Sections
Section · Reference
Power of Two
Data volumes in distributed systems scale by powers of two. Internalizing the five magnitudes below makes back-of-the-envelope arithmetic almost free — you can convert between bytes, megabytes, and petabytes without reaching for a calculator.
| Power | Approximate value | Full name | Short |
|---|---|---|---|
| 10 | ≈ 1 thousand | 1 kilobyte | 1 KB |
| 20 | ≈ 1 million | 1 megabyte | 1 MB |
| 30 | ≈ 1 billion | 1 gigabyte | 1 GB |
| 40 | ≈ 1 trillion | 1 terabyte | 1 TB |
| 50 | ≈ 1 quadrillion | 1 petabyte | 1 PB |
Estimate as a power of 2
Convert bytes
Section · Reference + Live Check
Latency Numbers
Latency in distributed systems spans nine orders of magnitude — from the half-nanosecond of an L1 cache hit to the 150 milliseconds of an intercontinental round-trip. The table below collects the points architects estimate against; the bar chart makes the magnitudes visceral; the live check measures your actual round-trip to this backend so you can compare reality against the reference.
Operation latencies — typical values, modern hardware
| Operation | Latency | Category |
|---|---|---|
| L1 cache reference | 0.50 ns | CPU |
| Branch mispredict | 5 ns | CPU |
| L2 cache reference | 7 ns | CPU |
| Mutex lock / unlock | 25 ns | CPU |
| Main memory reference | 100 ns | Memory |
| Compress 1 KB (snappy) | 3.0 µs | Compute |
| Send 1 KB over 1 Gbps net | 10 µs | Network |
| Read 4 KB random from SSD | 150 µs | Storage |
| Read 1 MB sequential from RAM | 250 µs | Memory |
| Round trip in same datacenter | 500 µs | Network |
| Read 1 MB sequential from SSD | 1.0 ms | Storage |
| Disk seek | 10 ms | Storage |
| Read 1 MB sequential from disk | 30 ms | Storage |
| Intercontinental round-trip | 150 ms | Network |
Log scale — bar width proportional to log₁₀(latency)
Real-world latency check
Fire 20 sequential pings to /api/chapters/ping/ and compare against the reference table's round-trip rows.
Section · Reference
Availability Numbers
Availability is measured in 9s. Each additional 9 cuts the allowed downtime by an order of magnitude — 99% buys you almost four days a year, 99.999% (“five nines”) leaves you about five minutes. Pick a target, see what you’re committing to.
SLA tiers — downtime budget per period
| Availability % | Per day | Per month | Per year |
|---|---|---|---|
| 99% | 14m 24s | 7h 18m | 3d 15h 36m |
| 99.9% | 1m 26s | 43m 49s | 8h 45m 36s |
| 99.99% | 8.64s | 4m 22s | 52m 33s |
| 99.999% | 0.86s | 26.30s | 5m 15s |
| 99.9999% | 0.086s | 2.63s | 31.56s |
Pick a target
99.9900%Per day
8.64s
Per month
4m 19s
Per year
52m 36s
Section · Workshop
Estimation Workshop
Take a small set of assumptions about a system — active users, post frequency, media payload — and derive the operationally-meaningful numbers: write QPS at peak, daily storage, multi-year storage. Walk the canonical microblog example one step at a time, or punch in your own numbers and watch every derivation update live.
Pre-loaded assumptions · canonical microblog
Daily active users
Formula · DAU = MAU × daily_fraction
Substituted · 300M × 0.5
→ 150M
Half of monthly users return on a given day. This fraction captures stickiness — chat apps run ~70%, news sites 10–20%. Get it wrong and your QPS estimate is off by the same factor.
Step 1 / 7
Sections
Section · Reference
Power of Two
Data volumes in distributed systems scale by powers of two. Internalizing the five magnitudes below makes back-of-the-envelope arithmetic almost free — you can convert between bytes, megabytes, and petabytes without reaching for a calculator.
| Power | Approximate value | Full name | Short |
|---|---|---|---|
| 10 | ≈ 1 thousand | 1 kilobyte | 1 KB |
| 20 | ≈ 1 million | 1 megabyte | 1 MB |
| 30 | ≈ 1 billion | 1 gigabyte | 1 GB |
| 40 | ≈ 1 trillion | 1 terabyte | 1 TB |
| 50 | ≈ 1 quadrillion | 1 petabyte | 1 PB |
Estimate as a power of 2
Convert bytes
Section · Reference + Live Check
Latency Numbers
Latency in distributed systems spans nine orders of magnitude — from the half-nanosecond of an L1 cache hit to the 150 milliseconds of an intercontinental round-trip. The table below collects the points architects estimate against; the bar chart makes the magnitudes visceral; the live check measures your actual round-trip to this backend so you can compare reality against the reference.
Operation latencies — typical values, modern hardware
| Operation | Latency | Category |
|---|---|---|
| L1 cache reference | 0.50 ns | CPU |
| Branch mispredict | 5 ns | CPU |
| L2 cache reference | 7 ns | CPU |
| Mutex lock / unlock | 25 ns | CPU |
| Main memory reference | 100 ns | Memory |
| Compress 1 KB (snappy) | 3.0 µs | Compute |
| Send 1 KB over 1 Gbps net | 10 µs | Network |
| Read 4 KB random from SSD | 150 µs | Storage |
| Read 1 MB sequential from RAM | 250 µs | Memory |
| Round trip in same datacenter | 500 µs | Network |
| Read 1 MB sequential from SSD | 1.0 ms | Storage |
| Disk seek | 10 ms | Storage |
| Read 1 MB sequential from disk | 30 ms | Storage |
| Intercontinental round-trip | 150 ms | Network |
Log scale — bar width proportional to log₁₀(latency)
Real-world latency check
Fire 20 sequential pings to /api/chapters/ping/ and compare against the reference table's round-trip rows.
Section · Reference
Availability Numbers
Availability is measured in 9s. Each additional 9 cuts the allowed downtime by an order of magnitude — 99% buys you almost four days a year, 99.999% (“five nines”) leaves you about five minutes. Pick a target, see what you’re committing to.
SLA tiers — downtime budget per period
| Availability % | Per day | Per month | Per year |
|---|---|---|---|
| 99% | 14m 24s | 7h 18m | 3d 15h 36m |
| 99.9% | 1m 26s | 43m 49s | 8h 45m 36s |
| 99.99% | 8.64s | 4m 22s | 52m 33s |
| 99.999% | 0.86s | 26.30s | 5m 15s |
| 99.9999% | 0.086s | 2.63s | 31.56s |
Pick a target
99.9900%Per day
8.64s
Per month
4m 19s
Per year
52m 36s
Section · Workshop
Estimation Workshop
Take a small set of assumptions about a system — active users, post frequency, media payload — and derive the operationally-meaningful numbers: write QPS at peak, daily storage, multi-year storage. Walk the canonical microblog example one step at a time, or punch in your own numbers and watch every derivation update live.
Pre-loaded assumptions · canonical microblog
Daily active users
Formula · DAU = MAU × daily_fraction
Substituted · 300M × 0.5
→ 150M
Half of monthly users return on a given day. This fraction captures stickiness — chat apps run ~70%, news sites 10–20%. Get it wrong and your QPS estimate is off by the same factor.
Step 1 / 7