Benchmark Results

How Dekaf compares to Confluent.Kafka, measured with BenchmarkDotNet in daily or on-demand GitHub Actions runs from main.

Last Updated: 2026-08-16 14:42 UTC

At a glance

Each scenario is the median Dekaf-vs-Confluent result over the last 10 benchmark runs (both clients measured on the same runner), aggregated across its parameter configurations. Memory compares heap allocations per operation from the latest run.

The charts use the representative high-load parameter set from the latest run and show its measured values. Multipliers in brackets are calculated from those displayed figures; the table below summarizes the broader cross-run range.

Reading producer results

The linger=0 scenario is the matched client comparison. The linger=5 ms scenario intentionally measures each client's app-limited batching policy and should not be read as general producer throughput. Legacy serial-awaited rows, when present, send a sole Dekaf record immediately while Confluent applies the configured linger; the old benchmark's unused BatchSize parameter also duplicated each payload result. Legacy batch and fire-and-forget rows, when present, use acks=all with idempotence for Dekaf and acks=leader without idempotence for Confluent.

Latest run · representative high-load caseExecution timeMeasured time per benchmark operation; shorter bars are better.

DekafConfluentLower is better

Produce — serial awaited (linger=5 ms; legacy)1000 B · app-limited

Dekaf116.51 μs (45× faster)

Confluent5.28 ms

Produce — batches (legacy: Dekaf acks=all/idempotent; Confluent acks=leader/non-idempotent)1000 B · batch 1000

Dekaf9.49 ms (1.2× slower)

Confluent8.02 ms

Produce — fire-and-forget (legacy: Dekaf acks=all/idempotent; Confluent acks=leader/non-idempotent)1000 B · batch 1000

Dekaf7.10 ms (1.3× slower)

Confluent5.32 ms

Consume — drain a topic1000 B · 1000 messages

Dekaf905.32 μs (1.3× slower)

Confluent683.96 μs

Consume — poll a single message1000 B

Dekaf613.1 ns (2.5× faster)

Confluent1.54 μs

Bars are scaled within each scenario for a direct client-to-client comparison.

Latest run · bytes per operationManaged allocationsManaged heap bytes allocated per operation; shorter bars are better.

DekafConfluentLower is better

Produce — serial awaited (linger=5 ms; legacy)1000 B · app-limited

Dekaf456 B (4.6× less)

Confluent2.06 KB

Produce — batches (legacy: Dekaf acks=all/idempotent; Confluent acks=leader/non-idempotent)1000 B · batch 1000

Dekaf50.51 KB (38× less)

Confluent1.85 MB

Produce — fire-and-forget (legacy: Dekaf acks=all/idempotent; Confluent acks=leader/non-idempotent)1000 B · batch 1000

Dekaf1.14 KB (1031× less)

Confluent1.15 MB

Consume — drain a topic1000 B · 1000 messages

Dekaf1.97 MB (1.2× less)

Confluent2.35 MB

Consume — poll a single message1000 B

Dekaf2.03 KB (1.2× less)

Confluent2.40 KB

Bars are scaled within each scenario for a direct client-to-client comparison.

Scenario	Speed vs Confluent	Memory vs Confluent	Confidence
Produce — serial awaited (linger=5 ms; legacy)	21×–22× faster	3.3× less	⚠ Noisy
Produce — batches (legacy: Dekaf acks=all/idempotent; Confluent acks=leader/non-idempotent)	on par to 2.4× faster	25× less	Mixed
Produce — fire-and-forget (legacy: Dekaf acks=all/idempotent; Confluent acks=leader/non-idempotent)	on par	1000× less	Mixed
Consume — drain a topic	1.8× slower to 1.2× faster	1.6× less	Mixed
Consume — poll a single message	3.7×–9.8× faster	1.6× less	⚠ Noisy

"On par" means within ±20% — differences that small are runner noise. A range means the result depends on message or batch size; the per-parameter tables below have the detail.

Full results

Cross-run comparison — last 10 runs, per parameter set

Each ratio pairs Dekaf and Confluent means from the same runner, then reports the median across recent comparable runs. Lower is better; < 1.0 means Dekaf is faster.

