AWS just made its most-used data integration service both cheaper and more capable in the same release. On August 21, 2026, Amazon Web Services shipped AWS Glue 6.0, cutting the per-DPU-hour compute rate by 30 percent and adding full support for Apache Iceberg v3. For teams running recurring extract-transform-load pipelines at scale, that combination changes the math on what a data lake actually costs to operate.
The timing matters. Iceberg v3 was ratified in June 2025, and the major warehouse and lakehouse vendors spent the first half of 2026 racing to support it. AWS Glue 6.0 arrives as the most complete implementation on a fully serverless managed Spark service, according to AWS’s own announcement, landing weeks after Snowflake and Databricks rolled out their own v3 support. This piece breaks down what actually changed, what it costs, and where it leaves engineering teams choosing between AWS, Databricks, Snowflake, and Google Cloud for their next lakehouse build.
AWS Glue 6.0 Lands With a 30% Price Cut
The headline number is simple. Glue 6.0 costs $0.308 per DPU-hour for Spark and Spark Streaming jobs, down from the $0.44 per DPU-hour rate that applied to Glue 5.1 and earlier versions in US East (N. Virginia). That’s a flat 30 percent reduction, and AWS says it requires no configuration changes. Existing job scripts point at the new Glue version and the lower rate applies automatically, billed per second with the usual one-minute minimum.
AWS frames the cut as compounding with performance gains. Because Glue 6.0 also runs faster on the same workloads, thanks to the underlying Spark 4.1 runtime, a job that used to take 100 DPU-hours might now finish in 70 to 80 DPU-hours at the new lower rate. AWS cites up to 36 percent better price-performance overall once both effects are combined, though actual gains vary by workload shape and how much of the job is bottlenecked on I/O versus compute.
What Actually Changed Under the Hood
Spark 4.1, Python 3.13, and Scala 2.13
Glue 6.0 rebuilds the runtime stack from the ground up. It moves to Apache Spark 4.1, Python 3.13, and Scala 2.13, replacing the older Spark 3.x-based stack that shipped with Glue 5.1. That’s a meaningful jump for teams that have been stuck on Python 3.10 or Spark 3.3 for compatibility reasons. Newer language runtimes bring their own performance improvements independent of anything AWS added on top, and they close the gap between what Glue supports and what teams already run on self-managed EMR or Databricks clusters.
Arrow-Native Python UDFs
One of the more practical additions is Arrow-native execution for Python user-defined functions and user-defined table functions. Historically, PySpark UDFs paid a serialization tax every time data crossed the boundary between the JVM and the Python process. Arrow-native execution removes much of that overhead, which matters most for teams running complex row-level transformations in Python rather than pushing everything into native Spark SQL. Combined with the pricing cut, this is where the real savings show up for Python-heavy pipelines.
Apache Iceberg v3 Support, Explained
The second half of the release is Apache Iceberg v3 support, built on Iceberg 1.11.0. AWS describes this as the most complete Iceberg v3 implementation available on any fully serverless managed Spark offering. The Iceberg v3 spec, ratified in June 2025, adds seven capabilities on top of v2: deletion vectors for faster row-level deletes, row lineage for native change data capture, the VARIANT type for semi-structured data, default column values, geometry and geography types, nanosecond-precision timestamps, and multi-argument partition transforms.
Deletion vectors alone are a meaningful upgrade for teams running frequent updates or GDPR-style deletes against Iceberg tables. Instead of rewriting entire data files to remove a handful of rows, deletion vectors mark rows as deleted in a separate lightweight file, which community benchmarks put at up to 10 times faster for DML-heavy workloads compared to the copy-on-write approach that v2 tables typically use.
The VARIANT Type and Shredding in Plain Terms
The feature getting the most attention from data engineers is the VARIANT data type. Before Iceberg v3, semi-structured data like JSON payloads had to live in a plain string column, which meant every query touching that field had to parse the whole blob at read time. VARIANT stores that data in a high-performance binary encoding instead, and it supports a technique called shredding.
Shredding pulls specific fields out of a VARIANT column into their own explicitly typed Parquet columns, while keeping a residual binary copy for anything not shredded. Those extracted columns get normal column-level statistics, meaning query engines can use min, max, and null-count metadata to skip files that can’t match a filter. A query looking for one field inside a nested JSON object no longer has to deserialize the entire payload just to check that field’s value.
