'downtime: {float(down):.2f} hours per year')

```

Output:

```text
checkout: 99.8001%
downtime: 17.51 hours per year

```

## Check Your Understanding

Try these two new cases. Each one needs the ideas above, so think through the numbers before you choose.

**Quick check:** Your login service calls an API server rated 99.9% and a database rated 99.95%. Both must work for a login to succeed. What is the approximate combined availability, and would adding a second identical database in parallel raise or lower it?

A. About 99.95%, and a second database would lower it
B. About 99.85%, and a second database would raise it
C. About 99.90%, and a second database would raise it
D. About 99.99%, and a second database would lower it

**Answer:** B. The two parts are in sequence, so 0.999 times 0.9995 gives about 0.9985, or 99.85%. A second database forms a parallel pair, which is far more reliable than one database. The chain then rises to about 99.90%, so redundancy raises the result.

**Quick check:** A payment step relies on one provider rated 99.5%. The team runs a second identical provider side by side, and the step works if either provider is up. About how much downtime per year does the step now allow?

A. About 44 hours
B. About 13 minutes
C. About 4 hours
D. About 1 hour

**Answer:** B. One provider alone allows about 0.5% of 8,760 hours, which is roughly 44 hours. With two providers, the step is down only when both are down: 0.005 times 0.005 is 0.000025. Multiply by 8,760 hours to get about 0.22 hours, or roughly 13 minutes per year. This assumes the providers fail separately.

## Common Mistakes

- **Looking only at the yearly number:** A yearly total can hide one long outage. Many teams also track monthly or weekly budgets, so a single bad incident does not go unnoticed.
- **Leaving out planned downtime:** Some sites skip planned maintenance when they add up downtime. A strict target counts it too, because users cannot use the service during those hours either.
- **Assuming spares always save you:** Parallel math needs copies that fail separately and a way to route work around a dead copy. Without both, a spare adds cost and little safety.

## Key Takeaways

- **Nines are a time budget:** An availability percentage becomes a set number of minutes or hours of allowed downtime. Each extra nine cuts that budget about tenfold.
- **Chains lose reliability:** Parts in sequence multiply together, so the total is always below the weakest part.
- **Spares gain reliability:** Parallel copies only fail together when all of them fail, so the group is far more reliable, as long as the copies fail separately.

> **Tip:** When you review a design, write the availability of each part, mark which parts are in sequence and which are in parallel, and calculate the total before you trust any target.

## Sources

- [Uptimeify: Uptime Calculator, SLA Nines and Downtime](https://uptimeify.io/blog/calculate-uptime-sla-nines)
- [OpenObserve: How Many Nines Do You Actually Need? A Practical Guide](https://openobserve.ai/blog/how-many-nines-do-you-actually-need/)
- [Control Plane: What Is High Availability? How It Works and How Many Nines You Need](https://controlplane.com/blog/post/what-is-high-availability-how-it-works-and-how-many-nines-you-actually-need)
- [Wikipedia: High availability](https://en.wikipedia.org/wiki/High_availability)
- [Wikipedia: Availability (series and parallel components)](https://en.wikipedia.org/wiki/Availability)
- [The System Design Interview Handbook: 3.2 Availability and Reliability](https://www.systemdesigninterview.com/guides/system-design-interview-handbook/32-availability-reliability)
- [Karan Pratap Singh: System Design course, Availability](https://www.karanpratapsingh.com/courses/system-design/availability)
- [EventHelix: System Reliability and Availability Calculation](https://www.eventhelix.com/RealtimeMantra/FaultHandling/system_reliability_availability.htm)
- [Ramly: System Availability Calculator, Series and Parallel](https://ramly.io/tools/availability-calculator)
- [Worldstream: Understanding uptime percentages, what the nines mean](https://www.worldstream.com/en/knowledge-base/articles/understanding-uptime-percentages)
- [Pingping: Uptime SLA Explained, 99.9% vs 99.99% vs 99.999%](https://pingping.io/guides/uptime-sla-explained)
- [arXiv survey: Reliability and Availability in Virtualized Networks (RBD series and parallel rules)](https://arxiv.org/pdf/2503.22034)
