Latency Numbers

Every architecture decision has a latency justification behind it

These are Jeff Dean's numbers — the standard reference every interviewer expects you to know. Memorise the order of magnitude, not the exact value.

The numbers#

Operation	Latency	In human terms
L1 cache hit	0.5 ns	fastest thing a CPU can do
L2 cache hit	7 ns	14× slower than L1
RAM access	100 ns	200× slower than L1
Read 1MB from RAM	250 µs	sequential, fast
SSD random read (4KB)	150 µs	1,500× slower than RAM access
Read 1MB from SSD	1 ms	sequential SSD read
HDD disk seek	10 ms	moving mechanical parts
Read 1MB from HDD	30 ms	sequential HDD read
Network — same datacenter	0.5 ms	fast, synchronous calls are fine
Network — cross-region (US–EU)	100–150 ms	too slow for synchronous critical path

What these numbers actually mean for design#

RAM is 1,000× faster than SSD. This is why you cache hot data in memory. A Redis lookup (RAM) at 0.1ms vs a Postgres disk read (SSD) at 5ms — that's a 50× difference in the real world. At 1M reads/sec, the difference between serving from cache vs going to disk is the difference between a system that works and one that falls over.

SSD is 100× faster than HDD. This is why databases run on SSDs, not spinning disks. HDD is only viable for cold storage (S3, archival) where you accept high latency in exchange for cheap cost.

Same-datacenter network is 0.5ms. Synchronous service-to-service calls within the same region are fine. An app server calling Redis, calling a DB shard, calling an auth service — all within the same DC, all sub-millisecond.

Cross-region is 100–150ms. A synchronous call from US to EU on the critical path adds 100ms of unavoidable physics latency. This is why read replicas exist in each region — you cannot serve users globally from a single datacenter without cross-region latency killing your p99.

The key ratios to remember#

RAM vs SSD:        ~1,000× faster
SSD vs HDD:        ~100× faster
Same DC vs cross-region: ~300× faster (0.5ms vs 150ms)

If an interviewer asks "why cache in Redis instead of reading from DB every time?" — the answer is 1,000×. That's the ratio. That's the justification.

Derived latency estimates for common operations#

These are not memorised — they are derived from the fundamentals above:

Redis read (cache hit):          0.1 – 0.5 ms   (RAM access + network within DC)

Redis write:                     0.1 – 0.5 ms
Postgres read (indexed, hot):    1   – 5 ms     (SSD + query execution)

Postgres write (INSERT):         5   – 10 ms    (SSD write + WAL flush)

Cassandra write:                 0.5 – 2 ms     (memtable, no disk on hot path)

Cassandra read:                  5   – 10 ms    (may check multiple SSTables)

MongoDB read (indexed):          1   – 5 ms

S3 GET (object storage):         50  – 200 ms   (network + S3 internals)

Kafka produce (ack=1):           5   – 15 ms    (broker write + network)

Kafka end-to-end (produce→consume): 10 – 50 ms

Interview framing

"Redis is sub-millisecond — 0.1 to 0.5ms. Postgres indexed read is 1–5ms. SSD is 1,000× faster than HDD, RAM is 1,000× faster than SSD. Cross-region adds 100–150ms of unavoidable physics — that's why you need regional replicas for global systems."