Back to Devoured - July 27, 2026
DEVOURED
What Zero-Copy Actually Costs

What Zero-Copy Actually Costs

Data Iceberg Lakehouse
The term 'zero-copy' is a dangerously overloaded marketing label that hides four distinct costs: egress, repeated scans, source load, and metadata churn.
What: Data architect Alex Merced breaks down 'zero-copy' into six patterns: federated query, format virtualization, sharing protocols, catalog federation, mirroring, and caching/materialization. Three of these patterns actually create copies, and the 'correct' one depends entirely on read frequency and geographic location.
Why it matters: This exposes the 'no-ETL' industry narrative as a trade-off exercise, clarifying that data movement is often the most economical choice when query volume is high or systems are separated by regional boundaries.
Takeaway: Run the arithmetic for your specific use case: calculate bytes scanned per day × cost of egress. If the result is high, replicate the data locally.
Deep dive
  • Federated Query: Queries live data via remote connections; dangerous for high-volume analytical workloads.
  • Format Virtualization: Translates metadata (e.g., OneLake serving Delta as Iceberg) without rewriting data files.
  • Sharing Protocols: Uses signed references (e.g., Delta Sharing) to access provider files; egress costs land on the provider.
  • Catalog Federation: Metadata moves (table schemas), but actual data reads stay local to the storage.
  • Pushdown: The critical performance metric; check query plans to ensure filters are running on the remote side, not locally.
  • Agent Amplification: Autonomous agents query in loops, turning predictable human dashboards into massive, unpredictable egress bills.
Decoder
  • Egress: The cost charged by cloud providers for data leaving a network region or availability zone.
  • Pushdown: The optimization where the primary processing engine delegates parts of a query (filters, aggregations) to the source system to reduce data transfer.
  • Reflection: A specialized, materialized view (often used by Dremio) designed to accelerate specific query shapes.
Original article

Full article content is not available for inline reading.

Read the original article →

Explore all 65 articles in today's digest

Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi