Storage Estimation

The Formula

Every storage estimate is one multiplication with three honest add-ons:

 raw      = records/day × record size × days of retention
 +indexes ≈ ×1.2–2.0 depending on secondary indexes
 +replication ≈ ×3 (RF=3) or ×1.5 (erasure coding)
 +growth headroom ≈ keep ≤70% disk utilization
 
 usable capacity ≈ raw × index-factor × replication ÷ 0.7

Worked Example: RideShare Trips

Assumptions first (illustrative, stated):

 trip record:
   ids, geo endpoints, timestamps, fare, status, ratings refs
   ≈ 1 KB as a row (with overhead)
 
 20M trips/day, retained 5 years

Arithmetic:

 daily raw        = 20M × 1 KB              = 20 GB/day
 yearly           = 20 GB × 365             ≈ 7.3 TB/year
 5-year retention = 7.3 TB × 5              ≈ 36.5 TB raw
 + indexes ×1.5                             ≈ 55 TB
 + replication ×3                            ≈ 165 TB
 ÷ 0.7 max utilization                       ≈ 235 TB provisioned at steady state

One table, quarter-petabyte class when fully loaded — numbers like this decide whether trips stay in the transactional cluster forever or age out to object storage.

The Other Tables (fast pass)

Data	Record	Volume/day	Notes
Users	2 KB	negligible growth	50M total ≈ 100 GB once
Location pings	~200 B	4B/day ≈ 800 GB/day	ephemeral — TTL minutes; net storage tiny
Trip events stream	~500 B	~60M events ≈ 30 GB/day	Kafka retention hours–days only

Location data is the lesson’s twist: highest write volume, near-zero stored bytes because expiry is a design decision made before sizing.

Growth Shape Matters

Linear accumulation vs compounding usage:

 linear (fixed users):          compounding (user growth 2x/yr):
 year 5 = 5 × year-1            year 5 ≈ 32 × year-1  ← 2^5
 
 storage plans that assume linearity die quietly in compounding products
 → re-run estimates quarterly against actuals

Retention Is an Architecture Decision

The cheapest storage is deleted storage:

 TIERED RETENTION (typical pattern)
 hot    0–90 days   Postgres/SSD     full query speed     ~0.5 TB tier
 warm   90d–2yr     compressed rows  occasional queries   ~10 TB tier
 cold   2yr+        object storage   compliance reads     pennies/TB/month
 
 same data, three price points — retention policy IS cost architecture

Legal/compliance floors set minimums (“financial records 7 years”); product need sets everything above the floor.

Interview Framing

Storage estimation scores through method visibility: state record size with a one-line justification (“ids + geo + fare ≈ 1KB”), run the multiplication aloud including all three multipliers, then use the result (“165TB replicated means trips archive to object storage after 12–18 months”). The follow-up almost always tests retention — have the tiered answer ready.