Kangaroo - Library

Description

The Kangaroo library is more than just a queue. It collects all data (exceptions, sensor data, ...) that should be collected. For each type a separate Kangaroo instance is created. Each instance has one or more export manager implementations (developed by user). These export manager can be extended by a converter (also developed by user). At the end it should standardize the collecting and exporting type. Updates on each export/converter implementations should have no impact in already implemented projects.

Installation

Add it to your .Net Framework or DotNet project. The library Kangaroo supports DotNet 10 and higher.

Benchmarks

The Kangaroo.Benchmarks project uses BenchmarkDotNet to measure Kangaroo's performance against equivalent approaches from the .NET BCL:

• TPL Dataflow (System.Threading.Tasks.Dataflow) – BufferBlock<T>, BatchBlock<T>, ActionBlock<T>, TransformBlock<T>, BroadcastBlock<T>
• System.Threading.Channels – raw Channel<T>.Writer.TryWrite
• System.Collections.Concurrent.ConcurrentQueue<T> – simple lock-free queue

Benchmarks were executed on:

BenchmarkDotNet v0.14.0, macOS 26.4.1 (25E253) [Darwin 25.4.0]
Apple M1 Pro, 1 CPU, 10 logical and 10 physical cores
.NET SDK 10.0.103  |  Runtime .NET 10.0.3 (Arm64 RyuJIT AdvSIMD)

Queueing throughput (QueueingBenchmarks)

Measures how fast N items can be added to each queue / buffer without any export being triggered. Kangaroo is the baseline (Ratio = 1.00).

Method	N	Mean	Ratio	Rank	Allocated
ConcurrentQueue Enqueue	100	1.782 µs	0.27	1	-
Channel TryWrite	100	3.690 µs	0.56	2	-
Dataflow BufferBlock.Post	100	4.457 µs	0.68	3	-
Kangaroo AddData	100	6.583 µs	1.00	4	-
ConcurrentQueue Enqueue	1 000	18.048 µs	0.28	1	34 720 B
Dataflow BufferBlock.Post	1 000	37.202 µs	0.58	2	19 200 B
Channel TryWrite	1 000	39.034 µs	0.61	2	34 720 B
Kangaroo AddData	1 000	64.700 µs	1.00	3	50 592 B
ConcurrentQueue Enqueue	10 000	122.434 µs	0.35	1	265 440 B
Channel TryWrite	10 000	185.497 µs	0.53	2	265 440 B
Dataflow BufferBlock.Post	10 000	327.192 µs	0.94	3	135 232 B
Kangaroo AddData	10 000	348.151 µs	1.00	4	396 000 B

Takeaway – The abstraction overhead of Kangaroo's AddData (category tagging, switch dispatch) costs roughly 2–4× compared to a raw Channel.TryWrite. This is the expected trade-off for a richer feature set.

---

Batch export pipeline (ExportPipelineBenchmarks)

Adds N items then drains and "exports" them all at once. Kangaroo manual export is the baseline.

Method	N	Mean	Ratio	Rank	Allocated
ConcurrentQueue manual drain	100	2.629 µs	0.80	1	6.48 KB
Kangaroo batch export	100	3.303 µs	1.00	2	11.52 KB
Kangaroo size-triggered export	100	4.131 µs	1.25	3	11.52 KB
Dataflow BatchBlock→ActionBlock	100	9.803 µs	2.97	4	7.03 KB
ConcurrentQueue manual drain	1 000	24.039 µs	0.84	1	49.85 KB
Kangaroo batch export	1 000	28.681 µs	1.00	2	75.94 KB
Dataflow BatchBlock→ActionBlock	1 000	30.447 µs	1.06	3	28.19 KB
Kangaroo size-triggered export	1 000	41.756 µs	1.46	4	75.94 KB

Takeaway – Kangaroo's synchronous batch export outperforms TPL Dataflow's BatchBlock + ActionBlock pipeline (3× faster for N=100, on par for N=1 000) because it avoids the async scheduling overhead. A bare ConcurrentQueue drain is faster still, but provides none of Kangaroo's categorisation, filtering, or converter features.

---

Direct / per-item export (DirectExportBenchmarks)

Each item is processed immediately (MaxStoredObjects = 1). Kangaroo direct export is the baseline.

