test: resolve CCM server flavor per cluster instead of from global properties

[nikagra]

Fixes #983.

Stacked on #982. The first three commits (f0ffa5ca14, 1119088802, 5b3588ebe0) belong to that PR — review only the last four here. Will be rebased onto scylla-3.x once #982 merges.

Why this is a separate PR

#982 exists to unblock driver-core's four CCM tests that Surefire never discovered. Reviewing it surfaced that CCMBridge resolves the server flavor from global system properties throughout, so a cluster configuring its own version gets the surrounding run's flavor instead of its own. Three fixes for that were written and live-verified on #982's branch, but they're a refactor of the CCM test harness rather than part of unblocking the tests — so they moved here, verbatim, and #982 went back to being the rename plus the ccm add fix it made necessary.

None of this is reachable from CI today (see #983 for why), so it is footgun removal that makes the harness behave as its API claims, not a live CI failure. Sibling to #800, the same bug class in 4.x.

099a169abf — pass --scylla to ccm create when Scylla mode is explicit

buildCreateCommand only ever received the resolved Cassandra and DSE versions, so its versionConfigured branch emitted -v <cassandraVersion> and never --scylla. For Scylla that Cassandra version is hardcoded to 3.0.8 — what Scylla reports in system.local, not an install target — so withVersion(...) under -Dscylla.version=... had CCM build an Apache Cassandra 3.0.8 cluster while add() passed --scylla and getScyllaVersion() reported Scylla. The requested version was discarded outright.

• buildCreateCommand now takes the whole resolution and emits --scylla -v release:<version>, the same shape the globally configured Scylla version already produces via CASSANDRA_INSTALL_ARGS.
• The version resolution moved out of build() into resolveVersions() returning a ResolvedVersions holder — a straight extraction, no behaviour change, but it makes the flavor decision assertable without starting a cluster.
• jmxAddressOfNode() and the Scylla-only yaml ports (prometheus_port, api_port, native_shard_aware_transport_port) now use the per-instance version.

9f34037f3b — SCYLLA_PRODUCT follows the version each cluster installs

ENVIRONMENT_MAP was built once in the static initializer and, being immutable and shared by every cluster, gave a builder that configures its own version the wrong repository in both directions:

• global OSS + explicit Enterprise version → SCYLLA_PRODUCT unset → Enterprise version installed from the OSS repository.

• global Enterprise + explicit OSS version → SCYLLA_PRODUCT=enterprise leaks in → OSS version looked up in the Enterprise repository.

• The map is split into a flavor-independent BASE_ENVIRONMENT_MAP and a per-cluster environment built by buildEnvironmentMap(ResolvedVersions), sibling to buildCreateCommand. It is threaded through the CCMBridge instance, so every ccm command a bridge issues — create, add, start — sees its own environment, not just the create.

• The 4-digit-year Enterprise check is now the shared isScyllaEnterpriseVersion helper, used by both the static initializer and the per-cluster path.

• withScyllaEnterprise copies into a HashMap rather than using ImmutableMap.Builder: the inherited process environment may already define SCYLLA_PRODUCT, and duplicate keys would make build() throw.

58bf916fc6 — client encryption follows the flavor, SCYLLA_PRODUCT can't be inherited

SCYLLA_PRODUCT inherited from the surrounding shell. The previous commit derived the variable from the resolved version, but BASE_ENVIRONMENT_MAP still copied the process environment verbatim — so an exported SCYLLA_PRODUCT=enterprise reached every cluster configuring its own version, installing an explicitly configured OSS version from the Enterprise repository. It is now stripped there, leaving withScyllaEnterprise as the only way the variable gets set.

Client encryption picked the wrong flavor's TLS yaml. withSSL()/withAuth() chose between Cassandra's JKS keystore/truststore and Scylla's PEM certificate/keyfile/truststore by reading the global scylla.version at builder-configuration time — before the flavor is resolved, and before withVersion() has any effect. Both directions were wrong.

• The two methods now only record the request. buildClientEncryptionOptions(ResolvedVersions) derives the yaml at build time, sibling to buildCreateCommand and buildEnvironmentMap.
• Explicitly configured entries still win over these defaults, preserving the precedence that applied when withCassandraConfiguration() ran after withSSL() — this matters for SSLSessionTicketsTest, the one test that sets a client_encryption_options.* key of its own.
• Builder.equals/hashCode now compare ssl/auth directly. This is required, not cosmetic: they no longer show up in cassandraConfiguration at configuration time, and CCMCache keys cached clusters on the builder, so without it an encrypted cluster could be handed to a test that asked for a plaintext one. hashCode also picks up scylla, which equals already compared.

