[Feature] Maintain primary-key full-text index archives on write and compaction

Start with src/paimon/core/index/pk/bucketed_primary_key_index_maintainer.cpp and the level-planning code in src/paimon/core/index/pksorted/, after reviewing dependencies #400 and #408. Use the Java-aligned PkFullTextIndexFileTest, PkFullTextDataFileReaderTest, PkFullTextBucketIndexStateTest, and BucketedFullTextIndexMaintainerTest as behavioral references; done means archives are maintained through restore, commit, compaction, and abort, with the listed cross-language tests passing.

Search before asking

• I searched in the issues and found nothing similar.

Motivation

Sub-issue of #399 (step 5: primary-key full-text index, write side).

BucketedPrimaryKeyIndexMaintainer keeps only BTree definitions (src/paimon/core/index/pk/bucketed_primary_key_index_maintainer.cpp:128-132), and restore keeps only BTree payloads (IsPrimaryKeyBTreePayload). This has two effects:

• Paimon C++ writers never build primary-key full-text archives.
• When C++ compacts a table that already has Java-written archives, those archives are neither rebuilt nor retired. They keep pointing at data files that compaction has removed, and the new compacted files are not indexed.

Java model (apache/paimon#8649, apache/paimon#8651, apache/paimon#8672, apache/paimon#8992):

• One archive per level. Each (partition, bucket, non-zero data level) has exactly one immutable archive. It covers all eligible files of that level, sorted by file name.
• Eligible files. A file is eligible when fileSource == COMPACT && level > 0 (PrimaryKeyIndexSourcePolicy).
• Row ids. An archive's row ids are the concatenated physical row positions of its source files, including rows deleted by deletion vectors and null rows.
• PkFullTextIndexFile builds one archive:
  • index type full-text;
  • every source has the same level > 0 and a positive row count;
  • it calls GlobalIndexSingleColumnWriter#write(text, sourceOffset + rowPos) for each row;
  • finish() must return exactly one entry whose row count equals the total source row count.
  • The resulting IndexFileMeta has GlobalIndexMeta(0, total - 1, fieldId, null, indexMeta, sourceMeta):
    • indexMeta is the flat JSON of the prefix-stripped options;
    • sourceMeta is PrimaryKeyIndexSourceMeta(level, sourceFiles).
  • The file is named index-{uuid}-{N} under the index directory, or in the bucket directory when index-file-in-data-file-dir is set.
• PkFullTextDataFileReader reads one text value per physical row, with no deletion-vector filtering. PkFullTextIndexBuilder builds an archive for one file or a list of files.
• PkFullTextBucketIndexState#fromActiveDataFiles classifies payloads:
  • A payload is current only if its (level, ordered source files) exactly matches the level's eligible active files and its row counts match.
  • A payload that fails this is stale. So is every payload of a level that has more than one match, and any payload whose metadata cannot be parsed.
• Level planning (PrimaryKeyIndexLevels, shared with the sorted indexes) picks the lowest level whose payload is missing or out of date. A plan with no source files removes the payload.
• BucketedFullTextIndexMaintainer:
  • Restore: stale payloads are retired and emitted as deletions in the next commit.
  • prepareCommit(append, compact, waitCompaction):
    1. Applies the data transition: removes compactBefore files and adds eligible compactAfter files.
    2. Finishes or starts a single background build.
    3. Atomically replaces the level's archive when the build is still valid (canAccept); otherwise deletes the generated file.
    4. Routes the result to the compact increment if there was a compact transition, and to the append increment otherwise.
    5. Rolls back and deletes generated files on failure. The returned commit carries an abort hook.
  • A failed build is not retried within the same call. The next prepareCommit plans again.
• Wiring:
  • BucketedPrimaryKeyIndexMaintainer.Factory creates the full-text maintainer for fixed-bucket writes and for postpone-bucket compaction (apache/paimon#8992).
  • IndexFileHandler#pkFullTextIndex(partition, bucket) provides the index file.
  • Writers are not closed while a build is pending.

Solution

• Port the classes above.
• Reuse the existing C++ PrimaryKeyIndexSourceMeta, PrimaryKeyIndexSourcePolicy and PrimaryKeyIndexSourceFile. Extract the level planning currently embedded in src/paimon/core/index/pksorted/ so it can be shared.
• Wire the full-text maintainer into BucketedPrimaryKeyIndexMaintainer: restore, prepareCommit, abort, and merging the increments. Create it through the full-text indexer from #400, using the options resolved in #408.
• Add tests aligned with Java PkFullTextIndexFileTest, PkFullTextDataFileReaderTest, PkFullTextBucketIndexStateTest and BucketedFullTextIndexMaintainerTest, plus:
  • C++ compaction of a table with Java-written primary-key full-text archives; stale archives are retired and new ones built;
  • Java reading archives written by C++.

Anything else?

• Depends on #400 and #408.
• The realtime path still rejects primary-key global indexes (src/paimon/core/utils/primary_key_table_utils.cpp:136-141). That is out of scope here.

Are you willing to submit a PR?

• I'm willing to submit a PR!