SDK Performance Boost by lisandroct · Pull Request #5840 · clockworklabs/SpacetimeDB

lisandroct · 2026-08-28T18:29:49Z

Description of Changes

It cleans up a couple of things to achieve better performance in the C# SDK:

It replaces the usage of LINQ (with direct for loops, for example);
Reduces the total amount of allocations by pooling arrays and avoiding some calls to .ToArray();
Reduces the initial size of websockets buffers and resizes them when necessary (this also reduces a lot of allocations).

API and ABI breaking changes

No changes. It's all internal.

Rollback safety impact

n/a

Expected complexity level and risk

We need to make sure BSATN decoding stays consistent for all paths (I added a unit test for this) and that websockets behavior stays the same as before (just less memory hungry).

Testing

Run benchmarks
Tests all still pass

Benchmarks

I added two benchmarks:

xUnit Benchmarks:

Benchmark	Rows / Variant	Master mean	PR mean	Time delta	Master alloc	PR alloc	Alloc delta
DecompressDecodeMessage transaction_update	1, uncompressed	0.49 us	0.41 us	-16.3%	1,544 B	1,200 B	-344 B
DecompressDecodeMessage transaction_update	100, uncompressed	2.87 us	2.74 us	-4.5%	2,128 B	1,592 B	-536 B
DecompressDecodeMessage transaction_update	10,000, uncompressed	89.01 us	85.28 us	-4.2%	81,328 B	41,192 B	-40,136 B
DecompressDecodeMessage transaction_update	1, brotli	2.22 us	1.46 us	-34.2%	67,232 B	1,672 B	-65,560 B
DecompressDecodeMessage transaction_update	100, brotli	5.36 us	4.61 us	-14.0%	67,816 B	2,256 B	-65,560 B
DecompressDecodeMessage transaction_update	10,000, brotli	197.16 us	195.17 us	-1.0%	147,017 B	81,457 B	-65,560 B
Reducer/procedure payload	1	0.23 us	0.25 us	+8.7%	936 B	904 B	-32 B
Reducer/procedure payload	100	4.55 us	4.61 us	+1.3%	5,840 B	5,408 B	-432 B
Reducer/procedure payload	10,000	233.89 us	300.59 us	+28.5%	571,663 B	531,630 B	-40,033 B
DecompressDecodeMessage initial_connection	uncompressed	0.13 us	0.09 us	-30.8%	728 B	384 B	-344 B
DecompressDecodeMessage initial_connection	brotli	1.46 us	0.45 us	-69.2%	66,416 B	856 B	-65,560 B
Brotli inflate transaction_update	1	1.13 us	1.13 us	0.0%	936 B	936 B	0 B
Brotli inflate transaction_update	100	1.98 us	1.96 us	-1.0%	1,128 B	1,128 B	0 B
Brotli inflate transaction_update	10,000	109.13 us	108.40 us	-0.7%	40,728 B	40,710 B	-18 B
MultiDictionary.Apply	1	0.15 us	0.15 us	0.0%	872 B	872 B	0 B
MultiDictionary.Apply	100	8.42 us	8.41 us	-0.1%	43,544 B	43,544 B	0 B
MultiDictionary.Apply	10,000	1,202.86 us	1,174.74 us	-2.3%	4,234,208 B	4,234,253 B	+45 B
ParseRowList	1	0.05 us	0.05 us	0.0%	224 B	224 B	0 B
ParseRowList	100	3.22 us	3.28 us	+1.9%	2,600 B	2,600 B	0 B
ParseRowList	10,000	244.97 us	276.34 us	+12.8%	240,200 B	240,200 B	0 B

BenchmarkDotNet:
I started with the xUnit benchmark from above and I liked the look of the allocation deltas a lot. But the reducer/procedure time delta didn't look so nice, so I also added this BenchmarkDotNet project that should be a little more accurate and the equivalent benchmark (the first three) now look a lot more flat.

