feat(namespace): add count_table_rows, insert_into_table, query_table

[XuQianJin-Stars]

close #6111

• Implement count_table_rows with optional version checkout and predicate filter
• Implement insert_into_table with append/overwrite modes via Arrow IPC
• Implement query_table supporting vector similarity search (with distance type, nprobes, refine factor, prefilter) and plain scan with filter/limit/offset/version
• Add lance-linalg dependency for DistanceType in vector search
• Add 8 unit tests covering all new methods and edge cases

Test plan

• cargo test -p lance-namespace-impls passes all new tests
• count_table_rows returns correct count with and without predicate filter
• insert_into_table correctly appends and overwrites data
• query_table returns correct results for vector search and plain scan

[github-actions]

PR Review

P0: multi_vector enters vector path but is never handled

In query_table, has_vector is true when multi_vector.is_some(), but the vector search path only handles single_vector:

let has_vector =
    request.vector.single_vector.is_some() || request.vector.multi_vector.is_some();
// ...
if has_vector {
    if let Some(ref single) = request.vector.single_vector {
        // only single_vector is handled
    }
    // multi_vector is silently ignored
}

If a caller passes multi_vector without single_vector, the code enters the vector path (skipping the plain-scan limit/offset logic) but never configures a nearest-neighbor query. This will silently produce wrong results. Either handle multi_vector or return an explicit "not supported" error when it's set.

P1: insert_into_table doesn't use session

count_table_rows and query_table both open datasets via DatasetBuilder with self.session, but insert_into_table uses Dataset::write with only ObjectStoreParams. If the namespace has a session configured, inserts won't use it. This is inconsistent and could cause failures in environments relying on session-based configuration.

P1: Empty-result schema may not reflect column projection

In query_table, when no batches are returned, the fallback creates an IPC stream using dataset.schema():

let schema: arrow_schema::Schema = dataset.schema().into();

This is the full dataset schema, not the projected schema. If the caller requested columns: ["id"], the empty-result IPC will contain the full schema instead. Consider capturing the projected schema from the scanner before streaming.

Minor: insert_into_table double-buffers all batches

The IPC stream is fully deserialized into Vec<RecordBatch> before being re-wrapped and passed to Dataset::write. Per project guidelines ("avoid collecting streams of RecordBatch into memory"), consider streaming directly or at minimum noting this as a known limitation for large payloads.

---

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.06950% with 15 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
rust/lance-namespace-impls/src/dir.rs	98.67%	8 Missing ⚠️
rust/lance-namespace-impls/src/rest_adapter.rs	95.95%	3 Missing and 4 partials ⚠️

📢 Thoughts on this report? Let us know!

[yanghua]

@XuQianJin-Stars Thanks for the contribution. Left two minor comments.

[yanghua]

To be more seriously, we may need to enumale the mode more carefully. For example, if it is not append, we need to throw an exception?

[yanghua]

Check the k arg? Should it be larger than 0?

[github-actions]

ACTION NEEDED
Lance follows the Conventional Commits specification for release automation.

The PR title and description are used as the merge commit message. Please update your PR title and description to match the specification.

For details on the error please inspect the "PR Title Check" action.

[jackye1995]

thanks for the work, I added the missing bindings all the way up