[SPARK-56802][SQL] Add bulk read+widen path for FLOAT to Double Parquet vector updater#55816
Closed
LuciferYang wants to merge 4 commits into
Closed
[SPARK-56802][SQL] Add bulk read+widen path for FLOAT to Double Parquet vector updater#55816LuciferYang wants to merge 4 commits into
LuciferYang wants to merge 4 commits into
Conversation
…et vector updater
LuciferYang
force-pushed
the
SPARK-56802-float-to-double
branch
from
308150a to
051b94b
Compare
…uet.ParquetVectorUpdaterBenchmark (JDK 17, Scala 2.13, split 1 of 1)
…uet.ParquetVectorUpdaterBenchmark (JDK 21, Scala 2.13, split 1 of 1)
…uet.ParquetVectorUpdaterBenchmark (JDK 25, Scala 2.13, split 1 of 1)
LuciferYang
commented
May 12, 2026
| DowncastLongUpdater (INT64 -> Decimal(9,2)) 2 2 0 455.0 2.2 0.4X | ||
| IntegerToLongUpdater 1 1 0 1280.6 0.8 1.0X | ||
| IntegerToDoubleUpdater 1 1 0 1537.9 0.7 1.2X | ||
| FloatToDoubleUpdater 1 1 0 1418.8 0.7 1.1X |
Contributor
Author
There was a problem hiding this comment.
This is the embodiment of this optimization.
Contributor
Author
yaooqinn
approved these changes
May 13, 2026
LuciferYang
added a commit
that referenced
this pull request
May 13, 2026
…et vector updater ### What changes were proposed in this pull request? Extend the bulk read+widen pattern introduced in SPARK-56791 to `FloatToDoubleUpdater` (parquet FLOAT read into Spark `DoubleType`). A new `readFloatsAsDoubles` default method on `VectorizedValuesReader` does the per-row fallback. `VectorizedPlainValuesReader` overrides it to fetch source bytes once via `getBuffer(total * 4)` and run a tight in-method conversion loop. `FloatToDoubleUpdater.readValues` becomes a one-line delegation. The widen is Java's primitive float-to-double conversion: exact for every finite and infinite float; a NaN float widens to a double NaN (the JVM may canonicalize the payload). ### Why are the changes needed? `FloatToDoubleUpdater.readValues` allocates a fresh `ByteBuffer` slice inside `getBuffer(4)` for every element on the legacy path, and that allocation dominates the loop. Collapsing N allocations into one is the same win SPARK-56791 delivered for the INT32 -> Long sibling, with the gain again amplifying on newer JDKs where escape analysis better optimizes the tight loop. Baselines and after-numbers are taken directly from the benchmark-result diffs committed on this branch: | JDK | Baseline | After | Speedup | |----:|---------:|----------:|--------:| | 17 | 480.2 M/s | 1418.8 M/s | 2.95x | | 21 | 489.5 M/s | 2527.2 M/s | 5.16x | | 25 | 483.3 M/s | 2570.7 M/s | 5.32x | Peer Updaters in the same benchmark group hold within run-to-run noise, confirming the change is local to `FloatToDoubleUpdater`. ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? New unit tests in `ParquetVectorUpdaterSuite` cover boundary batch lengths (0, 1, 7, 8, 9, 17, 1024, 4097), the singular `readValue` path, and special values (signed zeros, +/-Infinity, NaN, with raw-bit assertions on signed zeros). A NaN-safe `assertDoublesEqual` helper uses `java.lang.Double.compare` so NaN equality and signed-zero distinction are well-defined. A new end-to-end test in `ParquetIOSuite` round-trips FLOAT written to parquet and read back as `DoubleType`, exercising both REQUIRED columns (no def-levels) and OPTIONAL columns with interleaved nulls so that `readValue` and `readValues` are both invoked. Benchmark results on JDK 17, 21, and 25 are committed on the branch. ### Was this patch authored or co-authored using generative AI tooling? Generated-by: Claude Code Closes #55816 from LuciferYang/SPARK-56802-float-to-double. Authored-by: YangJie <yangjie01@baidu.com> Signed-off-by: yangjie01 <yangjie01@baidu.com> (cherry picked from commit bc4bf69) Signed-off-by: yangjie01 <yangjie01@baidu.com>
Contributor
Author
|
Merged into master. Thanks @yaooqinn |
Member
|
+1, LGTM. |
Contributor
Author
|
Thank you @dongjoon-hyun |
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
3 participants
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.
What changes were proposed in this pull request?
Extend the bulk read+widen pattern introduced in SPARK-56791 to
FloatToDoubleUpdater(parquet FLOAT read into SparkDoubleType).A new
readFloatsAsDoublesdefault method onVectorizedValuesReaderdoes the per-row fallback.VectorizedPlainValuesReaderoverrides it to fetch source bytes once viagetBuffer(total * 4)and run a tight in-method conversion loop.FloatToDoubleUpdater.readValuesbecomes a one-line delegation. The widen is Java's primitive float-to-double conversion: exact for every finite and infinite float; a NaN float widens to a double NaN (the JVM may canonicalize the payload).Why are the changes needed?
FloatToDoubleUpdater.readValuesallocates a freshByteBufferslice insidegetBuffer(4)for every element on the legacy path, and that allocation dominates the loop. Collapsing N allocations into one is the same win SPARK-56791 delivered for the INT32 -> Long sibling, with the gain again amplifying on newer JDKs where escape analysis better optimizes the tight loop. Baselines and after-numbers are taken directly from the benchmark-result diffs committed on this branch:Peer Updaters in the same benchmark group hold within run-to-run noise, confirming the change is local to
FloatToDoubleUpdater.Does this PR introduce any user-facing change?
No.
How was this patch tested?
New unit tests in
ParquetVectorUpdaterSuitecover boundary batch lengths (0, 1, 7, 8, 9, 17, 1024, 4097), the singularreadValuepath, and special values (signed zeros, +/-Infinity, NaN, with raw-bit assertions on signed zeros). A NaN-safeassertDoublesEqualhelper usesjava.lang.Double.compareso NaN equality and signed-zero distinction are well-defined.A new end-to-end test in
ParquetIOSuiteround-trips FLOAT written to parquet and read back asDoubleType, exercising both REQUIRED columns (no def-levels) and OPTIONAL columns with interleaved nulls so thatreadValueandreadValuesare both invoked.Benchmark results on JDK 17, 21, and 25 are committed on the branch.
Was this patch authored or co-authored using generative AI tooling?
Generated-by: Claude Code