Rows with run spread above 30% are marked low-confidence. Run spread is (maximum ratio - minimum ratio) / median ratio.

Benchmark	Parameters	Runs	Median Ratio	Ratio Range	Run Spread	Confidence
ConsumerBenchmarks.ConsumeAll	MessageCount: 100, MessageSize: 100	10	1.14	1.01–1.29	24%	Stable
ConsumerBenchmarks.ConsumeAll	MessageCount: 100, MessageSize: 1000	10	1.30	1.09–1.51	32%	⚠ Low
ConsumerBenchmarks.ConsumeAll	MessageCount: 1000, MessageSize: 100	10	0.81	0.65–0.97	39%	⚠ Low
ConsumerBenchmarks.ConsumeAll	MessageCount: 1000, MessageSize: 1000	10	1.79	1.17–2.40	69%	⚠ Low
ConsumerPollBenchmarks.PollSingle	PollsPerIteration: 400000, MessageSize: 100	10	0.10	0.05–0.11	56%	⚠ Low
ConsumerPollBenchmarks.PollSingle	PollsPerIteration: 400000, MessageSize: 1000	10	0.27	0.26–0.70	160%	⚠ Low
ProducerBenchmarks.FireAndForget	MessageSize: 100, BatchSize: 100	10	0.93	0.81–1.48	72%	⚠ Low
ProducerBenchmarks.FireAndForget	MessageSize: 100, BatchSize: 1000	10	0.98	0.89–1.47	59%	⚠ Low
ProducerBenchmarks.FireAndForget	MessageSize: 1000, BatchSize: 100	10	0.86	0.74–0.98	28%	Stable
ProducerBenchmarks.FireAndForget	MessageSize: 1000, BatchSize: 1000	10	0.83	0.75–1.31	67%	⚠ Low
ProducerBenchmarks.ProduceBatch	MessageSize: 100, BatchSize: 100	10	0.44	0.43–0.44	4%	Stable
ProducerBenchmarks.ProduceBatch	MessageSize: 100, BatchSize: 1000	10	0.50	0.48–0.51	6%	Stable
ProducerBenchmarks.ProduceBatch	MessageSize: 1000, BatchSize: 100	10	0.42	0.40–0.53	31%	⚠ Low
ProducerBenchmarks.ProduceBatch	MessageSize: 1000, BatchSize: 1000	10	1.09	1.01–1.28	25%	Stable
ProducerBenchmarks.ProduceSingle	MessageSize: 100, BatchSize: 100	10	0.05	0.02–0.06	82%	⚠ Low
ProducerBenchmarks.ProduceSingle	MessageSize: 100, BatchSize: 1000	10	0.05	0.02–0.06	81%	⚠ Low
ProducerBenchmarks.ProduceSingle	MessageSize: 1000, BatchSize: 100	10	0.05	0.02–0.06	83%	⚠ Low
ProducerBenchmarks.ProduceSingle	MessageSize: 1000, BatchSize: 1000	10	0.05	0.02–0.06	81%	⚠ Low

