Egress Estimation

What Egress Means and Why It Bills

Egress = bytes leaving a boundary you pay at. Three boundaries matter:

 1. REGION → INTERNET     ~$0.05–0.12/GB (illustrative market range)
    every byte delivered to users from origin region
 
 2. AZ → AZ (same region) ~$0.01–0.02/GB
    replica reads, chatty service hops pay silently
 
 3. REGION → REGION       ~$0.02–0.09/GB
    cross-region replication, global fan-out

 ingress is free almost everywhere — the asymmetry shapes architecture:
 pull data in freely; push it out thoughtfully

The Core Multiplication

 monthly egress = avg payload × requests/sec × seconds/month

 RideShare rider reads (illustrative):
   2 KB responses × 1,400 rps × 2.6M s ≈ 7 TB/month → hundreds of dollars ✓ fine

 media platform:
   3 MB average video segment × 50k streams/s sustained... 
   ≈ 12 PB/month → millions of dollars without CDN offload ✗✗

 same formula, opposite verdicts — payload class decides everything

CDN Offload Math

The standard fix, priced honestly:

 WITHOUT CDN:   100% of bytes at internet egress rates from origin
 WITH CDN:      ~90–95% served from edge (hit ratio);
                origin pays only misses + CDN's cheaper bulk pricing

 break-even is immediate for cacheable static content —
 CDNs aren't an optimization for media; they're the accounting structure

Cacheable-vs-uncacheable split drives everything: profile images cache forever; personalized API responses don’t — so their egress has no CDN escape hatch and needs payload discipline instead.

Payload Discipline for Non-Cacheable Traffic

Lever	Typical saving
gzip/brotli on JSON/text	60–80%
Field trimming (serve what the screen shows)	30–70%
Pagination + cursoring	caps worst-case payloads
Protocol choice (binary vs verbose JSON)	20–50%
Image formats (WebP/AVIF vs JPEG)	30–60%

API design reviews that ask “is this field consumed?” are egress cost control wearing an engineering hat.

The Hidden Internal Bill

User-facing egress gets attention; internal chatter burns silently:

 anti-pattern: microservice A calls B per item in a loop
   10k rps × 100 calls × 1 KB × cross-AZ pricing = real money/month
 fixes: batching, co-location (same-AZ affinity), caching responses,
        event-driven sync instead of polling loops
 
 rule of thumb: if internal bytes exceed user-facing bytes,
 the topology deserves a redesign conversation

Interview Framing

Egress fluency = running one multiplication per traffic class, then landing the consequence (“that volume means CDN-first” / “cross-AZ chatter argues for zone-aware routing”). For media designs, stating the PB-scale number before proposing components demonstrates why the architecture exists. Numbers terminating in decisions — always.