Numbers Worth Knowing for System Design Interviews
Published
A lot of system design prep treats numbers as trivia to memorise - as if reciting that S3 offers eleven nines of durability is the point. It isn't. The point is having a rough mental model of scale, so that when you say "we'll need about 40TB for this" or "that's going to add 100ms" you're reasoning from something, not guessing. Here are the ones that come up most often, with the context that actually makes them useful.
Availability - the nines
99.9% uptime sounds almost the same as 99.99%, but the gap between them is about 8.75 hours of downtime a year versus 52 minutes. The useful thing to know isn't the table itself, it's that each additional nine tends to cost disproportionately more engineering effort - going from three nines to four is often a redesign (multi-AZ, automated failover), not a config change.
| Availability | Downtime / year | Downtime / month | Downtime / day |
|---|---|---|---|
| 99% | 3.65 days | 7.31 hours | 14.4 minutes |
| 99.9% | 8.77 hours | 43.8 minutes | 1.44 minutes |
| 99.99% | 52.6 minutes | 4.38 minutes | 8.64 seconds |
| 99.999% | 5.26 minutes | 26.3 seconds | 864 milliseconds |
| 99.9999% | 31.5 seconds | 2.63 seconds | 86.4 milliseconds |
Storage estimation
The formula is simple - records per day × average record size × retention period - but the interview signal is whether you remember to multiply by the things that aren't in the formula: replication factor, indexes, and any derived data you're also storing. Forgetting the 3x from replication is a more common mistake than getting the base estimate wrong.
Byte sizes worth having memorised
Roughly: a tweet-sized message is a few hundred bytes, a typical JSON row is around 1KB, a compressed photo is somewhere between 200KB and 2MB, and a minute of video is tens of megabytes. These aren't precise, but they're enough to turn "500 million users posting twice a day" into a terabyte figure without stalling the conversation.
| Item | Approximate size |
|---|---|
| Tweet-length message | ~280 bytes |
| Typical JSON row / API response object | ~1 KB |
| Compressed photo | 200 KB - 2 MB |
| One minute of video | 10 - 50 MB |
Throughput and the read:write ratio
Daily active users gets you to an average requests-per-second figure, but peak load is usually several times the average, not the same number - plan for that multiplier explicitly rather than quietly forgetting it. The ratio that matters more than raw QPS, though, is read:write. Most systems are read-heavy, often by 100:1 or more, and that ratio is what actually drives whether you need a cache, a read replica, or both.
Latency numbers worth knowing
Memory access is nanoseconds, an SSD read is tens of microseconds, a round trip within the same datacentre is under a millisecond, and a cross-region round trip is 50-150ms. The context that matters: that cross-region number is a floor, not a target you can engineer away. If a design leans on synchronous cross-region calls, that latency is already baked in before you've written a line of code.
| Operation | Typical latency |
|---|---|
| Memory (RAM) access | ~100 nanoseconds |
| SSD read | 10 - 100 microseconds |
| Round trip, same datacentre | <1 millisecond |
| Round trip, cross-region | 50 - 150 milliseconds |
Durability and availability are different axes
Durability is "will the data still exist." Availability is "can I read it right now." S3's eleven nines of durability and roughly four nines of availability are answering two different questions, and it's a common tell when someone conflates them. A store can be extremely durable and briefly unavailable during a network partition - whether that's acceptable depends entirely on what you're building.
Network bandwidth is a hard ceiling
A single box's NIC - typically 1, 10, or 25 Gbps - puts a hard limit on how much traffic it can serve, independent of how much CPU or memory you throw at it. It's a useful sanity check before designing any kind of sharding or distribution: work out whether the traffic could physically fit through one machine before assuming it can't.
| NIC speed | Theoretical max throughput |
|---|---|
| 1 Gbps | ~125 MB/s |
| 10 Gbps | ~1.25 GB/s |
| 25 Gbps | ~3.1 GB/s |
None of this needs to be memorised to the decimal place. What's useful is having enough of a mental model that when a number gets thrown into a design - "we expect 10,000 requests per second" - you can immediately sense whether that's a one-box problem or a fifty-box problem, and why.