Latest run — producer benchmarks

Method	Categories	MessageSize	BatchSize	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Gen1	Allocated	Alloc Ratio	Confidence
Confluent_ProduceBatch	BatchProduce	100	100	5,481.38 μs	65.240 μs	43.152 μs	1.00	0.01	-	-	105170 B	1.00	Stable
Dekaf_ProduceBatch	BatchProduce	100	100	2,344.27 μs	12.635 μs	7.519 μs	0.43	0.00	-	-	5344 B	0.05	Stable
													—
Confluent_ProduceBatch	BatchProduce	100	1000	6,384.55 μs	83.715 μs	49.817 μs	1.00	0.01	62.5000	23.4375	1048384 B	1.00	Stable
Dekaf_ProduceBatch	BatchProduce	100	1000	3,224.80 μs	120.425 μs	79.654 μs	0.51	0.01	-	-	49812 B	0.05	Stable
													—
Confluent_ProduceBatch	BatchProduce	1000	100	5,865.54 μs	6.133 μs	3.208 μs	1.00	0.00	7.8125	-	194772 B	1.00	Stable
Dekaf_ProduceBatch	BatchProduce	1000	100	2,441.00 μs	47.392 μs	31.347 μs	0.42	0.01	-	-	6307 B	0.03	Stable
													—
Confluent_ProduceBatch	BatchProduce	1000	1000	8,016.78 μs	2,456.682 μs	1,461.932 μs	1.03	0.23	109.3750	46.8750	1944394 B	1.00	Stable
Dekaf_ProduceBatch	BatchProduce	1000	1000	9,488.78 μs	169.661 μs	100.962 μs	1.21	0.18	-	-	51720 B	0.03	Stable
													—
Confluent_FireAndForget	FireAndForget	100	100	58.71 μs	2.676 μs	1.770 μs	1.00	0.04	1.7700	-	30400 B	1.000	Stable
Dekaf_FireAndForget	FireAndForget	100	100	85.42 μs	20.084 μs	13.284 μs	1.46	0.22	-	-	24 B	0.001	Stable
													—
Confluent_FireAndForget	FireAndForget	100	1000	570.86 μs	33.310 μs	22.033 μs	1.00	0.05	17.5781	-	304000 B	1.000	Stable
Dekaf_FireAndForget	FireAndForget	100	1000	777.48 μs	280.071 μs	185.250 μs	1.36	0.31	-	-	298 B	0.001	⚠ Low
													—
Confluent_FireAndForget	FireAndForget	1000	100	552.50 μs	117.556 μs	61.484 μs	1.01	0.15	7.2021	-	120896 B	1.000	Stable
Dekaf_FireAndForget	FireAndForget	1000	100	523.16 μs	272.069 μs	161.904 μs	0.96	0.30	-	-	237 B	0.002	Stable
													—
Confluent_FireAndForget	FireAndForget	1000	1000	5,321.78 μs	647.157 μs	428.054 μs	1.01	0.11	70.3125	-	1206138 B	1.000	Stable
Dekaf_FireAndForget	FireAndForget	1000	1000	7,098.98 μs	2,794.574 μs	1,663.006 μs	1.34	0.32	-	-	1170 B	0.001	⚠ Low
													—
Confluent_ProduceSingle	SingleProduce	100	100	5,271.37 μs	8.384 μs	4.385 μs	1.00	0.00	-	-	1202 B	1.00	Stable
Dekaf_ProduceSingle	SingleProduce	100	100	116.84 μs	4.220 μs	2.791 μs	0.02	0.00	-	-	456 B	0.38	Stable
													—
Confluent_ProduceSingle	SingleProduce	100	1000	5,290.59 μs	52.814 μs	27.623 μs	1.00	0.01	-	-	1202 B	1.00	Stable
Dekaf_ProduceSingle	SingleProduce	100	1000	115.97 μs	1.139 μs	0.753 μs	0.02	0.00	-	-	456 B	0.38	Stable
													—
Confluent_ProduceSingle	SingleProduce	1000	100	5,276.22 μs	8.543 μs	5.651 μs	1.00	0.00	-	-	2113 B	1.00	Stable
Dekaf_ProduceSingle	SingleProduce	1000	100	118.39 μs	2.593 μs	1.543 μs	0.02	0.00	-	-	456 B	0.22	Stable
													—
Confluent_ProduceSingle	SingleProduce	1000	1000	5,277.37 μs	5.620 μs	3.344 μs	1.00	0.00	-	-	2113 B	1.00	Stable
Dekaf_ProduceSingle	SingleProduce	1000	1000	116.51 μs	1.887 μs	1.248 μs	0.02	0.00	-	-	456 B	0.22	Stable