Method	N	Mean	Ratio	Rank	Allocated
Plain delegate loop	100	426.0 ns	0.10	1	3.13 KB
Kangaroo direct export	100	4,391.7 ns	1.00	2	17.08 KB
Dataflow ActionBlock (sequential)	100	8,342.5 ns	1.90	3	3.36 KB
Plain delegate loop	1 000	4,202.8 ns	0.10	1	31.25 KB
Dataflow ActionBlock (sequential)	1 000	41,714.2 ns	0.99	2	15.82 KB
Kangaroo direct export	1 000	41,995.2 ns	1.00	2	143.64 KB

Takeaway – Kangaroo's synchronous per-item export is 1.9× faster than TPL Dataflow's sequential ActionBlock at N=100. At N=1 000 both are essentially neck-and-neck, while Kangaroo provides richer categorisation and converter features.

---

Burst export pipeline (BurstExportBenchmarks)

Simulates repeated burst ingestion: items arrive in batches of BatchSize, repeated BatchCount times. Kangaroo size-triggered burst is the baseline.

Method	BatchSize	BatchCount	Mean	Ratio	Rank	Allocated
ConcurrentQueue threshold drain	100	10	21.82 µs	0.72	1	21.24 KB
Channel threshold drain	100	10	22.05 µs	0.73	1	19.92 KB
Kangaroo manual burst export	100	10	27.02 µs	0.89	2	39.72 KB
Kangaroo size-triggered bursts	100	10	30.32 µs	1.00	3	39.72 KB
Dataflow BatchBlock bursts	100	10	35.07 µs	1.16	4	15.3 KB
ConcurrentQueue threshold drain	100	50	107.82 µs	0.74	1	86.87 KB
Channel threshold drain	100	50	109.42 µs	0.75	1	85.55 KB
Kangaroo manual burst export	100	50	131.80 µs	0.91	2	165.03 KB
Dataflow BatchBlock bursts	100	50	138.64 µs	0.95	3	51.44 KB
Kangaroo size-triggered bursts	100	50	145.33 µs	1.00	4	165.03 KB
Dataflow BatchBlock bursts	500	10	106.28 µs	0.73	1	52.74 KB
ConcurrentQueue threshold drain	500	10	108.66 µs	0.74	2	96.62 KB
Channel threshold drain	500	10	109.54 µs	0.75	2	89.3 KB
Kangaroo manual burst export	500	10	127.49 µs	0.87	3	150.31 KB
Kangaroo size-triggered bursts	500	10	145.97 µs	1.00	4	150.31 KB
Dataflow BatchBlock bursts	500	50	494.41 µs	0.69	1	214.92 KB
ConcurrentQueue threshold drain	500	50	539.66 µs	0.76	2	412.24 KB
Channel threshold drain	500	50	550.39 µs	0.77	2	404.92 KB
Kangaroo manual burst export	500	50	644.53 µs	0.90	3	642.5 KB
Kangaroo size-triggered bursts	500	50	712.62 µs	1.00	4	642.5 KB

Takeaway – Kangaroo's manual burst export stays within 0.87–0.91× of the automatic size-triggered baseline and consistently outperforms Dataflow's BatchBlock at small batch sizes. At larger batches (500×50), Dataflow's lower per-item allocation gives it an edge in raw throughput.

---

Converter pipeline (ConverterPipelineBenchmarks)

Adds N items, then filters and converts them during export. Kangaroo filter+convert pipeline is the baseline.

Method	N	Mean	Ratio	Rank	Allocated
Channel manual filter+convert	1 000	23.39 µs	0.71	1	34.01 KB
ConcurrentQueue manual filter+convert	1 000	24.21 µs	0.73	2	41.45 KB
Kangaroo filter+convert pipeline	1 000	33.04 µs	1.00	3	79.2 KB
Kangaroo filter+convert direct	1 000	98.20 µs	2.97	4	396.83 KB
Dataflow TransformBlock pipeline	1 000	107.50 µs	3.25	5	74.35 KB
ConcurrentQueue manual filter+convert	10 000	242.16 µs	0.65	1	450.64 KB
Channel manual filter+convert	10 000	246.52 µs	0.66	1	387.2 KB
Kangaroo filter+convert pipeline	10 000	371.46 µs	1.00	2	747.64 KB
Dataflow TransformBlock pipeline	10 000	793.77 µs	2.14	3	650.69 KB
Kangaroo filter+convert direct	10 000	968.92 µs	2.61	4	3 982.77 KB

Takeaway – Kangaroo's batch filter+convert pipeline is 2–3× faster than Dataflow's TransformBlock chain, thanks to avoiding async scheduling overhead. Manual implementations with raw queues are faster still but lack Kangaroo's built-in categorisation and converter infrastructure.

