Expand description
ColumnChunk: differential’s Chunk over Column-shaped updates.
A chunk is a sorted, consolidated run of (D, T, R) updates in the flat
columnar layout, in one of two homes:
- Resident: an
Rc-sharedColumnon the heap. Fresh input, merge output, and small tails live here. - Spilled: the serialized body in the process
Pool, with the record count and the first and last data items resident. The pool owns residency from there, with slots under a memory budget and compression and device pageout under pressure, and a body that dies before pressure reaches it is freed without I/O.
Reads of a spilled body are copy-out and scoped to the call that needs them: the body is read into caller-owned memory and no reference into pool memory ever exists outside the pool. That contract is what lets the pool evict with no reader accounting at all.
Spilling happens in Chunk::settle, the trait’s designated commit point:
chunks moved to settled output are handed to the pool when spilling is
enabled (see set_compute_spill_enabled and set_storage_spill_enabled).
The spill destination resolves per commit from three pieces of mutable
state. A thread-local pool override, for tests and benches, wins outright.
Otherwise the compute and storage gates, composed as an OR, route commits
to the process pool installed by crate::pool_config, and with no pool
installed chunks stay resident. A second thread-local holds the reusable
scratch that call-scoped reads of spilled bodies copy into.
Grading is by serialized bytes, the ship size
Column already targets, rather than by the record-count TARGET,
since record count does not bound bytes for variable-width data.
Chunks whose data is a (key, val) pair additionally implement
UnloadChunk, the bulk-read capability: sorted probe keys in, matching
updates appended to caller-owned staging, with locate answered from the
resident fence metadata so a probe set faults only the chunk bodies it
actually touches.
Structs§
- Chunk
Chunker - A chunker for
arrange_coreoverColumnChunks: sorts and consolidates raw input columns through aColumnChunkerand wraps its output chunks. - Lz4Codec
- The chunk-side
ExtentCodec: a little-endianu32body-length prefix followed by one lz4 block, the framinglz4_flex::block::compress_prepend_sizeproduces. Every chunk consumer passesLZ4_CODECat insert; the pool itself has no codec opinion. - Spilled
Body - A spilled chunk body: the serialized column in the pool, plus the resident
metadata every
Chunkmust answer without fetching. That metadata is the record count and the first and last data items (the fence entriesUnloadChunk::locateconsults). - Unchunk
Builder - A batch builder over
ColumnChunkinput that delegates to a builder overColumninput, loading each chunk’s body as it is pushed.
Enums§
- Column
Chunk - A sorted, consolidated run of
(D, T, R)updates, resident or spilled.
Statics§
- LZ4_
CODEC - The
Lz4Codecinstance chunk consumers pass toPool::insert_with.
Functions§
- set_
compute_ spill_ enabled - Enable or disable chunk spilling on behalf of compute’s arrangement batchers.
- set_
spill_ override - Set or unset the pool through which this thread’s chunk spills are
routed, taking precedence over the gates and the process pool.
Nonerestores the global resolution. - set_
storage_ spill_ enabled - Enable or disable chunk spilling on behalf of storage’s upsert dataflows.