Struct RunWriter 

pub struct RunWriter<'a> { /* private fields */ }

Builds and writes a sorted run from drained memtable rows.

rows must be sorted ascending by (row_id, epoch) (the memtable’s natural drain order). System columns _row_id, _epoch, _deleted are always emitted; each user column in schema is emitted, with Null for rows that don’t set it.

Implementations

impl<'a> RunWriter<'a>

pub fn new( schema: &'a Schema, run_id: u128, epoch_created: Epoch, level: u8, ) -> Self

pub fn uniform_epoch(self, uniform: bool) -> Self

Mark this run as uniform-epoch: its stored _epoch column is a placeholder and the real commit epoch is supplied at read time from the manifest RunRef.epoch_created (see RUN_FLAG_UNIFORM_EPOCH). Used by the large-transaction spill path, which writes the run before the commit epoch is assigned.

pub fn with_encryption( self, kek: &'a Kek, indexable_columns: Vec<(u16, u8)>, ) -> Self

Encrypt this run’s pages with a fresh per-file DEK wrapped by kek. indexable_columns are the ENCRYPTED_INDEXABLE (column_id, scheme) pairs whose column keys are derived+wrapped into the descriptor.

pub fn with_zstd_level(self, level: i32) -> Self

Override the zstd level for this run’s pages (Phase 14.4). The bulk ingest path uses level 1; background compaction upgrades cold runs to 3.

pub fn with_lz4(self) -> Self

Compress hot/mutable-run pages with LZ4 (Phase 15.3): 3–5× faster decode than zstd with ~10% worse ratio. The right default for runs that get scanned (bulk-loaded analytical runs).

pub fn with_plain(self) -> Self

Emit raw ALGO_PLAIN pages with no compression (Phase 14.4 bulk_load_fast): maximal encode throughput at the cost of ~3–4× size.

pub fn clean(self, clean: bool) -> Self

Mark this run as “clean” (one version per RowId, no tombstones, ascending row_ids). Only set when the caller constructs the system columns so by construction (typed bulk load of fresh, contiguous row_ids; compaction output that has collapsed versions and dropped tombstones). Stamps RUN_FLAG_CLEAN into the header so readers skip the MVCC pass.

pub fn with_native_endian(self) -> Self

Write fixed-width page payloads in little-endian (Phase 15.7): the decode path becomes a memcpy on little-endian targets. Only effective on the typed write_native path; no-op on big-endian writers (they keep the portable BE layout, since “native” would not be LE there).

pub fn write_native( self, path: impl AsRef<Path>, user_columns: &[(u16, NativeColumn)], n: usize, first_row_id: u64, ) -> Result<RunHeader>

Write a run straight from typed columns (no Value). user_columns are the schema’s user columns as [NativeColumn]s; the system columns (_row_id/_epoch/_deleted) are built from first_row_id..+n / epoch_created / all-false. Sorted Int64 columns (always the system _row_id, plus any sorted user Int64) use delta encoding unless RunWriter::with_plain forces raw ALGO_PLAIN everywhere.

pub fn write(self, path: impl AsRef<Path>, rows: &[Row]) -> Result<RunHeader>