Latest run — consumer benchmarks

Method	MessageCount	MessageSize	Mean	Error	StdDev	Median	Ratio	RatioSD	Allocated	Alloc Ratio	Confidence
Confluent_ConsumeAll	100	100	70.61 μs	29.938 μs	15.658 μs	62.65 μs	1.04	0.28	64.99 KB	1.00	Stable
Dekaf_ConsumeAll	100	100	63.39 μs	2.645 μs	1.384 μs	63.77 μs	0.93	0.16	26.45 KB	0.41	Stable
											—
Confluent_ConsumeAll	100	1000	68.54 μs	3.923 μs	1.399 μs	68.52 μs	1.00	0.03	240.77 KB	1.00	Stable
Dekaf_ConsumeAll	100	1000	79.49 μs	3.115 μs	1.383 μs	79.28 μs	1.16	0.03	202.23 KB	0.84	Stable
											—
Confluent_ConsumeAll	1000	100	609.14 μs	316.950 μs	165.771 μs	563.88 μs	1.06	0.36	648.59 KB	1.00	⚠ Low
Dekaf_ConsumeAll	1000	100	378.15 μs	63.777 μs	28.317 μs	367.30 μs	0.66	0.15	258.48 KB	0.40	Stable
											—
Confluent_ConsumeAll	1000	1000	683.96 μs	385.912 μs	171.347 μs	577.48 μs	1.05	0.33	2406.4 KB	1.00	⚠ Low
Dekaf_ConsumeAll	1000	1000	905.32 μs	954.022 μs	423.592 μs	1,199.00 μs	1.39	0.68	2016.3 KB	0.84	⚠ Low

Method	PollsPerIteration	MessageSize	Mean	Error	StdDev	Median	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio	Confidence
Confluent_PollSingle	400000	100	5,319.8 ns	15.39 ns	8.05 ns	5,317.6 ns	1.00	0.00	0.0375	654 B	1.00	Stable
Dekaf_PollSingle	400000	100	279.5 ns	25.09 ns	13.12 ns	274.6 ns	0.05	0.00	0.0150	270 B	0.41	Stable
												—
Confluent_PollSingle	400000	1000	1,536.3 ns	1,709.02 ns	1,130.41 ns	794.4 ns	1.46	1.32	0.1450	2454 B	1.00	⚠ Low
Dekaf_PollSingle	400000	1000	613.1 ns	205.59 ns	135.98 ns	552.5 ns	0.58	0.30	0.1225	2074 B	0.85	Stable

Protocol serialization — Dekaf internals

Wire protocol serialization/deserialization. Allocated = - means zero heap allocations — the goal of Dekaf's design.

Method	Mean	Error	StdDev	Gen0	Allocated
ReadDescribeGroupsV5	580.13 ns	13.871 ns	3.602 ns	0.0725	1224 B
WriteFindCoordinatorV6	27.57 ns	0.253 ns	0.039 ns	-	-
WriteDescribeGroupsV6	45.52 ns	0.245 ns	0.038 ns	-	-
WriteListConfigResourcesV1	20.78 ns	0.138 ns	0.036 ns	-	-

Method	Version	Mean	Error	StdDev	Allocated
WriteRequest	0	2.004 μs	0.0229 μs	0.0035 μs	-
WriteRequest	1	2.002 μs	0.0019 μs	0.0005 μs	-

Method	Version	Mean	Error	StdDev	Allocated
WriteRequest	8	2.395 μs	0.0058 μs	0.0009 μs	-
WriteRequest	9	2.604 μs	0.0548 μs	0.0085 μs	-
WriteRequest	10	2.413 μs	0.0117 μs	0.0030 μs	-
WriteRequest	11	2.411 μs	0.0185 μs	0.0048 μs	-

Method	Version	Mean	Error	StdDev	Allocated
WriteOffsetCommitRequest	9	103.10 ns	0.465 ns	0.072 ns	-
WriteOffsetFetchRequest	9	100.43 ns	0.586 ns	0.152 ns	-
WriteOffsetCommitRequest	10	96.36 ns	0.903 ns	0.140 ns	-
WriteOffsetFetchRequest	10	87.07 ns	0.505 ns	0.078 ns	-

