Stress Test Results

Long-running stress tests comparing sustained performance between Dekaf and Confluent.Kafka under real-world load.

Last Updated: 2026-09-06 03:46 UTC

info

The paired Dekaf vs Confluent comparison runs on demand (full_run=true from main) and updates this page. Manual dispatches stay Dekaf-only unless full_run explicitly requests the same paired publish path. Tests measure sustained performance over 15+ minutes with real Kafka instances.

At a glance

Each row is a like-for-like comparison: both clients run the same sustained workload sequentially on the same VM, and repeated samples are aggregated with a geometric mean across both run orders.

Paired same-VM stress runSustained throughputBroker-confirmed messages per second for the same workload.

DekafConfluentHigher is better

Produce — fire-and-forget

Dekaf1.45M msg/s (1.5×)

Confluent995.25K msg/s

Produce — fire-and-forget (3 brokers)

Dekaf1.17M msg/s (1.6×)

Confluent713.58K msg/s

Produce — acks=all

Dekaf1.54M msg/s (1.3×)

Confluent1.23M msg/s

Produce — acks=all (3 brokers)

Dekaf1.13M msg/s (1.5×)

Confluent751.18K msg/s

Produce — fire-and-forget, idempotent

Dekaf1.49M msg/s (1.2×)

Confluent1.26M msg/s

Produce — fire-and-forget, idempotent (3 brokers)

Dekaf1.12M msg/s (1.4×)

Confluent803.60K msg/s

Produce + consume round-trip

Dekaf2.63M msg/s (1.6×)

Confluent1.64M msg/s

Produce — transactional (exactly-once) (3 brokers)

Dekaf1.19K msg/s (7.1×)

Confluent169 msg/s

Consume — messages

Dekaf1.75M msg/s (1.3×)

Confluent1.33M msg/s

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

Median client CPU timeCPU cost per messageCPU time needed to deliver one message; shorter bars are better.

DekafConfluentLower is better

Produce — fire-and-forget

Dekaf0.72 μs/msg (2.4× less)

Confluent1.72 μs/msg

Produce — fire-and-forget (3 brokers)

Dekaf1.07 μs/msg (2.0× less)

Confluent2.11 μs/msg

Produce — acks=all

Dekaf0.70 μs/msg (2.0× less)

Confluent1.42 μs/msg

Produce — acks=all (3 brokers)

Dekaf0.95 μs/msg (2.2× less)

Confluent2.05 μs/msg

Produce — fire-and-forget, idempotent

Dekaf0.69 μs/msg (2.0× less)

Confluent1.41 μs/msg

Produce — fire-and-forget, idempotent (3 brokers)

Dekaf0.88 μs/msg (2.2× less)

Confluent1.97 μs/msg

Produce + consume round-trip

Dekaf0.88 μs/msg (2.1× less)

Confluent1.85 μs/msg

Produce — transactional (exactly-once) (3 brokers)

Dekaf228.84 μs/msg (1.1× less)

Confluent258.82 μs/msg

Consume — messages

Dekaf0.76 μs/msg (1.5× less)

Confluent1.14 μs/msg

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

Scenario	Dekaf	Confluent	Throughput	CPU per message
Produce — fire-and-forget	1,446,139 msg/s	995,249 msg/s	1.5× faster	2.4× less
Produce — fire-and-forget (3 brokers)	1,165,881 msg/s	713,584 msg/s	1.6× faster	2.0× less
Produce — acks=all	1,540,523 msg/s	1,228,208 msg/s	1.3× faster	2.0× less
Produce — acks=all (3 brokers)	1,129,651 msg/s	751,176 msg/s	1.5× faster	2.2× less
Produce — fire-and-forget, idempotent	1,490,271 msg/s	1,263,552 msg/s	1.2× faster	2.0× less
Produce — fire-and-forget, idempotent (3 brokers)	1,116,744 msg/s	803,597 msg/s	1.4× faster	2.2× less
Produce + consume round-trip	2,625,535 msg/s	1,641,424 msg/s	1.6× faster	2.1× less
Produce — transactional (exactly-once) (3 brokers)	1,193 msg/s	169 msg/s	7.1× faster	1.1× less
Consume — messages	1,751,563 msg/s	1,329,577 msg/s	1.3× faster	1.5× less
Consume — batches	1,732,327 msg/s	—	—	—
Consume — raw bytes	3,741,807 msg/s	—	—	—
Consume — raw byte batches	4,094,242 msg/s	—	—	—

"On par" means within ±5% — differences that small are run-to-run noise. "CPU per message" compares the client CPU cost of delivering one message; "less" means Dekaf needs less CPU. Rows showing "—" have no Confluent counterpart in this run (for example, batch and raw consume APIs that librdkafka does not expose). The full per-run data is below.

