AWS Glue 6.0 cuts prices by 30% and ships the full Apache Iceberg v3 spec
On August 21, 2026, AWS launched Glue 6.0 with 30% lower pricing and complete Apache Iceberg v3 support on serverless Spark, including the VARIANT type with shredding. Teams working with semi-structured data now have a dated reason to plan their migration.
August 21, 2026. AWS announced the general availability of Glue 6.0. 30% lower pricing than previous versions. Apache Spark 4.1, Python 3.12, and Scala 2.13 under the hood. And above all, the complete Iceberg v3 specification, built on Iceberg 1.11.0.
This is not a routine release. It makes AWS Glue the most complete serverless Spark service for Iceberg v3, at the exact moment the open table format finishes its standardization. For a data team, the question is no longer “should we adopt Iceberg?” but “when do we switch, and what does version 3 actually get us?”
A price cut and a modernized runtime
The 30% cut is the easiest signal to read. AWS is not nudging a price at the margin: it is repositioning Glue against serverless competitors — Databricks, BigQuery, and homegrown engines — where per-job ETL cost has become a deciding criterion. A straight reduction in unit price lands directly on the bill of recurring pipelines, with no architectural migration required.
The runtime follows. Spark 4.1 replaces the previous generation and brings its native optimizations, while Python 3.12 removes the friction of outdated versions. That matters for a managed service: a recent engine plus a lower price changes the calculation for every workload that had not yet justified moving to the latest generation.
Full Iceberg v3: the VARIANT type is the story
The heart of the announcement is one word: VARIANT. Glue 6.0 implements the VARIANT type with shredding, the flagship Iceberg v3 feature for semi-structured data.
Concretely: instead of storing JSON, logs, or events in a string column that must be parsed on every read, shredding breaks the fields into internal columns at write time. The result is significantly faster reads, with no manual schema flattening, no duplicated data copies, and no pipeline breakage when the schema evolves.
That is precisely the pain point for teams ingesting semi-structured data at scale: the schema changes upstream, the pipeline breaks downstream, and you end up writing bespoke parsing code. VARIANT with shredding removes that entire class of maintenance.
The other Iceberg v3 additions round out the picture. The Geometry and Geography types enable native spatial processing — GIS, location intelligence, geospatial pipelines — directly on managed Spark. Nanosecond-precision timestamps cover IoT sensors, scientific computing, and high-frequency finance, beyond standard millisecond precision. Finally, unknown type handling makes pipelines resilient to unexpected schemas instead of failing on the first upstream change.
An engine that simplifies how pipelines are written
The second half of the release concerns how transformations are written.
Spark Declarative Pipelines changes the authoring model: the engineer declares what the data should become, and the engine determines execution order and optimization itself. That is a direct reduction in orchestration complexity and sequencing errors — the kind of debt that accumulates silently in aging PySpark jobs.
Python UDFs and UDTFs become Arrow-native. Execution moves through the columnar Arrow format without round-trip serialization between Python and the JVM, lifting the historical bottleneck of Python functions in Spark — often the most expensive part of a real pipeline.
Finally, AWS touts real-time streaming with single-digit-millisecond latency, positioning Glue closer to stream-processing workloads previously reserved for dedicated engines.
Verdict
If you already run Glue, migrating to 6.0 is the lowest-risk decision of the year: you recover a 30% price cut and a modernized runtime without changing architecture, provided you validate your UDFs and dependencies on Python 3.12.
If you are mid-way through a lakehouse decision, Iceberg v3 — carried by VARIANT with shredding — is the moment to lock in your format choice. Teams ingesting semi-structured data at volume now have a measurable argument, not just a standards preference.
If your schema changes often upstream, the VARIANT plus unknown type combination targets your exact pain point. Prototype on a real pipeline before making it doctrine: the resilience promise must hold across your range of sources, not on a synthetic benchmark.