How Apache Iceberg Lakehouses Works Under the Hood
TL;DR
A complete, up-to-date breakdown of under the hood for developers and founders. It covers the core ideas, the trade-offs that matter, a practical workflow, real numbers, and the questions people ask most — written to be skimmed, applied, and shared.
Key takeaways
- Treat Kafka topics as an append-only log and a source of truth, not just a message queue, because retention and replay are what make event-driven architectures durable.
- Push data quality left with data contracts at the producer boundary, so schema and semantic breakages fail in CI rather than silently corrupting downstream dashboards.
- Choose orchestration by paradigm: Airflow for battle-tested task DAGs, Dagster when you want asset-centric lineage and typed, testable pipelines.
- Use reverse ETL to operationalize the warehouse by syncing modeled data back into Salesforce, HubSpot, and ad platforms instead of building bespoke one-off integrations.
- Adopt data mesh for organizational scaling, not for small teams, because its domain ownership and self-serve platform overhead only pays off past real coordination pain.
This is a practical, up-to-date guide to Under the Hood — what it is, why it matters in 2026, and how to apply it in real projects. It is written for developers and founders who want clear answers and proven best practices, not filler.
Whether you're just starting out or leveling up, treat this as a working reference you can return to. Every section is built to be skimmed, applied, and shared.
Batch versus streaming: how the two paradigms differ
Batch processing collects data over a window and processes it in bulk on a schedule, which is simpler to reason about and cheaper for large historical reprocessing. Stream processing instead handles events one at a time or in small micro-batches as they arrive, trading some simplicity for low latency and continuously fresh results. The practical distinction is latency and boundedness: batch works on a finite dataset that sits still, while streaming works on an unbounded, never-ending flow where you must decide how to window and when results are complete. Modern engines increasingly blur the line, with Apache Flink treating batch as a special case of streaming and Apache Spark offering Structured Streaming on top of its batch engine. Choosing between them comes down to whether the business genuinely needs sub-minute freshness or whether an hourly or daily refresh is good enough, since streaming carries real operational complexity.
Stream processing with Apache Flink
Apache Flink is a stateful stream-processing framework built for high throughput, low latency, and correct handling of time. Its defining strengths are event-time processing with watermarks, which lets it produce correct aggregations even when events arrive out of order, and robust exactly-once state consistency backed by periodic checkpoints to durable storage. Developers work through layered APIs, from the low-level DataStream API up to Flink SQL and the Table API, which make continuous queries feel like familiar SQL over an unbounded table. Flink handles large keyed state efficiently using RocksDB-backed state backends, which is what enables use cases like real-time fraud scoring, sessionization, and streaming joins that must remember prior events. Managed Flink is now available through Confluent, Amazon Managed Service for Apache Flink, and Ververica, lowering the barrier that historically made Flink harder to adopt than Kafka.
Getting started and avoiding common pitfalls
A pragmatic way into data engineering is to master SQL and Python first, then build one end-to-end pipeline that ingests a real source, transforms it with dbt, lands it in a warehouse or lakehouse, and runs on an orchestrator like Airflow or Dagster. Resist the temptation to reach for streaming and a data mesh on day one, because most teams are better served by a reliable batch pipeline with good tests than by a complex real-time system nobody can debug. The most common pitfalls are premature complexity, missing idempotency that makes retries dangerous, no data quality checks so bad data spreads silently, and treating pipelines as one-off scripts rather than versioned, tested software. Favor incremental models over full reloads once volume grows, and adopt observability and contracts before an outage forces the lesson. Above all, optimize for trust: a slightly slower pipeline that is always correct beats a fast one that is quietly wrong.
Apache Kafka and the event streaming backbone
Apache Kafka is a distributed, partitioned, replicated commit log that has become the default backbone for event streaming across the industry. Producers append events to topics, which are split into partitions for parallelism, and consumers read at their own pace while Kafka retains the data for a configurable period, enabling replay. This durable-log design is what separates Kafka from a traditional message queue: consumers do not destroy messages by reading them, so the same stream can feed many independent systems. Around the core broker sit Kafka Connect for source and sink integrations and Kafka Streams for stateful stream processing, and managed offerings from Confluent, Amazon MSK, and Redpanda reduce the operational burden of running it yourself. Notably, recent Kafka releases removed the ZooKeeper dependency in favor of the built-in KRaft consensus protocol, simplifying cluster operations considerably.
Data mesh as an organizational architecture
Data mesh, introduced by Zhamak Dehghani, is a decentralized approach that treats data as a product owned by the domain teams that understand it best, rather than funneling everything through a single central data team. It rests on four principles: domain-oriented ownership, data as a product with clear contracts and SLAs, a self-serve data platform that lets domains publish without deep infrastructure expertise, and federated computational governance that enforces global standards through automation. The motivation is organizational scaling, because a central team becomes a bottleneck as the number of sources and consumers grows past what one group can meaningfully understand. Importantly, data mesh is an operating model rather than a specific technology, so it is often implemented on top of a lakehouse plus contracts and observability tooling. It is best suited to large organizations feeling real coordination pain, and it tends to be overhead rather than benefit for a small team.
