Numbers to Know (cheatsheet) - Useful during System Design Interviews

numbers_to_know.md

Latency Numbers

Operation	$ns$ (nano)	$\mu s$ (micro)	$ms$ (milli)	notable	1Bx
L1 cache reference	$0.5 ns$				$0.5 s$
Branch mispredict	$5 ns$				$5 s$
L2 cache reference	$7 ns$			14x L1 cache	$7 s$
Mutex lock/unlock	$25 ns$				$25 s$
Main Memory reference	$100 ns$			20x L2 cache, 200x L1 cache	$1.6\ mins$
Compress 1K bytes with Zippy	$3,000 ns$	$3 µs$	$0.003ms$		$50\ mins$
Send 1K bytes over 1Gbps net	$10,000 ns$	$10 µs$	$0.01ms$		$5.5\ hours$
Read 4K randomly from SSD	$150,000 ns$	$150 µs$	$0.15ms$	$\approx1GB/s$ SSD	$1.7\ days$
Read 1MB sequentially from Mem	$250,000 ns$	$250 µs$	$0.25 ms$	$\approx4GB/s$ Mem	$2.9\ days$
Round trip within same datacenter	$500,000 ns$	$500 µs$	$0.5ms$		$5.8\ days$
Read 1MB sequentially from SSD	$1,000,000 ns$	$1,000 µs$	$1ms$	$\approx1GB/s$ SSD (4x Mem)	$11.6\ days$
Read DB row from cache			$1ms - 5ms$
Read DB row from disk			$5ms - 30ms$
Write DB row			$5ms - 15ms$
Disk seek	$10,000,000 ns$	$10,000 µs$	$10 ms$	20x datacenter roundtrip	$16.5\ weeks$
Read 1MB sequentially from Disk	$20,000,000 ns$	$20,000 µs$	$20 ms$	$50MB/s$ (80x Mem / 20s SSD)	$7.6\ months$
Send pack et CA->Netherlands->CA	$150,000,000 ns$	$150,000 µs$	$150 ms$		$4.75\ years$

Note

The 1Bx column represents the values multiplied by 1 billion, to help mentally visualize the scale and proportions.

Also notable

• Average server GOOD/BAD response time: $~250ms - ~1.5s$
• Average server requests: $1k/sec - 10k/sec$ - Dependent on what it does
• Internet users: $\approx5.5B$
• Average internet user daily internet: $\approx2.5h$
• Average size of internet packets: $\approx1.5KB$
• Average network bandwidth modern server: $\approx25\ Gbps$

Availability

%	Downtime/day	Downtime/week	Downtime/month	Downtime/year
$99%$	$14.40\ minutes$	$1.68\ hours$	$7.31\ hours$	$3.65\ days$
$99.99%$	$8.64s$	$1.01\ minutes$	$4.38\ minutes$	$52.60\ minutes$
$99.999%$	$864ms$	$6.05s$	$26.3s$	$5.26\ minutes$
$99.9999%$	$86.4ms$	$604.8ms$	$2.63s$	$31.56s$

Units

$ns$ nanosecond	$\mu s$ microsecond	$ms$ millisecond	$s$ second
$1ns$	$10^{-3}\mu s$	$10^{-6}ms$	$10^{-9}s$
$1,000ns$	$1\mu s$	$10^{-3}ms$	$10^{-6}s$
$1,000,000ns$	$1,000\mu s$	$1ms$	$10^{-3}s$
$1,000,000,000ns$	$1,000,000\mu s$	$1,000ms$	$1s$

Typical “components”

Redis

Configuration Type	ops/sec	read/write latency	notable
Single Instance	$100k/300k$	$\approx0.1ms - 1ms$	Not really relevant, as rarely one would deploy a single instance in production
Cluster Scaling (N instances)	$200k$ per instance	$\approx0.1ms - 1ms$	theoretically unbounded, but adds $\approx1ms - 2ms$ of latency per network hop
Redis Labs benchmark	$1M$		Redis Labs press release documents reaching $1.2M$ transactions/sec with $<1ms$ latency: avoid during interview

Postgres

Configuration Type	transactions/sec	read latency	write latency	large tables JOIN
Modern single Servers	$10k - 50k$	$\approx 10ms - 100ms$	$\approx1ms - 5ms$
Replication, partitioning, indexing, connection pooling	$100k+$	$\approx0.1ms - 5ms$	$\approx 0.5ms - 2ms$	$\approx 10ms - 100ms$

Kafka

Configuration type	rec/sec	Bytes/sec
Single Broker	$200k - 500k$	$\approx 100MB/s - 250MB/s$
Typical production Cluster	$1M - 5M$	$\approx 1GB/s - 2GB/s$

AWS SQS

Configuration type	msg/sec	notable
Standard	Unlimited: millions of msg/sec
FIFO	$\approx 300msg/s - 3000msg/sec$	Range depends on the use of batching (dramatically improves performance)

AWS DynamoDB

Configuration type	writes/sec	latency
on-demand	$100k - 1M+$	$\approx1ms - 10ms$	Auto-scaling, partitioned; effectively unlimited if no hot partitions.

Credit

Originally by Peter Norvig: http://norvig.com/21-days.html#answers