Apache Iceberg is a table format used in data lake architectures. A useful interview explanation separates the table format, the catalog that helps locate and manage tables, and the engine that executes queries. A shared format does not make permissions, cost or feature support identical across engines.

The current industry signal

Google’s July 2026 lakehouse announcement describes Iceberg-based catalog federation, with the cited catalog integrations in preview. Databricks’ September release notes report general availability for sharing foreign Iceberg tables. These are reasons to study interoperability, not evidence of a search-volume ranking. Google announcement; Databricks release notes.

Explain the architecture

Picture an orders table stored as data files with table metadata. A catalog helps a compatible engine discover the table. The engine plans and runs a query against an appropriate table state. Authentication and authorisation still determine who can discover and read it. A dashboard adds its own model, filters and refresh behaviour.

Ask who owns writes, which engine versions are supported, how credentials are issued and how schema changes are coordinated. Two engines reading the same files can return different business totals when their queries use different refund or timezone rules.

Work through a cost decision

Use these invented planning inputs, all in USD. Remote querying costs $0.40 per report run in combined variable charges. A local analytical copy costs $18 per day plus $0.10 per run. These are not vendor prices.

runs = 80
remote_usd = 0.40 * runs
local_usd = 18 + 0.10 * runs
print(round(remote_usd, 2), round(local_usd, 2))
print(round(18 / (0.40 - 0.10)))

At 80 runs, remote querying costs $32 and the local option costs $26 per day. The simplified break-even point is 60 runs, when both cost $24. At 20 runs, the figures are $8 and $20. Workload changes can therefore change the preferred option.

Add the missing tradeoffs

The estimate excludes engineering time, storage growth, commitments and failures. Add these before making a purchase decision. A local copy also introduces freshness delay and another pipeline. Remote reads can introduce latency and availability dependencies. Measure representative query latency, scanned bytes, update frequency and concurrency.

Ask what “zero copy” means for the specific feature. Avoiding a persistent ingestion copy does not establish that a service never uses caches, temporary results or materialisation. Check the current billing rules rather than interpreting the phrase as zero cost.

Practise an interview answer

“I would define freshness and response-time requirements, test compatible readers and writers, verify access boundaries and compare total costs under representative workloads.” Explain how you would reconcile row counts and totals between engines at the same cutoff. Include the evidence that would make you change your recommendation.

Next: data engineering and report reconciliation.

← All articles · Suggest a correction