When the network loses to a truck
Rule of thumb (often attributed to AWS):
Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway.
100 TB over 1 Gbps ≈ 9+ days theoretical — real world slower. Snowball ships appliance, loads at datacenter line rate, AWS ingests internally.
Cloud physical transfer
| Service | Scale | Pattern |
|---|---|---|
| AWS Snowcone / Snowball / Snowmobile | TB–PB | Encrypted appliance, chain of custody |
| Azure Data Box | TB | Similar appliance model |
| GCP Transfer Appliance | TB | Rackable storage device |
Use when: limited bandwidth, one-time migration, air-gapped environments, compliance constraints on public internet transfer.
Online large transfer (when physical is overkill)
- AWS DataSync / Transfer Family — managed agents, scheduled sync
- gsutil -m / azcopy — parallel object copy
- Cross-region replication — ongoing sync after initial seed
Initial seed often still physical; replication handles delta.
Bandwidth math (staff quick estimate)
time_seconds ≈ bytes × 8 / bits_per_second
100 TB = 8×10^14 bits
10 Gbps → ~22 hours ideal; ~2–3× real world
Factor: TCP overhead, encryption, small files, API list operations.
Security for bulk data
- Encryption at rest on appliance (KMS keys)
- Tamper-evident seals
- Logging chain of custody
- VPC endpoints / PrivateLink so data never traverses public internet after cloud ingest
Application API role
Your API does not move the bytes — it:
- Issues credentials / job tickets
- Tracks manifest and checksums
- Triggers pipeline when object lands in
s3://landing/ - Notifies via webhook when import completes
Pairs with Data Engineering ingest patterns (bronze layer landing zone).
Replication vs migration
- Migration — one-time cutover, validate counts/checksums
- Replication — ongoing; conflict resolution, lag monitoring
Edge cases
- Many small files — tar/parquet bundle before transfer; list operations dominate
- Geo-distributed users — CDN for download; regional upload endpoints
- Compliance — data residency; transfer within region only
Senior answer structure: quantify size/time, compare network vs physical, describe orchestration API, mention integrity verification and incremental sync.
Further Reading
Hands-On Tasks (Optional)
API design drills and whiteboard exercises — protocol selection, contract design, and bulk-transfer architecture. Assumes Networking and sibling tracks on the hub page (Distributed Systems, Databases, Concurrency, LLD).
- Network vs truck calculation15m
100TB migration, 10 Gbps dedicated line vs AWS Snowball. Estimate time, cost order-of-magnitude, and decision factors (security, urgency, ops burden).