Method	Mean	Error	StdDev	Gen0	Allocated
'Write 1000 Int32s'	1,640.0 ns	2.93 ns	1.75 ns	-	-
'Write 100 Strings (100 chars)'	2,059.3 ns	2.87 ns	1.71 ns	-	-
'Write 100 Strings (300 chars)'	2,310.4 ns	4.74 ns	2.82 ns	-	-
'Write 100 String spans (300 chars)'	2,284.1 ns	3.77 ns	2.24 ns	-	-
'Write 100 CompactStrings'	2,028.6 ns	2.50 ns	1.49 ns	-	-
'Write 100 CompactStrings (300 chars)'	3,891.2 ns	10.47 ns	6.23 ns	-	-
'Write 100 CompactString spans (300 chars)'	3,903.7 ns	2.65 ns	1.39 ns	-	-
'Write 1000 VarInts'	2,735.5 ns	6.56 ns	4.34 ns	-	-
'Read 1000 Int32s'	1,116.7 ns	0.98 ns	0.58 ns	-	-
'Read 1000 VarInts'	1,814.0 ns	2.47 ns	1.47 ns	-	-
'Write RecordBatch (10 records)'	774.0 ns	1.58 ns	1.05 ns	-	-
'Write RecordBatch pre-serialized (10 records)'	837.2 ns	8.38 ns	4.99 ns	-	-
'Read RecordBatch (10 records)'	144.2 ns	0.48 ns	0.25 ns	-	-
'Read Gzip RecordBatch (10 records)'	1,734.8 ns	11.83 ns	7.82 ns	0.0172	312 B
'Read + Iterate RecordBatch (10 records)'	1,295.9 ns	5.84 ns	3.48 ns	-	-

Serializers — Dekaf internals

Method	Mean	Error	StdDev	Ratio	RatioSD	Allocated	Alloc Ratio
'Prepare stable generic Avro schema'	3.646 ns	0.0066 ns	0.0058 ns	1.00	0.00	-	NA
'Prepare equivalent generic Avro schema instance'	234.849 ns	0.1540 ns	0.1366 ns	64.41	0.11	-	NA

Method	Mean	Error	StdDev	Allocated
SerializeCached	41.77 ns	1.091 ns	0.060 ns	-

Method	Categories	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
'Serialize 100 Messages (key+value)'	Batch	9,386.595 ns	3.7648 ns	1.9691 ns	?	?	-	-	?
'Serialize String (10 chars)'	Scalar	10.182 ns	0.0209 ns	0.0124 ns	?	?	-	-	?
'Serialize String (100 chars)'	Scalar	12.055 ns	0.0091 ns	0.0060 ns	?	?	-	-	?
'Serialize String (1000 chars)'	Scalar	26.905 ns	0.2238 ns	0.1332 ns	?	?	-	-	?
'Deserialize String'	Scalar	26.912 ns	0.4969 ns	0.3287 ns	?	?	0.0026	224 B	?
'Serialize Int32'	Scalar	7.142 ns	0.0227 ns	0.0135 ns	?	?	-	-	?
'ArrayBufferWriter + Copy'	Writer	91.927 ns	1.7350 ns	1.1476 ns	1.00	0.02	0.0106	896 B	1.00
'ReusableBufferWriter Direct'	Writer	50.440 ns	0.5176 ns	0.3080 ns	0.55	0.01	-	-	0.00

Compression — Dekaf internals

Method	Mean	Error	StdDev	Gen0	Allocated
'Snappy Compress 1KB'	287.1 ns	3.39 ns	2.24 ns	0.0029	48 B
'Snappy Compress 1MB'	97,120.8 ns	103.34 ns	61.49 ns	-	48 B
'Snappy Decompress 1KB'	221.0 ns	0.79 ns	0.52 ns	0.0048	80 B
'Snappy Decompress 1MB'	122,700.7 ns	95.01 ns	56.54 ns	-	80 B

How to read these tables

• Mean: Average execution time
• Error: Half of 99.9% confidence interval
• StdDev: Standard deviation of all measurements
• Ratio: Performance relative to that table's baseline row
  • Producer/Consumer tables: baseline is Confluent.Kafka, so < 1.0 = Dekaf is faster, > 1.0 = Confluent is faster
  • Dekaf-internals tables (Protocol/Serializer/Compression): baseline is an internal reference implementation, not Confluent
• RatioSD: BenchmarkDotNet's uncertainty for the latest run's ratio
• Confidence: ⚠ Low when latest RatioSD > 0.30 or rolling run spread exceeds 30%
• Allocated: Heap memory allocated per operation
  • - = Zero allocations (ideal!)

Benchmarks automatically run daily at 05:00 UTC from main and can also be run manually.