Drop DSE parts of the driver

[wprzytula]

This PR prunes DSE-related parts of the driver. It's fully AI-authored, so requires a keen and sceptical eye.

Contents

1. docs: add java-rs-driver design assessment and migration plan — background documents only.
2. refactor: move the reactive API out of the com.datastax.dse namespace — the reactive API is not DSE-specific, so it is relocated rather than deleted. No behaviour change.
3. Remove all DataStax Enterprise (DSE) support — graph, continuous paging, geometry types, proxy/DSE auth, DSE config options and version registry entries, and the DSE test/IT suites. HCD is dropped with it: it inherits DataStax's DseAuthenticator, so keeping it would have meant keeping DSE's SASL mechanism negotiation. Both removals are documented in upgrade_guide/README.md.

Review rounds

Round 1 (nine items, all addressed): the build: unbreak the build on JDK 25 commit was dropped rather than gated, so Error Prone stays at 2.31.0 and the [11, 17] matrix and <release>11</release> are untouched; fmt:format was run at every commit of the stack; the GraalVM reactive substitutions were restored with the graph dimension collapsed out; HCD was dropped rather than restoring DSE's SASL negotiation; and the pinned docs.datastax.com javadoc links were reverted to their com/datastax/dse paths, since this fork's rename does not affect DataStax's published 4.17 docs.

Round 2: the authorization-id wording is corrected everywhere it appears (the driver still encodes the id; Cassandra and ScyllaDB ignore it server-side), the upgrade guide gained an HCD section, getMechanism()/MECHANISM and the now-dead CcmBridge.V4_0_0 are removed, and the 42 graph-era tinkerpop/gremlin entries are gone from core/revapi.json.

revapi, worth knowing: the gate has never compared anything in this fork — or upstream. versionFormat is \d+\.\d+\.\d+ while every published com.scylladb:java-driver-core release is four-part, so RELEASE resolves to no baseline and check passes vacuously (failOnUnresolvedArtifacts defaults to false). Filed upstream as scylladb#1054. This PR makes the config correct — com.datastax.dse is excluded on both sides, DSE and graph-era baseline entries are deleted — but does not activate the gate, because fixing the regex surfaces a backlog of unbaselined breakages that needs its own triage.

Verification

CI is green on this PR (22/22): Build, Full verify (= make check = mvn verify -DskipTests, where fmt:check and revapi:check run) and Unit tests on JDK 11 and 17, plus the Cassandra 3/4-LATEST and Scylla LATEST/LTS integration suites.

Locally on JDK 17: mvn verify -DskipTests green across the reactor; mvn test green except 20 Timestamp/LocalTimestamp/ZonedTimestamp/TimestampMillisCodecTest.should_parse cases, which are parameterized over ZoneId.systemDefault() and therefore fail on any non-UTC machine and pass in CI — not a regression and not specific to this branch.

[nikagra]

Reviewed rs..drop-dse (4 commits, 573 files). The DSE removal itself is thorough: no dangling com.datastax.dse code references, the OptionsMap/reference.conf default counts still balance on both sides so MapBasedDriverConfigLoaderTest.should_fill_default_profile_like_reference_file still passes, and cross-checking the equivalent upstream change (scylladb#1042), the inverted DefaultCcmBridgeBuilderCustomizer condition and the renamed serialized Version.dsePatch field are both handled correctly here.

Requesting changes on three build gates. None have ever run: the workflow only triggers on PRs targeting scylla-4.*x, so no rs-based branch has had CI. I verified the formatting one by running the formatter rather than reading it — 49 of the 136 changed files fail, 0 on rs.

Two behaviour regressions are inline as well: the GraalVM reactive substitutions were deleted along with the graph ones, and plaintext auth lost the SASL mechanism negotiation that HCD — which this PR keeps — still needs.

Each comment is a one-line TL;DR with the detail folded underneath. Several of these are Java build machinery with no Cargo equivalent, so the folded part explains the mechanism rather than just naming the plugin.

The description records no verification; one make check run would have caught all three gates before review.

[nikagra]

Re-reviewed at 6e06019e73. All nine round-1 items are addressed, and the three build gates now pass rather than being unexercised. Approving. One Minor and two Nits below, none blocking.

The gates now run. Round 1 said none of them ever had — that was true when written and is obsolete now. 142cb944f on rs retargeted the workflow trigger, so this PR gets CI, and it is green 22/22: Build (11)/(17) (Error Prone 2.31.0 loads on JDK 11), Full verify (11)/(17) (= make check = mvn verify -DskipTests, which is where fmt:check and revapi:check live), plus the full Cassandra 3/4-LATEST and Scylla LATEST/LTS integration suites that round 1 flagged as never having run.

Two corrections of mine.

1. I wrote "google-java-format 1.7 (the version fmt-maven-plugin 2.13 bundles)". Wrong — com.coveo:fmt-maven-plugin:2.13 declares google-java-format 1.13.0. My "49 files" count was produced with the wrong formatter version. The defects I listed (import order, dangling imports, duplicates) are version-independent, were all real, and are all gone: I re-checked and 8/8 dangling imports are removed, with zero order/duplicate/unused violations across the 137 Java files this PR touches.
2. My TypedDriverOptionTest.java:52 item was a false positive, and I also gave the wrong path (it is core/src/test/.../api/core/config/). Collapsed to one line that statement is 112 columns, over the 100-column limit, so the formatter correctly keeps it split. Nothing to fix there.

On the per-commit claim. CI only ever gates the tip, so I checked the middle commit separately: all 73 modified Java files at debe1ce683 are clean on import order, duplicates and unused imports. The "formatted at each commit" claim holds for the part that can be checked statically.

Your revapi question — agreed, and it is the same config upstream. scylla-4.x has the identical <versionFormat>\d+\.\d+\.\d+</versionFormat> (pom.xml:665) against ${project.groupId}:${project.artifactId}:RELEASE (pom.xml:668), and every published com.scylladb:java-driver-core is four-part (4.19.0.1, 4.18.1.0, ...) or -scylla-N — none match. With failOnUnresolvedArtifacts defaulting to false, revapi:check has been passing vacuously there too. Filed as scylladb#1054, credited to you. Flagged there that fixing the regex activates a gate that has never run, so it will surface a backlog of unbaselined breakages to triage.

Two things I did not verify, so as not to imply otherwise: no native-image build was run — the substitution is structurally correct and byte-identical to the working rs paths, but nothing here proves a native image builds without reactive-streams; and no Maven run, so every revapi conclusion is read off the config plus the Maven Central version list.

[Nit] 🟢 The description needs a refresh before merge (it becomes the merge-commit body). Lines 23 and 25 say the workflow only triggers on scylla-4.*x so CI does not run, and that ITs were not run — both are now false. The 20 should_parse failures on line 22 are also not a shared baseline: that test is parameterized over ZoneId.systemDefault(), so it fails on any non-UTC machine and passes in CI. No diff anchor, hence up here.

Also cosmetic: every reply says "New head: a3ba97a", which is one force-push stale — the live head is 6e06019e73. The content is a pure rebase so nothing is wrong, but it makes the threads harder to re-review.

[nikagra]

👍