This release is basically v2.1.0,
with many post-v2.1.0 bugs fixed (in about 150
commits) and a v3 in Go package paths.
The major version increment to v3 puts a new
floor on a stable API for plugin developers
(both Go plugin developers and exec plugin
developers who happen to use Go).
We made a mistake - v2.1.0 should have been v3.0.0. Per the Go modules doc (which have improved a great deal recently), a release that's already tagged v2 or higher should increment the major version when performing their first Go module-based release.
This advice applies to kustomize, since it was already at major version 2 when it began using Go modules to state its own dependencies in v2.1.0.
But the more important reason for v3 is a change
to the kustomize versioning policy, forced by
the introduction of plugins.
Historically, kustomize's versioning policy
didn't involve Go modules and addressed only the
command line tool's behavior and the fields in a
kustomization file. The underlying packages were
an implementation detail, not under semantic
versioning, because they weren't intended for
export (and should have all been under
internal). Thus although the v2.1.0 CLI is
backward compatible with v2.0.3, the underlying
package APIs are not.
With Go modules, the go tool must assume that Go
packages respect semantic versioning, so it can
perform minimal version selection.
With the introduction of alpha plugins, kustomize
sub-packages - in particular loader and
resmap - become part of an API formally exposed
to plugin authors, and so must be semantically
versioned. This allows plugins defined in other
repositories to clarify that they depend on
kustomize v3.0.0, and not see confusing errors
arising from incompatibilities between v2.1.0 and
v2.0.3. Hence, the jump to v3.
Aside - the set of kustomize packages outside
internal is too large, and over time, informed
by package use, this API surface must shrink.
Such shrinkage will trigger a major version
increment.