feat: add multi-version Flink module structure (1.18/1.19/1.20)

[fightBoxing]

Summary

Convert the project to a multi-module Maven build supporting 3 actively-supported Flink versions: 1.18.1 / 1.19.1 / 1.20.0. All modules share one source tree under root src/, so per-version forks of the source code are not needed.

This is a clean re-slice of #17 — it drops the 4 PR #14 catalog commits (which will land via their own PR) and removes the originally-included lance-flink-1.17 module (rationale below).

Closes #17.

Module Layout

.
├── pom.xml                  ← parent (packaging=pom, dependencyManagement)
├── src/                     ← single shared source tree
└── lance-flink-{1.18,1.19,1.20}/
    ├── pom.xml              ← references root src/ via build-helper-maven-plugin
    └── src/test/resources/log4j2-test.xml

Each submodule:

• Pulls source from root src/ via build-helper-maven-plugin (add-source / add-test-source / add-resource / add-test-resource)
• Overrides flink.version to its target (${flink118.version} / ${flink119.version} / ${flink120.version})
• Inherits all plugin & dependency versions from parent dependencyManagement
• Compiles to Java 8 bytecode

Plugin versions kept aligned with main (#39):

• maven-compiler-plugin 3.15.0
• maven-shade-plugin 3.6.2
• maven-jar-plugin 3.5.0
• maven-source-plugin 3.4.0
• jacoco-maven-plugin 0.8.14

Why no lance-flink-1.17?

Originally this PR included a lance-flink-1.17 module. Local verification on JDK 8 surfaced two unrelated bugs inside Flink 1.17.x's bundled (shaded) Calcite, both of which prevent TableEnvironment creation:

1. MetadataDef.<init>:48 — an assert on a generic-type bound that is always false under -ea (default for surefire), throwing AssertionError.
2. CacheGeneratorUtil$CacheKeyStrategy$1.safeArgList:213 — IllegalArgumentException: fromIndex(2) > toIndex(0), surfaces even with -ea disabled.

Same source tree + same JDK on Flink 1.18 / 1.19 / 1.20 runs cleanly, confirming this is a Flink-side issue, not a connector issue.

Combined with Flink 1.17 reaching end-of-life on 2024-10, supporting it would require carrying workarounds for an upstream-broken planner on a no-longer-maintained branch. We chose to drop 1.17 to keep the matrix lean.

Verification (local, JDK 8 / Tencent Kona 8.0.14)

mvn validate            BUILD SUCCESS  (4 modules)
mvn compile             BUILD SUCCESS
mvn test-compile        BUILD SUCCESS
mvn package -DskipTests BUILD SUCCESS  (3 × 163 MB shaded jar)
mvn test                                                                 (per module)
  lance-flink-1.18:  Tests run: 178, Failures: 1, Errors: 0, Skipped: 17
  lance-flink-1.19:  Tests run: 178, Failures: 1, Errors: 0, Skipped: 17
  lance-flink-1.20:  Tests run: 178, Failures: 1, Errors: 0, Skipped: 17

The single failure is LanceReadOptimizationsTest$FilterPushDownTests.testInPredicatePushDown — a pre-existing main bug that reproduces identically across every Flink version, unrelated to this PR.

Diff Summary

8 files changed, 1148 insertions(+), 312 deletions(-)

• New: lance-flink-{1.18,1.19,1.20}/{pom.xml, src/test/resources/log4j2-test.xml}
• Modified: pom.xml (now parent pom)
• Modified: LanceDynamicTableSource.java (remove unused TableSchema import)

Out of scope (intentionally excluded from #17)

• The 4 namespace-catalog commits from PR feat: add catalog namespace support and refactor adapter implementation #14 (will land via their own PR)
• 1.17 support (see rationale above)

[empathy87]

There are significant changes in how Flink works with sinks starting from version 1.19 (SinkV2). Many classes from 1.18 are marked as deprecated and removed in Flink 2+. Maybe it would be good for the connector to support new Flink 1.19+ API and skip supporting Flink 1.18-.

I already did some implementation of Lance Sink V2 in my repo
empathy87@811db16

[fightBoxing]

Thanks @empathy87 for the pointer and for the prototype work — really nice that you already have a SinkV2 PoC running. 🙏

Two thoughts to keep things clean:

On scope of this PR. #48 is intentionally narrow — it only restructures the build into a parent + per-version submodule layout that share the same src/. It doesn't touch the Sink code path at all (the existing LanceSink is reused unchanged across 1.18/1.19/1.20). So the SinkV2 migration and the version-matrix question are orthogonal to landing this skeleton.

On dropping 1.18. SinkV2 is actually available since Flink 1.15 (FLIP-191), so adopting SinkV2 wouldn't, by itself, force us to drop 1.18. The EOL argument also cuts both ways — 1.19 reached EOL in April 2025; only 1.20 is current LTS. If we want to trim the matrix on EOL grounds, that's a separate (and bigger) discussion worth its own issue, where we can also weigh existing user adoption.

Could I suggest we open a dedicated issue ("Migrate sink to SinkV2 / revisit supported Flink versions") to track that thread? Happy to review your branch there in detail — it looks like a substantial body of work that deserves its own focused PR rather than being folded into the build-system change.

[empathy87]

@fightBoxing, that would be great! I will open the issue "Migrate to SinkV2". And will prepare MR for it.