[Feature] Support table-level full-text search for data-evolution tables #404

Search before asking

• I searched in the issues and found nothing similar.

Motivation

Sub-issue of #399 (step 4: table-level full-text search).

Paimon C++ only exposes the reader primitive. To run a full-text search, a caller has to:

1. call GlobalIndexScan::CreateReaders;
2. call VisitFullTextSearch on the readers;
3. pass the result to ScanContextBuilder::SetGlobalIndexResult.

Along the way the caller also has to choose index types and shard ranges, and there is no serializable plan for running scan and read in different processes. Java offers a table-level API instead:

GlobalIndexResult result =
        table.newFullTextSearchBuilder()
                .withQuery("content", "{\"match\":{\"query\":\"paimon lake\"}}")
                .withLimit(10)
                .executeLocal();
TableScan.Plan plan = table.newReadBuilder().newScan().withGlobalIndexResult(result).plan();

The Java components live in paimon-core org.apache.paimon.table.source, plus one method in paimon-common:

Solution

1. API. Add C++ equivalents of the builder, scan, read and splits, following the existing builder conventions (ScanContextBuilder / ReadContextBuilder). Splits must be serializable so scan and read can run in different processes. The read result plugs into ScanContextBuilder::SetGlobalIndexResult, and scores stay readable through _INDEX_SCORE.
2. Capability. Add SupportsFullTextSearch() to GlobalIndexerFactory or GlobalIndexer, returning true for full-text. Decide whether lucene-fts returns true as well.
3. Scan, following DataEvolutionFullTextScan:
   • Pin the snapshot: the latest one, or the time-travel target.
   • Keep index manifest entries that have GlobalIndexMeta and pass the partition filter.
   • Treat a file as a full-text input when the column is its index_field_id or one of its extra_field_ids, and its factory supports full-text search.
   • Select ranges across overlapping and multi-field indexes ([global-index] Fix full-text range selection and scalar-filter coverage for multi-field indexes paimon#8547):
     1. Group candidates by column → physical range → index identity, where the identity is type|index_field_id|extra_field_ids.
     2. Sweep the elementary intervals and give each one to the best active candidate. A dedicated index comes first, then order by identity, from and to.
     3. Merge adjacent intervals.
     4. Emit one IndexFullTextSearchSplit per selection. It keeps the physical range, which is needed for the row-id offset, and the assigned searchRowRanges.
4. Read, following DataEvolutionFullTextRead:
   • Run on an executor sized by global-index.thread-num.
   • For each split:
     • Create the indexer from the split's own file meta, including any extra fields, and wrap its reader in OffsetGlobalIndexReader(start, end).
     • Search with FullTextSearch(column, query, limit), restricted to searchRowRanges ∧ liveRows. Pass no restriction when this covers the whole physical range, and skip the split when it is empty.
     • Close the reader afterwards.
   • Merge the per-split results, keeping the first score for a duplicate row id. Then apply the final topK(limit), ordered by score descending and then row id ascending. Each split only needs limit candidates ([core] Request only top-k candidates from data-evolution full-text indexes paimon#9955).
   • Pin the live-row computation ([Bug] Deleted rows take full-text top-k slots in data-evolution tables #403) and all reads to the plan snapshot ([core] Support full-text indexes on DV tables paimon#9031).
5. Tests. Cover:
   • multiple shards
   • overlapping index ranges
   • a text column that is an extra field of another index
   • partition filters
   • snapshot pinning
   • deletion-vector tables
   • a split serialization round trip

Anything else?

• Depends on [Feature] Replace the tantivy FFI with the paimon-full-text-index native engine #400, [Feature] Align FullTextSearch with the Java JSON DSL query model #401, [Bug] Full-text search over multiple index shards returns up to shards × limit rows #402 and [Bug] Deleted rows take full-text top-k slots in data-evolution tables #403.
• Java's IndexFullTextSearchSplit uses Java serialization, so there is no cross-language wire format to match. C++ should define its own versioned serialization.
• Java rejects search on tables under query authorization ([core] Fix column masking correctness in query-auth reads paimon#8570). C++ should do the same if it adds query auth.

Are you willing to submit a PR?

• I'm willing to submit a PR!