ListLayout baseline#8706
Open
mhk197 wants to merge 48 commits into
Open
Conversation
Signed-off-by: "Matthew Katz" <katz@spiraldb.com> Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matthew Katz" <katz@spiraldb.com> Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matthew Katz" <katz@spiraldb.com> Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: "Matt Katz" <mhkatz97@gmail.com> Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Signed-off-by: Matt Katz <mhkatz97@gmail.com>
…ST_LAYOUT Rework ListLayoutStrategy into a streaming transpose driven by the TableStrategy dispatcher (global u64 offsets, memory-bounded), add bounded element reads and is_null pruning routing in ListReader, and gate the whole thing off by default behind the VORTEX_EXPERIMENTAL_LIST_LAYOUT env var / with_list_layout(). Signed-off-by: Matt Katz <mhkatz97@gmail.com>
Merging this PR will not alter performance
Warning Please fix the performance issues or acknowledge them on CodSpeed. Performance Changes
Tip Investigate this regression by commenting Comparing 
Footnotes
|
Signed-off-by: Matt Katz <mhkatz97@gmail.com>
mhk197
added
changelog/skip
github-actions
Bot
removed
the
action/benchmark-sql
Contributor
Polar Signals Profiling ResultsLatest Run
Powered by Polar Signals Cloud |
Contributor
Benchmarks: PolarSignals ProfilingVortex (geomean): 1.034x ➖ How to read Verdict and Engines
datafusion / vortex-file-compressed (1.034x ➖, 0↑ 0↓)
File Size Changes (1 files changed, -0.0% overall, 0↑ 1↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=1 on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.993x ➖, 0↑ 0↓)
datafusion / parquet (1.002x ➖, 0↑ 0↓)
datafusion / arrow (1.012x ➖, 0↑ 1↓)
duckdb / vortex-file-compressed (1.002x ➖, 0↑ 0↓)
duckdb / parquet (0.989x ➖, 3↑ 1↓)
File Size Changes (17 files changed, -44.5% overall, 4↑ 13↓)
Totals:
|
Contributor
Benchmarks: FineWeb NVMeVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.024x ➖, 0↑ 0↓)
datafusion / parquet (0.975x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (1.010x ➖, 0↑ 0↓)
duckdb / parquet (0.982x ➖, 0↑ 0↓)
File Size Changes (3 files changed, -46.3% overall, 0↑ 3↓)
Totals:
|
Contributor
Benchmarks: TPC-DS SF=1 on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.019x ➖, 0↑ 4↓)
datafusion / parquet (1.028x ➖, 0↑ 4↓)
duckdb / vortex-file-compressed (1.013x ➖, 2↑ 2↓)
duckdb / parquet (1.004x ➖, 0↑ 1↓)
File Size Changes (30 files changed, -43.5% overall, 2↑ 28↓)
Totals:
|
Contributor
Benchmarks: Clickbench Sorted on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.965x ➖, 1↑ 0↓)
datafusion / parquet (1.007x ➖, 1↑ 0↓)
duckdb / vortex-file-compressed (0.949x ➖, 2↑ 0↓)
duckdb / parquet (0.988x ➖, 0↑ 0↓)
File Size Changes (201 files changed, -42.6% overall, 53↑ 148↓)
Totals:
|
Contributor
Benchmarks: FineWeb S3Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.084x ➖, 0↑ 1↓)
datafusion / parquet (1.055x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (1.071x ➖, 0↑ 1↓)
duckdb / parquet (1.029x ➖, 0↑ 0↓)
|
Contributor
Benchmarks: Statistical and Population GeneticsVerdict: No clear signal (low confidence) How to read Verdict and Engines
duckdb / vortex-file-compressed (1.047x ➖, 0↑ 0↓)
duckdb / parquet (1.038x ➖, 0↑ 0↓)
File Size Changes (3 files changed, -32.3% overall, 0↑ 3↓)
Totals:
|
Contributor
Benchmarks: Clickbench on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.002x ➖, 0↑ 0↓)
datafusion / parquet (0.992x ➖, 2↑ 0↓)
duckdb / vortex-file-compressed (0.989x ➖, 5↑ 1↓)
duckdb / parquet (0.997x ➖, 1↑ 0↓)
File Size Changes (201 files changed, -39.1% overall, 55↑ 146↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=10 on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.001x ➖, 0↑ 0↓)
datafusion / parquet (0.998x ➖, 0↑ 0↓)
datafusion / arrow (1.022x ➖, 0↑ 1↓)
duckdb / vortex-file-compressed (0.980x ➖, 0↑ 0↓)
duckdb / parquet (0.997x ➖, 0↑ 1↓)
File Size Changes (47 files changed, -44.4% overall, 10↑ 37↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=1 on S3Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.082x ➖, 0↑ 2↓)
datafusion / parquet (1.132x ➖, 0↑ 5↓)
duckdb / vortex-file-compressed (1.096x ➖, 0↑ 0↓)
