Struct DeltaDataset 

pub struct DeltaDataset { /* private fields */ }

LSM-style delta dataset for a single edge type and direction.

Stores L1 sorted runs that accumulate edge mutations before compaction merges them into the base CSR.

Implementations

impl DeltaDataset

pub fn new(_base_uri: &str, edge_type: &str, direction: &str) -> Self

Create a new DeltaDataset for the given edge type and direction.

_base_uri is ignored: physical path resolution belongs to the storage backend; this type holds only the table’s logical identity.

pub fn new_branched( base_uri: &str, edge_type: &str, direction: &str, branch: impl Into<String>, ) -> Self

Construct a delta dataset that reads from a Lance branch.

pub fn get_arrow_schema(&self, schema: &Schema) -> Result<Arc<ArrowSchema>>

Build the Arrow schema for this delta table using the given graph schema.

pub fn build_record_batch( &self, entries: &[L1Entry], schema: &Schema, ) -> Result<RecordBatch>

Serialize entries into an Arrow RecordBatch using the given graph schema.

pub async fn open_or_create( &self, backend: &dyn StorageBackend, schema: &Schema, ) -> Result<()>

Open or create a delta table via the storage backend.

pub async fn write_run( &self, backend: &dyn StorageBackend, batch: RecordBatch, ) -> Result<()>

Write a run to a delta table.

Creates the table if it doesn’t exist, otherwise appends to it. Race-safe under async-flush — see crate::storage::manager::write_batch_with_lance_conflict_retry.

pub fn build_partial_record_batch( &self, entries: &[L1Entry], touched_keys: &HashSet<String>, schema: &Schema, ) -> Result<RecordBatch>

Build a partial-column RecordBatch for Lance MergeInsert. Includes the join key (eid), the system columns (src_vid, dst_vid, op=0/Insert, _version, _updated_at), and ONLY the schema-defined property columns whose name appears in touched_keys. If any of the entry’s properties are NOT in the label schema, overflow_json is regenerated with all overflow properties present in the entry (the JSONB blob is one column; it must be rewritten in full because we can’t merge JSON fragments at the storage layer).

Used by Round-12 §A: edge SETs route through this builder so the per-edge-type delta tables receive only the touched schema columns; untouched columns retain their previous-version value via Lance’s MVCC WhenMatched::UpdateAll semantics.

pub async fn merge_insert_partial_run( &self, backend: &dyn StorageBackend, batch: RecordBatch, ) -> Result<()>

MergeInsert a partial-column batch via Lance. Join key is eid. Matched rows have WhenMatched::UpdateAll applied; unmatched source rows are dropped (a partial SET can only update an edge that was previously CREATEd — the full-row Append path landed the original row).

pub async fn ensure_eid_index(&self, backend: &dyn StorageBackend) -> Result<()>

Ensure a BTree index exists on the ‘eid’ column.

pub fn table_name(&self) -> String

Get the table name for this delta dataset.

pub async fn scan_all_backend( &self, backend: &dyn StorageBackend, schema: &Schema, ) -> Result<Vec<L1Entry>>

Scan all entries from the backend table.

Returns an empty vector if the table doesn’t exist.

pub async fn scan_all_backend_with_limit( &self, backend: &dyn StorageBackend, schema: &Schema, max_rows: usize, ) -> Result<Vec<L1Entry>>

Scan all entries from the backend table with a configurable row limit to prevent OOM.

pub async fn replace( &self, backend: &dyn StorageBackend, batch: RecordBatch, ) -> Result<()>

Replace the delta table with a new batch (atomic replacement).

This is used during compaction to clear the delta table after merging into L2.

pub async fn delete_up_to_version( &self, backend: &dyn StorageBackend, hwm: u64, ) -> Result<()>

Delete every delta row at or below a version high-water-mark.

Compaction uses this to clear ONLY the deltas it actually merged into L2 — the ones present at read time, whose max _version is hwm. Unlike a full table wipe (Self::replace with an empty batch), this preserves rows a concurrent flush appended after the compaction read them: those carry a strictly higher _version (flush versions are monotonic and a flush’s deltas land atomically), so _version <= hwm never matches them. This is what makes the clear safe without depending on an instantaneous flush_in_progress check. (review H11)

pub async fn read_deltas( &self, backend: &dyn StorageBackend, vid: Vid, schema: &Schema, version_hwm: Option<u64>, ) -> Result<Vec<L1Entry>>

Read delta entries for a specific vertex ID.

Returns an empty vector if the table doesn’t exist or no entries match.

pub async fn read_deltas_batch( &self, backend: &dyn StorageBackend, vids: &[Vid], schema: &Schema, version_hwm: Option<u64>, ) -> Result<HashMap<Vid, Vec<L1Entry>>>

Read delta entries for multiple vertex IDs in a single batch query.

Returns a HashMap mapping each vid to its delta entries. VIDs with no delta entries will not be in the map.

Trait Implementations

impl Debug for DeltaDataset

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.