TL;DR: Map the flows that pay before touching anything, because cross-region replication, chatty cross-region service calls, user-facing egress, backups and log shipping each have a different lever. Price each path using its provider, service, region pair and billing direction; there is no fixed ratio between cross-region and internet rates. Moving computation closer to data can reduce both transfer volume and latency.
How to approach it
Ask for the transfer line grouped by flow rather than by service, because the invoice aggregates it into one forgettable number. Then walk the taxed paths in descending order of typical damage and attach a lever to each.
A strong answer
Six flows account for nearly all of it. Database replication: every primary write ships to the replica region, so write-heavy systems pay twice for durability they may rarely use. Chatty service calls: synchronous fan-out across regions multiplies brutally. User-facing egress to the internet, with rates that depend on service, destination, volume tier and discounts. Backups and disaster-recovery copies. Observability pipelines shipping raw logs cross-region for centralisation. Object transfers somebody scripted once and forgot.
The chatty-call arithmetic is worth doing aloud. Suppose each user request causes 50 cross-region calls, each carrying 4,000 billable bytes, at a steady 100 requests per second. That is 50 × 4,000 × 100 × 86,400 = 1,728,000,000,000 bytes per day: 1,728 decimal GB. At an illustrative $0.02 per GB, the cost is $34.56 per day, or $1,036.80 for 30 days. This example counts one billed direction and excludes response payloads, retries and protocol overhead. Apply the actual billing unit and rates to each direction; AWS EC2 pricing separates transfer paths and volume tiers. Co-location removes this cross-region leg, though cross-zone or other network charges may remain.
Levers ranked by impact:
| Flow | Lever | Effect |
|---|---|---|
| Chatty cross-region calls | Co-locate coupled services | Removes the tax rather than discounting it |
| Replication | Async, delta-based, compressed sync | Often 5 to 10 times less volume |
| Internet egress | CDN and edge caching | Cheaper per GB and origin does less work |
| Logs and traces | Aggregate and sample locally | Ship summaries instead of raw firehoses |
| Predictable bulk flows | Private interconnect | Lower per-GB above a break-even point |
Two nuances separate operators from tourists. Compression and deltas apply far beyond databases: text-heavy log pipelines compress five to ten times, and block-level delta replication turns full-copy thinking into trickle traffic. And CDN offload is not free egress, it is cheaper egress with a cache-hit-ratio dependency; a poor hit ratio can leave you paying both sides, so watch origin fetches alongside the CDN bill.
The honest counterweight: some of this spend is the product. Data residency, disaster recovery and latency floors for users far from the primary region are reasons, not accidents. The goal is deliberate spend with an owner per flow, not zero. Collapse regions below resilience requirements and the next outage reprices this entire exercise.
What interviewers probe next
"Does cross-zone traffic inside one region matter?" Yes. Check the exact service and path, including whether one or both endpoints incur charges. Compare zone-aware routing savings with the resilience and load-distribution requirements; do not collapse replicas into one zone to save transfer fees.
"When does private interconnect pay?" Model fixed port cost against per-GB savings; it wins at sustained multi-terabyte monthly volumes and adds contract lead time.
"How do you stop it creeping back?" Per-flow owners with unit metrics such as bytes transferred per million requests, alerted on ratio since absolute growth can be legitimate.
Common mistakes
Optimising compute while the transfer line grows untouched, because nobody grouped the bill that way.
Shipping raw debug logs cross-region to satisfy a centralisation mandate invented before the cost existed.
Treating egress as one undifferentiated line with no owner, which guarantees its return within two quarters.