There’s a trade-off worth knowing about before turning shredding on everywhere. Independent benchmarks using GitHub event data found shredding added roughly 35 percent to write time across repeated append operations. That’s a real cost, and it means shredding makes sense for fields that get queried often and written less frequently, not for every VARIANT column by default.
How the New DPU-Hour Pricing Works
The table below lays out the practical difference between Glue 5.1 and Glue 6.0 pricing and capability, using the US East (N. Virginia) rate as the baseline.
| Attribute | Glue 5.1 | Glue 6.0 |
|---|---|---|
| Price per DPU-hour (Spark/Spark Streaming) | $0.44 | $0.308 |
| Spark version | 3.x | 4.1 |
| Python version | 3.10 | 3.13 |
| Scala version | 2.12 | 2.13 |
| Apache Iceberg support | v2 only | Full v2 and v3 |
| Python UDF execution | Standard JVM bridge | Arrow-native |
| Streaming latency mode | Micro-batch | Real-time, single-digit ms |
| Billing granularity | Per second, 1-min minimum | Per second, 1-min minimum |
Note that AWS also offers Flex execution, which drops the rate further for batch jobs that can tolerate a delayed start. Flex pricing existed before Glue 6.0 and stacks with the new base rate, so teams running non-urgent nightly batch jobs can combine the version 6.0 discount with Flex scheduling for additional savings.
A Worked Example: What 30% Actually Saves
Percentages are easy to state and hard to feel. Here’s what the Glue 6.0 discount looks like in dollar terms across a few common job sizes, assuming the same DPU-hour count on both versions and ignoring any additional speedup from the Spark 4.1 runtime.
| Monthly DPU-hours | Glue 5.1 cost | Glue 6.0 cost | Monthly savings |
|---|---|---|---|
| 500 | $220 | $154 | $66 |
| 2,000 | $880 | $616 | $264 |
| 10,000 | $4,400 | $3,080 | $1,320 |
| 50,000 | $22,000 | $15,400 | $6,600 |
| 200,000 | $88,000 | $61,600 | $26,400 |
At the low end, the savings barely register. At the scale a mid-size data platform team actually runs, the difference adds up to a real line item. And because AWS says the Spark 4.1 runtime also finishes jobs faster in many cases, the effective DPU-hour count for a given workload often drops too, which means these figures likely understate the real savings for teams that migrate their heaviest jobs first.
Real-Time Streaming Gets Single-Digit Millisecond Latency
Glue 6.0 introduces a real-time streaming mode built on Spark 4.1’s Real-Time Mode, layered with Glue-specific execution optimizations. AWS says it achieves single-digit millisecond latency, aimed at use cases like fraud scoring, time-sensitive routing, and low-latency transformation pipelines that previously had to run outside Glue on something like Kinesis Data Analytics or a self-managed Flink cluster.
This closes a gap that’s existed since Glue’s streaming ETL jobs first shipped. Micro-batch streaming, which is what earlier Glue versions relied on, trades latency for throughput efficiency and works fine for dashboards refreshing every few seconds. It never worked for genuinely real-time decisioning. Real-Time Mode changes that calculus, at least for teams willing to stay inside the Glue and Spark ecosystem rather than reaching for a dedicated stream processor.
Migrating From Glue 5.1: What Changes
Version upgrades on a managed Spark service are rarely free. The jump from Spark 3.x to Spark 4.1 and from Python 3.10 to 3.13 can break UDFs, dependency pins, and API calls that quietly changed behavior between major versions. AWS’s own migration documentation flags these as the areas most likely to need script changes before a job runs cleanly on Glue 6.0.
The AI-Powered Spark Upgrade Agent
To soften that pain, AWS paired the Glue 6.0 release with a generative AI upgrade agent for Spark jobs. The tool scans existing PySpark and Scala jobs, flags API and behavioral incompatibilities automatically, and iterates on fixes before validating the result against data quality checks. AWS positions this as cutting migration projects that used to take months down to a matter of days, available through AWS Glue Studio or via API for teams that want to script the upgrade across dozens of jobs at once.
