Fundamentals of data lake architecture — storage layout, partitioning, format selection, and lifecycle management.
37 articles
How STACKIT built a managed lakehouse offering on Kubernetes — custom operators, CRD-driven provisioning, multi-tenant Iceberg catalogs, and the engineering lessons from bringing a sovereign lakehouse service to market in Europe.
AI agents fail in production because they are overwhelmed with data but starved for context. The bottleneck is not the model — it is the data stack. Apache Iceberg turns lakehouse storage into a live, versioned context layer that powers structured RAG, schema-aware agents, and governed reasoning grounded in truth.
How to migrate Delta Lake tables to Apache Iceberg without copying or rewriting data files. Covers zero-copy metadata conversion, the mapping between Delta transaction logs and Iceberg manifest trees, Iceberg V3 spec compatibility, practical tooling (XTable, UniForm, Iceberg Delta module), and the post-migration operational discipline — compaction, sort optimization, statistics — that determines whether converted tables actually perform.
Iceberg deliberately avoids embedding governance into its table format — access control, classification, and policy enforcement belong in the catalog and policy engine layers. This article lays out the three-layer model: table format for data portability, catalog control plane for enforcement, and pluggable policy engines for rules. How Polaris, OPA, and Ranger fit together in production multi-engine lakehouses.
Apache Iceberg does not ship with observability — data platform teams need table health, engine metrics, cost attribution, and lineage across Spark, Trino, and Flink. The seven pillars of lakehouse monitoring and how a control plane makes it operational.
LinkedIn runs Iceberg CDC on 10,000+ tables — billions of upserts daily, 2+ PB throughput. Equality vs position deletes, delete-file compaction, budgeted maintenance, and WAP branching lessons for any team running MERGE INTO at scale.
A comprehensive migration guide covering three source patterns — Hive/HMS tables, raw Parquet on S3, and Delta Lake — with three migration approaches (in-place, CTAS, shadow), post-migration operations, validation checklists, and common pitfalls. Includes production SQL, config examples, and the operational discipline teams underestimate after conversion.
Data quality and table health are different failure modes — one breaks business trust, the other breaks performance silently. A practical guide to the metrics, monitoring queries, classification frameworks, and automated remediation that keep Iceberg tables reliable in production.
A practical guide for migrating Apache Iceberg catalogs — from Hive Metastore to REST (Polaris, Gravitino), AWS Glue, Nessie, or Unity Catalog. Covers in-place metadata registration, dual-catalog access, validation, rollback strategies, and multi-catalog federation with LakeOps.
Orphan files are invisible to Iceberg but fully billable by cloud storage. They accumulate silently from failed writes, crashed compaction, and concurrent conflicts — and on mature lakes they can account for 25–40% of storage spend. This guide covers how orphan files are created, how to detect them safely, the retention window that prevents table corruption, and how to automate cleanup at lake scale without listing millions of objects.
Every Iceberg commit creates a snapshot. Left unmanaged, snapshots pin storage, inflate metadata, and slow query planning. This guide covers what snapshots actually cost, retention strategies by workload type, the retain_last vs older_than tradeoffs, compliance via tags and branches, the cascade from snapshots to orphan files to manifest bloat, and how to set automated retention policies that match your operational reality.
Every concurrent write to an Apache Iceberg table risks a commit conflict. This guide covers how Iceberg's optimistic concurrency works, what triggers CommitFailedException, the common conflict scenarios in streaming and maintenance workloads, and the strategies — from partition isolation to branch-based writes — that eliminate conflicts in production.
A production-ready runbook for Iceberg incidents: queries suddenly slow, planning takes minutes, write conflicts spike, storage grows uncontrolled, compaction OOMs, time travel breaks, and delete files degrade reads. Each incident follows Symptom → Root Cause → Diagnosis → Fix → Prevention.
Apache Iceberg ships the maintenance primitives — compaction, snapshot expiration, orphan cleanup, and manifest rewriting — but none of them run themselves. This guide covers why each operation matters, the correct execution order, the limitations of scripts and cron jobs, and how to automate the full lifecycle with policies, observability, and a purpose-built control plane.
AWS Glue provides native Iceberg support for cataloging, ETL, and built-in table maintenance — but production lakehouses hit limitations fast. This guide covers Glue catalog configuration, ETL best practices, compaction tuning, common pitfalls, and how a dedicated control plane fills the operational gaps.
Flink streaming into Iceberg creates thousands of small files per hour. This guide covers checkpoint tuning, write distribution modes, Flink SQL patterns, and why external maintenance is essential for production streaming tables.
Delete files let Iceberg avoid rewriting data on every UPDATE or DELETE — but every unresolved delete file forces readers to reconcile at query time. A deep guide to position deletes, equality deletes, measuring overhead, and resolving accumulation before it tanks performance.
