Glue 6.0's 30% cut is version-gated, and the discount arrives with a Spark 4.1 migration bill

The 30% is attached to a version, not to an account. AWS describes it as 30% lower pricing than previous Glue versions ^[1], which means the saving lands on jobs that have actually been moved to 6.0 and on nothing else. Run the arithmetic the other way and it is a compute budget story: at 70% of the old rate, the same monthly spend buys roughly 43% more Glue capacity ^[10]. That is the size of the prize for finishing a migration, and it is the size of the money left on the table for every pipeline that stays where it is.

The bill is legible from AWS's own tooling. Glue 6.0 moves the runtime to Apache Spark 4.1, Python 3.12 and Scala 2.13 ^[12], and AWS ships a Spark Upgrade Agent plus an auto-upgrade option for existing jobs ^[5] alongside a dedicated migration guide for getting Spark jobs onto version 6.0 ^[6]. Vendors do not write migration guides for parameter flips. Note also what the reassurance covers: the "no API changes are required" line is about selecting the version through the existing `--glue-version` parameter on `create-job` and `update-job` ^[9], which is the control plane. It says nothing about the dependencies compiled into your job.

Which matters, because the two AWS posts announcing the same release do not agree on the interpreter. The launch blog says Python 3.12 ^[12]; the What's New entry says Python 3.13 ^[13]. One of those is wrong, and a one-minor-version gap is the difference between a wheel that loads and a wheel that does not ^[16]. The streaming numbers diverge too: the blog promises single-digit millisecond latency for stateless streaming ^[14], the What's New post says sub-second ^[15], and the published upper bounds sit more than 100x apart ^[17]. The feature lists are also asymmetric, with nanosecond-precision timestamps appearing only in the blog and deletion vectors and DEFAULT column values only in the What's New entry ^[11]. Pin your version matrix off the runtime, not off either page.

On the build-versus-buy side, the interesting item is not the price. It is VARIANT with shredding, which AWS says lets you store and query JSON, logs and event data without flattening schemas, removing duplicate copies and custom parsing code ^[2]. If that holds under load, the thing you delete is an entire stage of hand-written pipeline plus the breakage that follows an upstream schema change, and that is the part of a self-managed lakehouse that costs staff time rather than instance hours. Geometry and geography types and nanosecond timestamps remove two more categories of workaround ^[3].

Treat the superlative with more care. AWS claims the most complete Iceberg v3 implementation on any fully serverless managed Spark service ^[18], and the qualifier does the work: the comparison set excludes anything you run yourself. The spec support is upstream, in Iceberg 1.11.0, which Glue 6.0 is built on ^[2]. What AWS is selling here is the packaging and the 30%, and the packaging is the part you cannot get by upgrading a library.