Improve VariableWidthBlock encoding by pettyjamesm · Pull Request #27377 · trinodb/trino

pettyjamesm · 2025-11-19T21:09:40Z

Description

Avoids converting between offsets and lengths when serializing and deserializing VariableWidthBlock instances, which enables a fast-path conversion for blocks without nulls present. When nulls are present, the compaction and expansion of offsets still outperforms the length to offset conversion.

Also removes an additional int per serialized block by referencing the offsets array to determine the slice length instead of sending it as a separate field.

Benchmarks

jmh.morethan.io - Significant improvements

# Before
Benchmark                                   (nullChance)  (offset)  Mode  Cnt  Score   Error  Units
BenchmarkBlockSerde.deserializeSliceDirect             0      true  avgt   10  1.139 ± 0.134  ns/op
BenchmarkBlockSerde.deserializeSliceDirect             0     false  avgt   10  1.097 ± 0.056  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .01      true  avgt   10  1.476 ± 0.026  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .01     false  avgt   10  2.066 ± 0.081  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .10      true  avgt   10  2.196 ± 0.038  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .10     false  avgt   10  2.228 ± 0.197  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .50      true  avgt   10  1.251 ± 0.013  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .50     false  avgt   10  1.234 ± 0.021  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .90      true  avgt   10  1.434 ± 0.023  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .90     false  avgt   10  1.624 ± 0.132  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .99      true  avgt   10  1.423 ± 0.069  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .99     false  avgt   10  1.443 ± 0.011  ns/op
BenchmarkBlockSerde.serializeSliceDirect               0      true  avgt   10  0.659 ± 0.010  ns/op
BenchmarkBlockSerde.serializeSliceDirect               0     false  avgt   10  0.645 ± 0.001  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .01      true  avgt   10  1.151 ± 0.039  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .01     false  avgt   10  1.162 ± 0.017  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .10      true  avgt   10  1.104 ± 0.040  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .10     false  avgt   10  1.174 ± 0.008  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .50      true  avgt   10  1.002 ± 0.014  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .50     false  avgt   10  0.962 ± 0.012  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .90      true  avgt   10  0.869 ± 0.001  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .90     false  avgt   10  0.901 ± 0.059  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .99      true  avgt   10  0.868 ± 0.012  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .99     false  avgt   10  0.796 ± 0.011  ns/op


# After
Benchmark                                   (nullChance)  (offset)  Mode  Cnt  Score   Error  Units
BenchmarkBlockSerde.deserializeSliceDirect             0      true  avgt   10  0.648 ± 0.013  ns/op
BenchmarkBlockSerde.deserializeSliceDirect             0     false  avgt   10  0.665 ± 0.013  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .01      true  avgt   10  1.302 ± 0.071  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .01     false  avgt   10  1.302 ± 0.017  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .10      true  avgt   10  1.267 ± 0.030  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .10     false  avgt   10  1.282 ± 0.074  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .50      true  avgt   10  0.973 ± 0.002  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .50     false  avgt   10  0.973 ± 0.017  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .90      true  avgt   10  0.743 ± 0.008  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .90     false  avgt   10  0.760 ± 0.058  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .99      true  avgt   10  0.688 ± 0.001  ns/op
BenchmarkBlockSerde.deserializeSliceDirect           .99     false  avgt   10  0.699 ± 0.021  ns/op
BenchmarkBlockSerde.serializeSliceDirect               0      true  avgt   10  0.538 ± 0.010  ns/op
BenchmarkBlockSerde.serializeSliceDirect               0     false  avgt   10  0.399 ± 0.033  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .01      true  avgt   10  1.035 ± 0.041  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .01     false  avgt   10  1.034 ± 0.013  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .10      true  avgt   10  1.003 ± 0.005  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .10     false  avgt   10  1.007 ± 0.004  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .50      true  avgt   10  0.909 ± 0.071  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .50     false  avgt   10  0.858 ± 0.014  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .90      true  avgt   10  0.754 ± 0.004  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .90     false  avgt   10  0.755 ± 0.001  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .99      true  avgt   10  0.725 ± 0.040  ns/op
BenchmarkBlockSerde.serializeSliceDirect             .99     false  avgt   10  0.718 ± 0.019  ns/op

Release notes

( ) This is not user-visible or is docs only, and no release notes are required.
( ) Release notes are required. Please propose a release note for me.
( ) Release notes are required, with the following suggested text:

## Section
* Improve performance of data exchanges involving variable width data ({issue}`27377`)

core/trino-spi/src/main/java/io/trino/spi/block/EncoderUtil.java

core/trino-spi/src/main/java/io/trino/spi/block/VariableWidthBlockEncoding.java

core/trino-spi/src/main/java/io/trino/spi/block/EncoderUtil.java

SongChujun · 2025-11-20T21:20:49Z

...faulttolerant-tests/src/test/java/io/trino/faulttolerant/BaseFaultTolerantExecutionTest.java

        @Language("SQL") String createTableSql =
                """
                CREATE TABLE test_table_writer_skew_mitigation WITH (%s = ARRAY['returnflag']) AS
-                SELECT orderkey, partkey, suppkey, linenumber, quantity, extendedprice, discount, tax, linestatus, shipdate, commitdate, receiptdate, shipinstruct, shipmode, comment, returnflag


Why do we want to delete column comment?

The test is highly dependent on the serialized page size, so adding 4 extra bytes per serialized page breaks the test. I'm duplicating the workaround from #15760 but this is indeed a hack.