duckdb / parquet (1.108x ➖, 0↑ 1↓)
|
- Extract transpose_list_column with named senders in the writer - Fold use_experimental_list_layout into the builder default - Reader: fetch_raw_offsets/elements helpers, clearer names, doc fixes Signed-off-by: Matt Katz <mhkatz97@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
List Layout
Decomposes a
list<T>column into three independently-written sub-columns —elements,offsets, and (when nullable)validity— each backed by its own layout, under a singlevortex.listnode.There are three pieces: the dispatcher enables using
ListLayoutfor arrays withDType::List, the writer shreds a list stream into the three child streams, and the reader reassembles only the children an expression actually needs.Note that this a prototype and is not a performance improvement over flattening lists... yet. As such, this PR does not yet enable writing
ListLayoutfor benchmarks.1.
Recursive Writer DispatchTableStrategyinspects the dtype of the stream it's handed and routes:StructStrategyListLayoutStrategy(when list decomposition is on)To enable recursive handling of nested lists (elements are lists) or structs it hands a descended copy of itself as
the child strategy. So a
list<struct<{ a, list<i32> }>>recurses with no manual wiring — eachlevel dispatches on its own dtype.
List decomposition is gated (
use_experimental_list_layout()/VORTEX_EXPERIMENTAL_LIST_LAYOUT,or an explicit builder call). When off, lists fall through to the leaf strategy unchanged.
2.
ListLayoutStrategyA structural writer: it does not compress or inspect element dtypes; it shreds a list stream into
its three sub-streams and hands each to its own downstream strategy, producing one
ListLayout.To do so it must:
Listparts. A gapped/reorderedListViewis rebuiltinto a gapless
Listfirst (list_from_list_view).u64. Each chunk's local offsets are shifted byelement_base(the running count of elements emitted so far) and widened to
u64, and the duplicated boundaryoffset is dropped on every chunk after the first — so the concatenation of all chunks is one
monotonic
[0 … total_elements]array of lengthrow_count + 1that indexes into the singleconcatenated
elementschild.elements,offsets, andvalidityonto three bounded (capacity-1) channels; eachchild is written concurrently by its own strategy via
spawn_nested. The producer future isjoined with the child writers (
try_join) so a producer error surfaces instead of being hiddenas an early channel close.
Non-list input is forwarded to a
fallbackstrategy unchanged.Because each child is written through an independent strategy, the children's physical shape is
decoupled: routing
elementsthrough the recursive dispatcher (→ repartition → chunked) makes it aChunkedLayoutchunked in element space, whileoffsets/validitycan be written with a different(e.g. row-space) strategy.
3.
ListReaderTo read as little as possible,
projection_evaluationfirst classifies theexpression (
get_necessary_list_children) into the minimal set of children it needs, then routesto a matching read path:
ListChildrenNeeded)Validityis_null(x)/is_not_null(x)project_validityOffsetsAndValiditylist_length(x)project_offsets_validityAllx, or anything over elementsproject_allproject_allmaterializes the list, and itself picks between two sub-paths:project_all_concurrent) — used for a full-column range, or whenelementsis a single flat segment (nothing to skip). Fires theelements,offsets, andvalidityreads concurrently, assembles the list, slices to the requested range, and filters.No offsets→elements ordering dependency.
project_all_elements_bounded) — used for a strict sub-range over chunkedelements. Reads
offsets[range], decodes the first/last offset to bound the elements read to[first … last), fetches only the element chunks that range overlaps, and rebases the offsetsto index into the sliced buffer. Costs one offsets→elements round-trip but avoids reading the
whole elements column for a partial scan split.
TODOs:
Status / open decisions
ListLayoutinstead of just chunking children? Doing aligns the row range covered by element chunks and potentially offset chunks within a list chunk.