What data engineering actually is
Data engineering is the discipline of building and operating the systems that move, store, transform, and serve data reliably at scale. Where a data scientist asks questions of data, a data engineer builds the pipelines, storage layers, and infrastructure that make those questions answerable in the first place. The core responsibilities span ingestion from operational systems and APIs, transformation into clean modeled tables, storage in warehouses or lakehouses, and orchestration that ties it all together on a schedule or in response to events. In practice the job has converged on a common toolkit: SQL and Python as the working languages, dbt for transformation, an orchestrator like Airflow or Dagster, and a cloud warehouse or lakehouse as the destination. The unifying goal is trustworthy, timely data that analysts, machine learning models, and applications can depend on.
Under the Hood: Key Facts and Data
According to recent industry research and the official documentation linked below:
- Apache Iceberg reached broad vendor support by 2025, with Snowflake, Amazon (S3 Tables and Athena), Google BigQuery, Databricks, Dremio, and Confluent all offering native or managed Iceberg integration.
- The open table format landscape consolidated sharply after Databricks acquired Tabular (the company founded by Iceberg's original creators) in 2024, pushing the industry toward Iceberg and Delta Lake interoperability rather than a single winner.
- Apache Kafka is used by a large share of the Fortune 100, and its own project materials have long claimed adoption at more than 80% of that group, making it the de facto backbone for event streaming as of 2025.
Quick-Reference Summary
A map of what this guide covers:
| Topic | What you'll learn |
|---|---|
| Batch versus streaming: how the two paradigms differ | Batch processing collects data over a window and processes it in bulk on a schedule |
| Stream processing with Apache Flink | Apache Flink is a stateful stream-processing framework built for high throughput |
| Getting started and avoiding common pitfalls | A pragmatic way into data engineering is to master SQL and Python first |
| Apache Kafka and the event streaming backbone | Apache Kafka is a distributed, partitioned, replicated commit log that has become the default backbone for event |
| Data mesh as an organizational architecture | Data mesh, introduced by Zhamak Dehghani, is a decentralized approach that treats data as a product owned by the domain |
| What data engineering actually is | Data engineering is the discipline of building and operating the systems that move |
How to Get Started with Under the Hood
A simple path that works:
- Learn the fundamentals of Under the Hood from primary sources, not just tutorials.
- Build one small, real project end to end.
- Get feedback, refactor, and add tests.
- Ship it publicly and document what you learned.
- Repeat with a slightly harder project each time.
Build It with a World-Class Full Stack Developer
Sandeep Kumar Chaudhary is a full stack world-class developer. If you want to turn this into a real, production-ready product, get in touch — message directly on WhatsApp at +9779802348957 for a fast, no-pressure consult.
You can also explore the projects already shipped to thousands of users, or start a conversation here.
Final Thoughts
Treat Kafka topics as an append-only log and a source of truth, not just a message queue, because retention and replay are what make event-driven architectures durable. The developers and teams who win in 2026 pair strong fundamentals with consistent shipping. Start small, stay curious, build in public, and revisit this guide as your skills grow.
Sources and Further Reading
Frequently Asked Questions
What is under the hood?
Apache Flink is a stateful stream-processing framework built for high throughput, low latency, and correct handling of time. Its defining strengths are event-time processing with watermarks, which lets it produce correct aggregations even when events arrive out of order, and robust exactly-once state consistency backed by periodic checkpoints to durable storage. This guide covers under the hood end to end — core concepts, best practices, concrete data, and a step-by-step approach you can apply right away.
Do I need a data mesh?
Probably not unless you are a large organization where a central data team has become a genuine bottleneck across many domains. Data mesh is an operating model built on domain ownership, data as a product, a self-serve platform, and federated governance, and its overhead only pays off at real organizational scale. Small and mid-size teams usually get more value from a well-run centralized lakehouse with good contracts and observability.
Airflow or Dagster: which orchestrator should I choose?
Choose Airflow if you want the most mature ecosystem, the widest set of integrations, and a well-understood task-based DAG model. Choose Dagster if you prefer an asset-centric approach that gives you built-in lineage, data-aware scheduling, and stronger local testing and typing. Both are capable; the decision usually comes down to whether you want the orchestrator to understand your data assets or simply run your tasks.
How is data observability different from data quality testing?
Data quality testing asserts specific expectations you already know to check, such as a column being non-null or a value falling in a set, often via tools like dbt tests or Great Expectations. Data observability is broader and more continuous, monitoring freshness, volume, schema, distribution, and lineage to surface anomalies you did not anticipate. The two are complementary: explicit tests catch known failure modes, while observability catches the unknown ones and speeds up root-cause analysis.
What is the difference between ETL and ELT?
ETL extracts data, transforms it in a separate processing step, and then loads the cleaned result into the destination. ELT instead loads raw data into a powerful modern warehouse or lakehouse first and transforms it in place using SQL, typically with a tool like dbt. ELT has become the dominant pattern because cloud warehouses make in-database transformation cheap and scalable, and it keeps the raw data available for reprocessing.
Sandeep Kumar Chaudhary
Full Stack Software Developer· Nepal's SEO, AEO, GEO & AIO expert and share-market educator. More about me