I understand, hacky part is from the test setup itself though.

I'm actually going to avoid changing the serialized size- we don't need to add the extra 4 bytes anyway since we can just send the non-null ending offsets and not send the starting offset of 0 in the serialized representation.

...and it didn't work. Reintroducing the hack.

I actually think this is not a good fix to the spurious test, since we will probably always change block serde, another code change may fail on the case where comment is deleted, but pass on the case where comment is added back. I think we should change the assertion here instead https://github.com/trinodb/trino/blob/master/testing/trino-faulttolerant-tests/src/test/java/io/trino/faulttolerant/BaseFaultTolerantExecutionTest.java#L62C20-L62C39, the two numbers just happen to be equal and is not guaranteed to be equal on all cases.

starburstdata-automation · 2025-11-21T18:26:28Z

Started benchmark workflow for this PR with test type = iceberg/sf1000_parquet_part.

Building Trino finished with status: success
Benchmark finished with status: failure
Comparing results to the static baseline values, follow above workflow link for more details/logs.
Status message: Found regressions for:<br/>(presto/tpcds, q09, totalCpuTime, over by 33.2%)
Benchmark Comparison to the closest run from Master: Report

starburstdata-automation · 2025-11-21T18:26:39Z

Started benchmark workflow for this PR with test type = iceberg/sf1000_parquet_unpart.

Building Trino finished with status: success
Benchmark finished with status: success
Comparing results to the static baseline values, follow above workflow link for more details/logs.
Status message: NO Regression found.
Benchmark Comparison to the closest run from Master: Report

raunaqmorarka

lgtm % minor comments

core/trino-spi/src/main/java/io/trino/spi/block/VariableWidthBlockEncoding.java

raunaqmorarka · 2025-11-26T06:30:26Z

core/trino-spi/src/main/java/io/trino/spi/block/VariableWidthBlockEncoding.java

+                    .writeInts(rawOffsets, baseOffset + 1, positionCount);
+        }
+        else if (valueIsNull == null) {
+            // Subtract starting offset from each offset value to translate them to start from zero, no null suppression required


Is it possible to reuse io.trino.spi.block.BlockUtil#compactOffsets ?

Not quite, compactOffsets leaves the initial starting offset value of 0 in the resulting array whereas this branch omits it. I think we're better off with slight duplication to emphasize that we're working only with the ending offset for each position here.

core/trino-spi/src/main/java/io/trino/spi/block/VariableWidthBlockEncoding.java

Avoids converting between offsets and lengths when serializing and deserializing VariableWidthBlock instances, which enables a fast-path conversion for blocks without nulls present. When nulls are present, the compaction and expansion of offsets still outperforms the length to offset conversion.

Avoids encoding the slice length separately from the offsets, which can be used to infer the value directly.

cla-bot bot added the cla-signed label Nov 19, 2025

pettyjamesm force-pushed the improve-variable-width-block-encoding branch 5 times, most recently from a842e17 to f1d3e60 Compare November 20, 2025 14:50

pettyjamesm requested a review from radek-kondziolka November 20, 2025 15:29

pettyjamesm marked this pull request as ready for review November 20, 2025 16:24

pettyjamesm requested review from SongChujun, dain and sopel39 November 20, 2025 16:25

raunaqmorarka reviewed Nov 20, 2025

View reviewed changes

raunaqmorarka requested review from losipiuk and piotrrzysko and removed request for radek-kondziolka November 20, 2025 18:19

pettyjamesm force-pushed the improve-variable-width-block-encoding branch 2 times, most recently from 10a9db5 to e08afb0 Compare November 20, 2025 20:06

SongChujun reviewed Nov 20, 2025

View reviewed changes

pettyjamesm force-pushed the improve-variable-width-block-encoding branch 4 times, most recently from b076058 to 85afc34 Compare November 20, 2025 22:45

pettyjamesm requested a review from raunaqmorarka November 21, 2025 16:40

pettyjamesm force-pushed the improve-variable-width-block-encoding branch 3 times, most recently from db6b908 to ddaef15 Compare November 24, 2025 18:36

raunaqmorarka reviewed Nov 26, 2025

View reviewed changes

losipiuk approved these changes Nov 26, 2025

View reviewed changes

pettyjamesm added 4 commits November 26, 2025 10:57

Fix spurious test failure

e2f4d7a

Add offsets to BenchmarkBlockSerDe

3892452

Remove redundant size in serialized variable width block

99fa73f

Avoids encoding the slice length separately from the offsets, which can be used to infer the value directly.

pettyjamesm force-pushed the improve-variable-width-block-encoding branch from ddaef15 to 99fa73f Compare November 26, 2025 16:21

pettyjamesm merged commit cd128dc into trinodb:master Nov 26, 2025
101 checks passed

pettyjamesm deleted the improve-variable-width-block-encoding branch November 26, 2025 17:13

github-actions bot added this to the 479 milestone Nov 26, 2025

chenjian2664 mentioned this pull request Nov 28, 2025

Add Trino 479 release notes #27153

Merged

Conversation

pettyjamesm commented Nov 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Benchmarks

Release notes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

starburstdata-automation commented Nov 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

starburstdata-automation commented Nov 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

raunaqmorarka left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

5 participants

pettyjamesm commented Nov 19, 2025 •

edited

Loading

starburstdata-automation commented Nov 21, 2025 •

edited

Loading

starburstdata-automation commented Nov 21, 2025 •

edited

Loading