CONCEPT Cited by 2 sources
Egress Cost¶
Egress cost is the per-byte charge a cloud provider levies when data leaves a region, a cloud, or (in some cases) an availability zone. It is one of the strongest shaping forces in cross-cloud / multi-region architecture, because unlike compute and storage, egress is linear in access pattern: every cross-cloud read pays, every time.
Why it reshapes architectures¶
- Live cross-cloud reads are a pay-per-query surface. Even when the remote system is technically capable of serving the query directly, each read is a cross-cloud data movement event. At TB-scale datasets + many consumers, this dominates the other cost lines.
- Replication trades one recurring cost for one bounded cost. Copying the remote dataset into the local cloud pays egress once per sync window instead of once per query. When reads-per-sync >> 1, replication wins even if storage is "paid twice".
- Freshness becomes a price knob. The sync cadence is the dial: short windows → more egress, fresher data; long windows → cheaper, staler. Neither extreme is "right"; the architecture has to expose the choice per use case. This is exactly the knob Mercedes-Benz's mesh exposes to consumers.
- Who pays matters as much as how much. Egress bills that land on the consumer team disincentivise cross-team data sharing; patterns/chargeback-cost-attribution as used at Mercedes-Benz moves the bill to the producer so the data-modelling team bears the cost of "inefficient data products".
The Mercedes-Benz numbers¶
- Live cross-cloud consumption of 60 TB of after-sales data against dozens of Azure use cases was expensive enough to force an architectural redesign.
- Pairing Delta Sharing with incremental Delta-Deep-Clone replication reduced egress cost by 66 % on the first 10 data products; projected ~93 % annual reduction scaled across 50 use cases.
- Compute cost of the replication Sync Jobs themselves was tracked but not quoted as offsetting the savings — i.e. the egress reduction, not compute reduction, was the business case.
(Source: sources/2026-04-20-databricks-mercedes-benz-cross-cloud-data-mesh)
Seen in¶
- sources/2026-04-20-databricks-mercedes-benz-cross-cloud-data-mesh — egress as the forcing function behind the hybrid replication tier.
- sources/2024-08-15-flyio-were-cutting-l40s-prices-in-half — GPU-inference instance. Fly.io frames the hyperscaler cost squeeze on inference workloads as a two-axis tax: GPU-instance surcharges on the compute side, plus egress fees on the object-storage side — "The hyperscaler public clouds strangle these customers, first with GPU instance surcharges, and then with egress fees for object storage data when those customers try to outsource the GPU stuff to GPU providers." The architectural counter-play is GPU + object-storage co-location on one platform. Egress is the shaping force that makes cross-cloud GPU-inference topologies economically unviable at scale.