Unlike the other two, this one is reachable without a withScylla() call: CCMTestsSupport.ccmBuilder calls withVersion() before withSSL(), so @CCMConfig(version = ..., ssl = true) is enough. It still isn't reachable from CI, because the only SSL tests are SSLTestBase subclasses, which use @CCMConfig(ssl = true, createCluster = false) and never configure a version.

Testing

Carried over from #982, where these commits were originally verified:

• CCMBridgeCreateCommandTest (new, unit group, no CCM) pins the explicit-version branches of ccm create — Scylla gets --scylla -v release:<version> and never 3.0.8, Cassandra gets a bare -v, DSE gets --dse — plus the execution environment and the flavor-specific client-encryption yaml. Each case sets the flavor explicitly so it doesn't depend on the surrounding run's system properties. 9/9 pass, with and without -Dscylla.version=2026.1.0 and with or without SCYLLA_PRODUCT exported.
• Falsified each fix rather than just observing green:
  • Making buildEnvironmentMap return ENVIRONMENT_MAP unconditionally fails the Enterprise case under a global OSS run, and the OSS + Cassandra cases under -Dscylla.version=2026.1.0 — one failure per leak direction.
  • Dropping the SCYLLA_PRODUCT strip makes the two doesNotContainKey assertions fail under SCYLLA_PRODUCT=enterprise.
  • Restoring the global read in buildClientEncryptionOptions fails the Scylla case under a Cassandra run and the Cassandra case under -Dscylla.version=2026.1.0.
  • Removing the ssl/auth comparison from equals fails the cache-key test.
• Live-ran the short-group SSL tests against Scylla 2026.1.0 — SSLEncryptionTest + SSLAuthenticatedEncryptionTest, 9/9 pass. These are the tests CI actually exercises through this code, so this confirms the PEM path is unchanged end-to-end. The JKS path is left to the Cassandra IT legs; local CCM can't start Apache Cassandra here.
• Re-ran ZeroTokenNodesTest live (7 real clusters, plaintext): 7/7 — a regression check on the builder-equality change.
• Full driver-core unit group green, so the Builder.equals/hashCode change breaks nothing else.
• CI: 099a169abf and 9f34037f3b each passed all 10 checks including the four IT legs.

82028f1425 — the last two flavor decisions

The two findings raised against the commits above. Both are pre-existing scylla-3.x behavior rather than regressions from them, and neither is reachable from CI.

SCYLLA_PRODUCT could still be inherited for the globally configured version. 9f34037f3b stripped an inherited value from BASE_ENVIRONMENT_MAP, so no cluster configuring its own version could pick one up, but ENVIRONMENT_MAP was copied before that strip and kept whatever the surrounding shell exported. That is what scylla-3.x always did — one env map, no strip at all — but the comment justifying it only covered a branch spec, while the code also applied it to a numeric OSS version, a pure Cassandra run, and no configured version. In each of those the repository is already known from the resolved flavor, so a stale SCYLLA_PRODUCT=enterprise would install an OSS version from the Enterprise repository.

The assembly moved into buildGlobalEnvironmentMap, which strips the inherited value unless the version is a branch spec — the one case isScyllaEnterpriseVersion cannot classify, where exporting the variable is the only way to select the repository. Extracting it is what makes the decision testable at all: it is otherwise reachable only through a static initializer reading the live process environment, which is why the previous global-path test could assert nothing but isSameAs.

@CCMConfig could not declare Scylla. Builder.scylla defaults to GLOBAL_SCYLLA_VERSION_NUMBER != null and withVersion() doesn't touch it, so under a Scylla run an explicit Cassandra version resolves as a Scylla release. Base silently installed C* 3.0.8 and discarded the requested version; since 099a169abf started honouring the flag it fails loudly instead. Neither is right — and a test could not fix itself, because @CCMConfig forwards dse but had no Scylla equivalent, and version()'s javadoc still called itself "the C* or DSE version to use".

Added the scylla attribute, forwarded it from CCMTestsSupport alongside dse and ahead of withVersion, and documented on both withVersion() and @CCMConfig.version() that the flavor flags decide what the version names. The two call sites that configure a version without a flavor now say which they mean — ProtocolVersionRenegotiationTest's 2.1.16 and RecommissionedNodeTest's 2.1.20 are both Cassandra versions.

Testing:

• CCMBridgeCreateCommandTest grows to 13 cases, covering all four SCYLLA_PRODUCT paths against a synthetic inherited environment, so they hold regardless of the surrounding run. 13/13 with and without -Dscylla.version=2026.1.0 and with SCYLLA_PRODUCT=enterprise exported.
• Falsified by making buildGlobalEnvironmentMap always keep the inherited value: exactly the two strip cases fail, the Enterprise and branch-spec cases still pass.
• Full driver-core unit group green — 640 tests, 0 failures — so the @CCMConfig addition breaks nothing.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: d106d455-51c7-466e-9535-f3b2431e79dd

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}