---

Category routing (CategoryRoutingBenchmarks)

Routes N items to category-specific exporters. Kangaroo category routing + export is the baseline.

Method	N	Mean	Ratio	Rank	Allocated
ConcurrentQueue dictionary routing	1 000	22.18 µs	0.52	1	62.31 KB
Channel dictionary routing	1 000	27.64 µs	0.65	2	48.96 KB
Kangaroo category routing + export	1 000	42.51 µs	1.00	3	100.5 KB
Kangaroo category direct export	1 000	96.16 µs	2.26	4	167.29 KB
Dataflow BroadcastBlock predicate routing	1 000	760.91 µs	17.90	5	94.29 KB
ConcurrentQueue dictionary routing	10 000	209.27 µs	0.40	1	602.27 KB
Channel dictionary routing	10 000	271.54 µs	0.52	2	476.92 KB
Kangaroo category routing + export	10 000	526.26 µs	1.00	3	959.54 KB
Kangaroo category direct export	10 000	960.30 µs	1.82	4	1 643.91 KB
Dataflow BroadcastBlock predicate routing	10 000	7,976.20 µs	15.16	5	1 069.64 KB

Takeaway – Kangaroo's category routing is 15–18× faster than Dataflow's BroadcastBlock with predicate filtering. The BroadcastBlock approach copies every item to every target and filters afterwards, whereas Kangaroo dispatches directly to the correct exporter.

---

Concurrent producers (ConcurrentProducerBenchmarks)

Multiple threads write ItemsPerProducer items concurrently. Kangaroo concurrent AddData is the baseline.

Method	Producers	Items/Producer	Mean	Ratio	Rank	Allocated
Channel concurrent TryWrite	4	1 000	127.0 µs	0.98	1	69.6 KB
Kangaroo concurrent AddData	4	1 000	130.1 µs	1.00	1	102.67 KB
Dataflow BufferBlock concurrent Post	4	1 000	133.1 µs	1.02	1	38.14 KB
ConcurrentQueue concurrent Enqueue	4	1 000	268.7 µs	2.07	2	69.18 KB
Kangaroo concurrent AddData	4	10 000	1,179.0 µs	1.00	1	1 550.37 KB
Channel concurrent TryWrite	4	10 000	1,289.9 µs	1.09	2	1 031.67 KB
Dataflow BufferBlock concurrent Post	4	10 000	1,362.9 µs	1.16	3	520.11 KB
ConcurrentQueue concurrent Enqueue	4	10 000	3,097.9 µs	2.63	4	1 031.25 KB
Channel concurrent TryWrite	8	1 000	279.1 µs	0.94	1	134.91 KB
Kangaroo concurrent AddData	8	1 000	297.7 µs	1.00	2	200.01 KB
Dataflow BufferBlock concurrent Post	8	1 000	306.3 µs	1.03	2	71.44 KB
ConcurrentQueue concurrent Enqueue	8	1 000	1,881.9 µs	6.32	3	134.49 KB
Channel concurrent TryWrite	8	10 000	2,234.6 µs	0.96	1	2 060.57 KB
Kangaroo concurrent AddData	8	10 000	2,325.3 µs	1.00	2	3 086.93 KB
Dataflow BufferBlock concurrent Post	8	10 000	2,593.8 µs	1.12	3	1 035.09 KB
ConcurrentQueue concurrent Enqueue	8	10 000	20,059.1 µs	8.63	4	2 057.06 KB

Takeaway – Under contention (4–8 producers), Kangaroo matches Channel and BufferBlock throughput while dramatically outperforming ConcurrentQueue, which suffers from lock contention at higher thread counts (6–9× slower).

---

Multi-exporter fan-out (MultiExporterFanOutBenchmarks)

Each item is delivered to multiple exporters (fan-out). Kangaroo multi-exporter fan-out is the baseline.

