Bandwidth Estimation

The Two Multipliers

Bandwidth = QPS × payload size. Every path has both directions, and they rarely match:

 INGRESS (into system)              EGRESS (out to users)
 client requests, uploads           responses, media, pages
 
 typically small for read apps      typically 5–50x ingress
                                    ← the number that shows up on invoices

Worked Example: RideShare

 RIDER READ PATH (egress-dominant):
   GET /trip/status response ≈ 2 KB × 900 peak rps  ≈ 1.8 MB/s egress
   
 LOCATION INGESTION (ingress-heavy — unusual!):
   ping ≈ 200 B × 100k/sec                          ≈ 20 MB/s ingress
   fan-out to nearby drivers' match views adds internal egress
   
 INTERNAL (server↔server):
   replica reads, cache fills, CDC streams — often exceeds
   user-facing traffic; cross-AZ transfer is PRICED per GB

The surprise pattern: a location service with modest user counts moves more bytes than its web tier — telemetry payloads are tiny but relentless.

Units Discipline

The classic estimation error lives here:

 network speaks BITS:    1 Gbps NIC = 125 MB/s  (÷8!)
 cloud links quoted in Gbps
 payloads measured in BYTES
 
 conversion slips cause order-of-magnitude errors → always divide by 8 first

What the Numbers Get Compared Against

Link	Approx capacity	Implication
Cloud instance NIC (typical)	1–25 Gbps	Per-box ceiling before LB spreads load
Cross-AZ transfer	priced ~$0.01–0.02/GB	Replica placement is a cost decision
Internet egress	~$0.05–0.12/GB	CDN offload pays for itself fast

(Prices illustrative; orders of magnitude are the point.)

Example consequence: 20 MB/s sustained = ~52 TB/month. At internet egress rates that’s real money monthly; served from CDN edge it’s a fraction — the estimate just made an architecture argument.

Media Changes Everything

Text systems measure bandwidth in KB/s; media in TB/s:

 video platform sanity check (illustrative):
   100k concurrent viewers × 3 Mbps stream
   = 300k Mbps = 300 Gbps sustained egress
   → physically impossible from one origin region
   → CDN is not an optimization here; it is the architecture

This is why bandwidth estimation precedes component choice for any media-adjacent design.

Interview Framing

Run one direction of math per traffic class, convert bits/bytes correctly out loud, then land the implication (“that egress volume means CDN-first delivery” or “cross-AZ chatter argues for replica-aware reads”). Bandwidth numbers score when they terminate in a decision — a bare Gbps figure floating in the air scores nothing.