Amazon Q Developer is also wired into the Glue console for ad hoc migration questions, which is a smaller addition but useful for teams without deep Spark internals expertise in-house. Whether the upgrade agent handles complex custom UDFs as cleanly as AWS’s marketing implies is something only wide production use over the next few months will confirm.
The Competitive Field: Databricks, Snowflake, and BigQuery
Glue 6.0 doesn’t land in a vacuum. Every major warehouse and lakehouse vendor spent 2026 building toward Iceberg v3 support, and the competitive picture now looks like this.
| Platform | Iceberg v3 status | Pricing model | Notable strength |
|---|---|---|---|
| Glue 6.0 | Full support, GA Aug 2026 | $0.308/DPU-hour | Serverless, no cluster management |
| Snowflake | GA since May 7, 2026 | Credit-based compute | Deep SQL tooling, governance |
| Databricks | GA on Runtime 18.0+ | DBU-based compute | Unified ML and data engineering |
| Google BigQuery | Native Iceberg table support | Slot or on-demand query pricing | Cross-cloud lakehouse querying |
The practical takeaway is that Iceberg has become the neutral ground everyone agrees to build on. A table written by Glue 6.0 can, in principle, be queried by Snowflake, Databricks, or BigQuery without a data migration, since they all speak the same open table format. That interoperability is arguably a bigger structural shift than any single vendor’s price cut, because it weakens the lock-in argument that used to favor picking one platform and staying there for the life of a data warehouse.
Historical Context: AWS Glue’s Pricing Story
AWS Glue launched in 2017 as a serverless ETL service built around Apache Spark, and its pricing has drifted downward in stages rather than all at once. The DPU-hour model has stayed consistent since launch, but the rate per DPU-hour and the minimum billing granularity have both moved in the customer’s favor over the years, and Flex execution was added later as a second, cheaper tier for delay-tolerant jobs. The 30 percent cut in Glue 6.0 is the largest single-version price reduction the service has taken, at least based on publicly documented rate changes.
It also lands amid a broader AWS pattern this year of using price as a competitive lever against Databricks and Snowflake, both of which have faced pressure over usage-based billing that many customers find hard to predict month to month. A flat, transparent DPU-hour rate is easier to forecast than credit or DBU consumption that varies with query complexity, and AWS appears to be leaning into that predictability as a selling point.
Market Impact: Why This Matters for the Data Lake Economy
Data integration and ETL spend is one of the largest recurring line items in a typical cloud data platform budget, often larger than the storage costs it feeds. A 30 percent cut to a service that runs thousands of recurring jobs a day across AWS’s customer base is not a rounding error. For FinOps teams already under pressure to show measurable savings this year, Glue 6.0 is close to a free win: the discount applies automatically on version upgrade, with no re-architecture required to see it show up on next month’s bill.
The Iceberg v3 support carries a longer-term impact. Deletion vectors and row lineage make Iceberg tables viable for use cases that used to require a dedicated OLTP-adjacent system, like tracking row-level change history for compliance audits. That pulls more workloads toward the open lakehouse model and away from proprietary warehouse formats, which in turn increases the leverage customers have when negotiating with any single vendor, AWS included.
What AWS Is Saying
AWS’s own announcement puts the release in blunt terms. In the AWS News Blog, AWS states: “Today, we are announcing the general availability of AWS Glue 6.0, delivering 30% lower pricing than previous AWS Glue versions and introducing full support for Apache Iceberg v3 features.”
The official AWS What’s New post frames the scope of the release beyond just Iceberg: “AWS Glue 6.0 is now generally available, delivering a 30% price reduction and introducing full support for Apache Iceberg v3, newer versions of Apache Hudi and Delta Lake, and new capabilities to improve developer productivity.”
On the mechanics of how the savings actually reach customers, the AWS Big Data Blog is direct: “This means every job you run on AWS Glue 6.0 costs 30% less per DPU-hour compared to AWS Glue 5.1, with no changes required to your workload configuration.” AWS’s own documentation is equally plain about format coverage, noting simply that “AWS Glue 6.0 supports Iceberg table format versions 2 and 3,” per the AWS Glue documentation.