Method	Exporters	N	Mean	Ratio	Rank	Allocated
ConcurrentQueue sequential fan-out	2	1 000	23.86 µs	0.83	1	49.85 KB
Kangaroo multi-exporter fan-out	2	1 000	28.87 µs	1.00	2	83.8 KB
Channel cloned fan-out	2	1 000	38.81 µs	1.34	3	37.74 KB
Kangaroo multi-exporter direct	2	1 000	53.80 µs	1.86	4	143.66 KB
Dataflow BroadcastBlock fan-out	2	1 000	777.22 µs	26.93	5	65.75 KB
ConcurrentQueue sequential fan-out	2	10 000	248.06 µs	0.79	1	415.86 KB
Kangaroo multi-exporter fan-out	2	10 000	312.48 µs	1.00	2	801.92 KB
Channel cloned fan-out	2	10 000	375.73 µs	1.20	3	264.37 KB
Kangaroo multi-exporter direct	2	10 000	537.32 µs	1.72	4	1 409.29 KB
Dataflow BroadcastBlock fan-out	2	10 000	7,329.01 µs	23.45	5	692.57 KB
ConcurrentQueue sequential fan-out	5	1 000	23.94 µs	0.76	1	49.85 KB
Kangaroo multi-exporter fan-out	5	1 000	31.72 µs	1.00	2	107.38 KB
Kangaroo multi-exporter direct	5	1 000	83.17 µs	2.63	3	143.73 KB
Channel cloned fan-out	5	1 000	95.11 µs	3.01	4	94.32 KB
Dataflow BroadcastBlock fan-out	5	1 000	3,647.36 µs	115.25	5	299.97 KB
ConcurrentQueue sequential fan-out	5	10 000	248.46 µs	0.77	1	415.86 KB
Kangaroo multi-exporter fan-out	5	10 000	323.00 µs	1.00	2	1 036.43 KB
Kangaroo multi-exporter direct	5	10 000	817.64 µs	2.53	3	1 409.36 KB
Channel cloned fan-out	5	10 000	939.70 µs	2.91	4	660.88 KB
Dataflow BroadcastBlock fan-out	5	10 000	36,404.35 µs	112.71	5	3 082.89 KB

Takeaway – Kangaroo is 23–115× faster than Dataflow's BroadcastBlock fan-out, delivering the largest relative advantage in the entire benchmark suite. As the exporter count grows from 2 to 5, Kangaroo's advantage widens because BroadcastBlock duplicates work per target while Kangaroo iterates exporters directly.

---

Conclusion

Across 102 benchmarks covering 8 distinct scenarios, Kangaroo demonstrates competitive performance against raw .NET primitives while providing significantly richer functionality (categorisation, converters, multi-export, size-triggered batching).

Scenario	vs. Dataflow	vs. Raw Primitives
Queueing throughput	~1.1× faster	~2–4× slower (expected)
Batch export	~3× faster (N=100)	~1.2× slower
Direct per-item export	~1.9× faster (N=100)	~10× slower vs. plain loop
Burst export	~1.2× faster (small batch)	~1.1–1.4× slower
Converter pipeline	~2–3× faster	~1.4–1.5× slower
Category routing	~15–18× faster	~1.9–2.5× slower
Concurrent producers	on par	1× (matches Channel)
Multi-exporter fan-out	~23–115× faster	~1.2–1.3× slower

The overall pattern: Kangaroo trades modest raw queueing throughput for a richer feature set, but recovers that cost and more in every end-to-end pipeline benchmark where those features matter. In the scenarios that most closely resemble real-world usage — batch export, converter pipelines, category routing, concurrent ingestion, and multi-exporter fan-out — Kangaroo matches or significantly outperforms the TPL Dataflow equivalent.

---

Running the benchmarks yourself

dotnet run --project Kangaroo.Benchmarks -c Release -- --filter "Kangaroo.Benchmarks*" --job short

Full BenchmarkDotNet artifacts (HTML, CSV, JSON) are written to BenchmarkDotNet.Artifacts/results/.

Contribution

Create a fork of the project into your own reposity. Make all your necessary changes and create a pull request with a description on what was added or removed and details explaining the changes in lines of code. If approved, project owners will merge it.

Licensing

This project is under MIT-Licence (see LICENSE file).

Support

Please file bugs and issues on the Github issues page for this project. This is to help keep track and document everything related to this repo. The code and documentation are released with no warranties or SLAs and are intended to be supported through a community driven process.