Replication Multiplier

Replication Is Never Free

Every copy of data multiplies three costs simultaneously:

 REPLICATION FACTOR (RF) = number of copies held

 RF multiplies:
   storage          ×RF
   write bandwidth  ×RF   (every write fans to every copy)
   failure domains consumed ×RF

 RF=1: no redundancy — a node loss is a data-loss event
 RF=2: survives one loss barely; maintenance windows get scary
 RF=3: the industry floor for serious systems
       tolerates 1 loss + 1 under maintenance concurrently

The Arithmetic on RideShare Storage

 raw trips over 5 years:            ≈ 36.5 TB

 RF=3 replication:      ×3   →     ~110 TB stored
 + indexes (~×1.5):             → ~165 TB provisioned-class
 ÷ 0.7 max utilization:         → ~235 TB fleet capacity

 same data with erasure coding (8+4-ish):
   overhead ≈ ×1.5 instead of ×3 → roughly HALF the storage bill

Erasure Coding vs Replication

The trade behind the savings:

 REPLICATION (RF=3)                 ERASURE CODING (k+m)
 store 3 full copies                split object into k shards +
                                    m parity shards (any k rebuild)

 write cost: ×3 bandwidth           write cost: k+m fragments written
 read latency: any replica answers  read: must gather k fragments —
                                      cross-node reads per request!
 rebuild: copy from surviving       rebuild: distributed reconstruction
              replica (cheap)                 across many nodes
 
 wins: latency-sensitive reads      wins: cold/archive storage economics
 
 typical pattern: hot data replicated; cold data EC'd

Write Bandwidth Amplification

Storage is only half the multiplier story:

 application sends 10k writes/sec × 1 KB

 primary absorbs:        10 MB/s logical
 cluster actually moves (RF=3): 30+ MB/s internal write traffic
 quorum W=2: still ≥20 MB/s before acks return

 this amplification lands on the SAME network as your user traffic —
 capacity plans must include it or NICs saturate mysteriously

Quorum Numbers Interact With RF

Consistency settings ride on top of the factor choice:

 N=3 copies with:
   W=1, R=1 : fastest, weakest — eventual consistency territory
   W=2, R=2 : quorum consistency; tolerates 1 node loss
   W=3, R=1 : write-durable but reads may lag
 
 each step toward stronger quorums adds latency = slowest-quorum-member
 → placement (same-rack? cross-AZ?) decides what that latency IS

Interview Framing

The tested reflexes: multiply storage AND write bandwidth by RF when sizing anything stateful (“110TB at RF=3, not 36”), mention erasure coding as the cold-data alternative with its latency caveat, and connect quorum values to both consistency and latency. One sentence pattern covers it all: “RF=3 for the hot path because reads answer locally; EC for archive because bytes dominate there.”