Benchmark	Rows	Master mean	PR mean	Time delta	Master alloc	PR alloc	Alloc delta
XUnitReducerProcedureShape	1	103.57 ns	112.14 ns	+8.3%	936 B	904 B	-32 B
XUnitReducerProcedureShape	100	1.394 us	1.422 us	+2.0%	5,840 B	5,408 B	-432 B
XUnitReducerProcedureShape	10,000	130.733 us	131.045 us	+0.2%	571,663 B	531,631 B	-40,032 B
SerializeAndDecodeFromByteArray	1	85.93 ns	82.72 ns	-3.7%	840 B	840 B	0 B
SerializeAndDecodeFromByteArray	100	1.359 us	1.376 us	+1.3%	4,944 B	4,944 B	0 B
SerializeAndDecodeFromByteArray	10,000	124.907 us	133.552 us	+6.9%	491,566 B	491,566 B	0 B
SerializeToListOnly	1	56.57 ns	57.33 ns	+1.3%	536 B	536 B	0 B
SerializeToListOnly	100	830.12 ns	845.00 ns	+1.8%	1,872 B	1,872 B	0 B
SerializeToListOnly	10,000	75.081 us	77.035 us	+2.6%	211,288 B	211,288 B	0 B
DecodeFromListOnly	1	45.78 ns	53.04 ns	+15.9%	400 B	368 B	-32 B
DecodeFromListOnly	100	582.62 ns	584.81 ns	+0.4%	3,968 B	3,536 B	-432 B
DecodeFromListOnly	10,000	56.498 us	52.541 us	-7.0%	360,373 B	320,340 B	-40,033 B
DecodeFromByteArrayOnly	1	41.55 ns	41.80 ns	+0.6%	368 B	368 B	0 B
DecodeFromByteArrayOnly	100	565.05 ns	566.06 ns	+0.2%	3,536 B	3,536 B	0 B
DecodeFromByteArrayOnly	10,000	50.594 us	55.436 us	+9.6%	320,340 B	320,340 B	0 B

JasonAtClockwork

I ran into troubles running the PerformanceBenchmarks locally on Windows and WSL. I'll try next week on Linux and reach out if I still have troubles.

JasonAtClockwork · 2026-09-04T21:48:46Z

+  </PropertyGroup>
+
+  <ItemGroup>
+    <PackageReference Include="BenchmarkDotNet" Version="0.14.0" />


.\sdks\csharp\tests~\test.csproj is currently recursive in loading the tests but doesn't have access to BenchmarkDotNet which causes builds to fail. It might be best to just flag PerformanceBenchmarks as removed in .\sdks\csharp\tests~\test.csproj, such as:

<ItemGroup> <Compile Remove="PerformanceBenchmarks/**/*.cs" /> </ItemGroup>

JasonAtClockwork · 2026-09-04T21:50:16Z

                        var reducedMultiplicity = (int)my.Multiplicity + before.Delta;
                        if (reducedMultiplicity != 0)
                        {
                            PseudoThrow($"Attempted to apply {their} to {my}, but this resulted in a multiplicity of {reducedMultiplicity}, failing to correctly remove the row before applying the update");


It looks like marking PseudoThrow as readonly or static should be fine and clear up all the warnings.

lisandroct marked this pull request as draft August 28, 2026 18:29

lisandroct force-pushed the lisandro/sdk-performance branch from 7b86fbb to 4739cb5 Compare August 31, 2026 14:29

lisandroct marked this pull request as ready for review September 1, 2026 19:23

lisandroct requested a review from JasonAtClockwork September 3, 2026 13:28

lisandroct added 9 commits September 4, 2026 09:33

Add performance test

18b6773

Remove LINQ usage

616d717

Remove unnecessary allocations with .ToArray()

ec9d5a2

Start websockets with smaller buffers and grow only if needed

a6d106d

Improve tests

1d1391a

Use pooled arrays when decoding BSATN-serializable types from lists

544e1e1

Increase amount of iterations (to remove noise)

a95fb96

Add PerformanceBenchmarks.csproj

670bea5

Clean SpacetimeDBClient error decoding

a4d089f

lisandroct force-pushed the lisandro/sdk-performance branch from 62fb4b7 to a4d089f Compare September 4, 2026 14:33

JasonAtClockwork reviewed Sep 4, 2026

View reviewed changes

Navigation Menu

Uh oh!

SDK Performance Boost - #5840

SDK Performance Boost#5840
lisandroct wants to merge 9 commits into
masterfrom
lisandro/sdk-performance

lisandroct commented Aug 28, 2026 •

edited

Loading

Uh oh!

JasonAtClockwork left a comment

Uh oh!

JasonAtClockwork Sep 4, 2026

Uh oh!

JasonAtClockwork Sep 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

lisandroct commented Aug 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description of Changes

API and ABI breaking changes

Rollback safety impact

Expected complexity level and risk

Testing

Benchmarks

Uh oh!

JasonAtClockwork left a comment

Choose a reason for hiding this comment

Uh oh!

JasonAtClockwork Sep 4, 2026

Choose a reason for hiding this comment

Uh oh!

JasonAtClockwork Sep 4, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

lisandroct commented Aug 28, 2026 •

edited

Loading