Full results

Each section holds the measured per-run data behind the summary: repeated same-VM samples in both client orders, CPU per message and per request, and throughput drift across the run.

Producer (Fire-and-Forget) (15 minutes, 1000B messages)

Order-Balanced Aggregate

Client	Samples	Geomean comparison msg/s	Sample range	Median CPU μs/msg	Comparison Ratio
Dekaf	2	1,446,139	1,395,659–1,498,445	0.72	1.45x
Confluent	2	995,249	917,674–1,079,382	1.72	1.00x

The aggregate uses the geometric mean across balanced same-VM samples run in both dekaf-first and confluent-first order. Raw ordered samples remain below.

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf (3conn)	0.62	629.06	2,653,207	2,746,307	+16.7%	+1.80%	2530.30	2,653,207	0	1.65
Dekaf (adaptive)	0.65	645.15	2,327,972	2,309,915	+14.2%	+1.66%	2220.13	2,327,972	0	1.51
Dekaf (confluent-first)	0.69	679.68	1,487,235	1,498,445	-0.4%	-0.06%	1418.34	1,487,235	0	1.03
Dekaf (dekaf-first)	0.75	736.54	1,394,493	1,395,659	-8.8%	-0.64%	1329.89	1,394,493	0	1.04
Confluent (dekaf-first)	1.64	-	1,027,057	1,079,382	-4.1%	-0.68%	979.48	1,027,057	0	1.68
Confluent (confluent-first)	1.80	-	954,617	917,674	+23.4%	+2.35%	910.39	954,617	0	1.72

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Producer (Fire-and-Forget), 3 Brokers (15 minutes, 1000B messages)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf (adaptive)	1.01	946.53	1,251,558	1,211,244	+17.1%	+1.69%	1193.58	1,251,558	0	1.27
Dekaf (3conn)	1.04	957.63	1,229,414	1,188,344	-19.2%	-1.73%	1172.46	1,229,414	0	1.27
Dekaf	1.07	1019.22	1,151,570	1,165,881	+5.8%	+0.59%	1098.22	1,151,570	0	1.23
Confluent	2.11	-	723,211	713,584	-9.1%	-0.72%	689.71	723,211	0	1.53

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Producer (Acks All) (15 minutes, 1000B messages)

Order-Balanced Aggregate

Client	Samples	Geomean comparison msg/s	Sample range	Median CPU μs/msg	Comparison Ratio
Dekaf	2	1,540,523	1,515,172–1,566,298	0.70	1.25x
Confluent	2	1,228,208	1,153,504–1,307,751	1.42	1.00x

The aggregate uses the geometric mean across balanced same-VM samples run in both dekaf-first and confluent-first order. Raw ordered samples remain below.

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf (confluent-first)	0.70	714.82	1,557,326	1,566,298	-1.1%	-0.11%	1485.18	1,557,326	0	1.09
Dekaf (dekaf-first)	0.70	709.76	1,486,940	1,515,172	-6.5%	-0.58%	1418.06	1,486,940	0	1.04
Confluent (confluent-first)	1.37	-	1,288,518	1,307,751	-3.5%	-0.27%	1228.83	1,288,518	0	1.77
Confluent (dekaf-first)	1.46	-	1,161,401	1,153,504	+9.4%	+0.94%	1107.60	1,161,401	0	1.70

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Producer (Acks All), 3 Brokers (15 minutes, 1000B messages)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf	0.95	937.60	1,117,845	1,129,651	+7.7%	+0.60%	1066.06	1,117,845	0	1.06
Confluent	2.05	-	751,552	751,176	+25.0%	+2.03%	716.74	751,552	0	1.54

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Producer (Fire-and-Forget, Idempotent) (15 minutes, 1000B messages)

Order-Balanced Aggregate

Client	Samples	Geomean comparison msg/s	Sample range	Median CPU μs/msg	Comparison Ratio
Dekaf	2	1,490,271	1,459,918–1,521,256	0.69	1.18x
Confluent	2	1,263,552	1,233,956–1,293,858	1.41	1.00x

The aggregate uses the geometric mean across balanced same-VM samples run in both dekaf-first and confluent-first order. Raw ordered samples remain below.