Risks and Trade-offs Engineers Should Watch
None of this is risk-free. Runtime upgrades of this size, moving three major language and framework versions at once, tend to surface edge cases that only show up under production load. Teams with heavily customized UDFs, third-party JAR dependencies, or scripts that lean on Spark internals rather than public APIs should test on Glue 6.0 in a staging environment before flipping production jobs over, even with the AI upgrade agent doing the heavy lifting.
The shredding trade-off deserves its own line of caution. Turning on VARIANT shredding indiscriminately across every semi-structured column will slow down writes without necessarily speeding up the queries that matter. Treat it as an optimization to apply selectively, based on which fields actually get filtered on in production queries, not a default to enable everywhere.
There’s also a subtler cost risk. Faster jobs on a cheaper rate sound like pure upside, but teams that don’t right-size their DPU allocation after upgrading can end up over-provisioning compute that finishes early and sits idle within its billed minute. Reviewing DPU counts after migration, not just accepting the old sizing, is where a chunk of the theoretical 36 percent price-performance gain either gets captured or left on the table.
What Comes Next: Predictions Through 2027
- Expect Databricks and Snowflake to respond with their own pricing adjustments or bundled discounts within the next two quarters, following the pattern where one hyperscaler’s cost cut forces competitive moves from the others.
- Iceberg v3 adoption will likely become the default for new lakehouse builds by early 2027, with v2 tables increasingly treated as legacy the way Parquet-on-S3-without-a-table-format is today.
- AWS will probably extend the AI-powered upgrade agent beyond Spark version migrations into broader ETL script optimization, given how central generative AI assistance has become across the Glue product line this year.
- Real-Time Mode adoption will stay limited to a narrow set of latency-sensitive use cases at first, since most existing Glue streaming jobs don’t need single-digit millisecond latency and won’t be worth re-architecting purely for that gain.
- Expect further DPU-hour price movement tied to Graviton-based compute options, following AWS’s broader push to shift managed services onto its own silicon for cost efficiency.
Frequently Asked Questions
How much cheaper is Glue 6.0 than previous versions?
Glue 6.0 costs $0.308 per DPU-hour for Spark and Spark Streaming jobs, a 30 percent reduction from the $0.44 per DPU-hour rate on Glue 5.1, applied automatically with no configuration changes.
Does Glue 6.0 support Apache Iceberg v3?
Yes. Glue 6.0 supports both Iceberg table format versions 2 and 3, built on Iceberg 1.11.0, including the VARIANT type, deletion vectors, row lineage, and nanosecond timestamps that the v3 spec introduced.
What is VARIANT shredding and should I enable it?
Shredding extracts specific fields from a VARIANT column into typed, queryable Parquet columns, which speeds up filtered reads. It also adds roughly 35 percent to write time in community benchmarks, so it works best applied selectively to fields queried often rather than every semi-structured column by default.
Will my existing Glue 5.1 jobs break on Glue 6.0?
Possibly, since the runtime moves to Spark 4.1, Python 3.13, and Scala 2.13. AWS offers a generative AI upgrade agent through Glue Studio that scans jobs for incompatibilities and proposes fixes, but scripts with heavy custom dependencies should still be tested in staging first.
How does Glue 6.0 compare to Databricks and Snowflake on Iceberg v3?
Snowflake reached Iceberg v3 general availability on May 7, 2026, and Databricks supports it on Runtime 18.0 and later. AWS says Glue 6.0 offers the most complete v3 implementation among fully serverless managed Spark services, though all three now support the format well enough for cross-platform interoperability.
Does the Glue 6.0 price cut apply to Flex execution jobs too?
Flex execution pricing is a separate, already-discounted tier for delay-tolerant batch jobs. The Glue 6.0 base rate cut and Flex pricing both apply independently, so delay-tolerant jobs on Glue 6.0 with Flex execution get the benefit of both.
What is the new real-time streaming mode in Glue 6.0?
It’s a streaming execution mode built on Spark 4.1’s Real-Time Mode with Glue-specific optimizations, designed to achieve single-digit millisecond latency for use cases like fraud scoring and time-sensitive event routing, as opposed to the micro-batch approach used in earlier Glue streaming jobs.
Where is Glue 6.0 available?
Glue 6.0 is generally available in all AWS Regions where AWS Glue operates, as of the August 21, 2026 announcement.