Partitioning determines how much data every query must scan. Apache Iceberg's hidden partitioning and partition evolution change the game — but choosing the wrong strategy still creates performance cliffs. A practical guide to transforms, sizing, evolution, and avoiding the small-files trap.
Iceberg tables degrade silently in production — small files multiply, snapshots accumulate, orphans waste storage, and manifests fragment. A comprehensive guide to the five maintenance operations, why sequencing matters, the metrics that reveal problems early, and how to automate the full lifecycle.
Apache Iceberg on AWS S3 is the standard architecture for open lakehouses. This guide covers how Iceberg's metadata hierarchy maps to S3 objects, the AWS services ecosystem (Glue, Athena, EMR, Redshift, S3 Tables), configuration best practices, performance optimization, table maintenance, and the operational components needed for production deployments.
AWS S3 bills for Iceberg lakehouses are inflated by small files, orphan data, retained snapshots, metadata overhead, and scan amplification. This guide quantifies each cost vector with S3 pricing mechanics and walks through five strategies — compaction, expiration, layout optimization, storage tiering, and engine routing — to cut storage and query spend.
How to route queries across Trino, Spark, DuckDB, Snowflake, Athena, and Flink on shared Iceberg tables — covering the architecture of a SQL routing proxy, dialect translation, routing strategies, table-aware optimization, and the tooling that makes it work.
A technical comparison of the seven major Apache Iceberg catalogs — Hive Metastore, AWS Glue, Apache Polaris, Project Nessie, Databricks Unity Catalog, Apache Gravitino, and Lakekeeper — across protocol support, access control, multi-engine interoperability, credential vending, and production readiness.
Eight optimization layers for data platform engineers running BI, ad-hoc SQL, and aggregation pipelines on Apache Iceberg — from partition design and file sizing through compaction, routing, and continuous maintenance.
A deep technical guide to managing the metadata layer that makes Apache Iceberg fast — snapshots, manifests, metadata.json files, and Puffin statistics — covering expiration, consolidation, orphan cleanup, and the sequencing that prevents production incidents.
Iceberg tables degrade silently — small files from streaming, unsorted data, fragmented manifests, accumulated delete files. Each one caps query speed regardless of engine. Six concrete optimization layers, how they interact, and how autonomous maintenance keeps every table at peak performance.
Data lakes, warehouses, and lakehouses are not interchangeable — each has hard limits the others cannot cover. A practical guide for platform leaders: where each architecture wins, where it fails, cost and governance trade-offs, and how to choose (or combine) them in 2026.
Compaction keeps Apache Iceberg lakehouses fast and lean — but every tool approaches it differently. A side-by-side look at nine production options: LakeOps, AWS Glue, Amazon S3 Tables, Snowflake, Google BigLake, Cloudera, Starburst, Dremio, and Databricks.
Every data lake starts with a promise of unlimited flexibility — and most end up as a swamp. Stale files, broken schemas, no observability, and engineers spending more time maintaining pipelines than analyzing data. Apache Iceberg fixed the reliability gap. A lakehouse control plane fixes everything else. A practical guide to the full transition — component by component.
Compaction is the most impactful operation in an Apache Iceberg lakehouse — and the hardest to get right at scale. File merging is the easy part. Knowing when to trigger it, what sort strategy to apply per table, how to avoid conflicting with other maintenance, and how to do it without spinning up expensive JVM clusters — that is the real problem. A breakdown of what modern compaction actually requires.
AI agents issue SQL iteratively, repeat query templates at high frequency, and need sub-second responses from tables designed for batch workloads. This post covers what breaks when agents hit a production Iceberg lake — and the five infrastructure layers that fix it: MCP connectivity, guardrails, multi-engine routing, self-optimizing storage, and closed-loop feedback.
ExternalHow we compacted 4.5 TB across 10 real production tables, achieved up to 99.8% file reduction, and made Apache Spark OOM on a job we finished in 11 minutes.
ExternalHow we built a high-performance, distributed compaction engine for Apache Iceberg using Rust and DataFusion—architecture, design choices, and lessons learned.
See how a 350TB data lake shrank to 230TB in 10 minutes by removing stale data—saving 34% in AWS S3 costs and proving the need for a control plane.
Apache Iceberg delivers speed, but without a control plane snapshots pile up, costs surge, query take more time — starting with expiration.
ExternalA detailed comparison between Delta Lake and Apache Iceberg, exploring their architectures, performance characteristics, and ideal use cases to help you make the right choice.
ExternalLearn how to implement efficient incremental processing with Apache Iceberg and Spark, including best practices for data lake optimization and performance tuning.