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf (adaptive)	0.61	611.45	2,425,532	2,479,418	+11.5%	+1.13%	2313.17	2,425,532	0	1.48
Dekaf (3conn)	0.60	587.96	2,274,289	2,280,285	-4.5%	-0.39%	2168.93	2,274,289	0	1.36
Dekaf (confluent-first)	0.70	704.59	1,508,119	1,521,256	-1.3%	-0.12%	1438.25	1,508,119	0	1.06
Dekaf (dekaf-first)	0.68	658.45	1,452,680	1,459,918	+0.9%	+0.08%	1385.38	1,452,680	0	0.99
Confluent (dekaf-first)	1.40	-	1,222,768	1,293,858	+2.1%	-0.08%	1166.12	1,222,768	0	1.71
Confluent (confluent-first)	1.42	-	1,211,858	1,233,956	+0.2%	-0.02%	1155.72	1,211,858	0	1.72

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Producer (Fire-and-Forget, Idempotent), 3 Brokers (15 minutes, 1000B messages)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf (adaptive)	0.82	815.42	1,244,391	1,255,058	-1.7%	-0.09%	1186.74	1,244,391	0	1.02
Dekaf (3conn)	0.83	817.60	1,239,171	1,250,776	-0.6%	-0.06%	1181.77	1,239,171	0	1.02
Dekaf	0.88	864.89	1,110,101	1,116,744	+0.9%	+0.09%	1058.67	1,110,101	0	0.97
Confluent	1.97	-	800,917	803,597	+4.5%	+0.41%	763.81	800,917	0	1.58

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Producer → Consumer Round-Trip Steady State (15 minutes, 128B messages)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf	0.88	3066.00	1,466,307	2,625,535	+59.2%	+571.55%	178.99	1,466,307	0	1.29
Confluent	1.85	-	127,287	1,641,424	+10.5%	+107.74%	15.54	127,287	0	0.24

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Round-Trip Validation

Client	Expected	Consumed	Missing	Duplicates	Corrupt	Out of Order	Wrong Partition	Unexpected	Timed Out	Result
Confluent	19,792,477	19,792,477	0	0	0	0	0	0	no	PASS
Dekaf	19,792,477	19,792,477	0	0	0	0	0	0	no	PASS

Producer (Transactional EOS), 3 Brokers (15 minutes, 1000B messages)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf	228.84	228.84	889	1,193	+0.8%	+0.09%	0.85	1,185	0	0.27
Confluent	258.82	-	126	169	+10.1%	+0.94%	0.12	168	0	0.04

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Messages/sec counts broker-confirmed deliveries (end-offset delta). Accepted msg/s is the client-side append rate — a large gap means messages were buffered or dropped without ever reaching the broker.

Transaction Verification

Client	Accepted	Committed	Aborted	Delivered	Duplicates	Shortfall	Aborted leaks	Unexpected	Missing sentinels	Status
Confluent	151,100	113,400	37,700	113,400	0	0	0	0	0	PASS
Dekaf	1,066,700	800,100	266,600	800,100	0	0	0	0	0	PASS

Consumer (15 minutes, 1000B messages, 16,384B seed batches)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf	0.76	-	1,744,457	1,751,563	-1.1%	-0.10%	1663.64	-	0	1.32
Confluent	1.14	-	1,302,386	1,329,577	+5.9%	+0.59%	1242.05	-	0	1.48

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Consumer (Batch) (15 minutes, 1000B messages, 16,384B seed batches)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf	0.78	-	1,740,015	1,732,327	-6.8%	-0.67%	1659.41	-	0	1.36

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Consumer (Raw Bytes) (15 minutes, 1000B messages, 16,384B seed batches)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf	0.42	-	3,718,641	3,741,807	+2.6%	+0.25%	3546.37	-	0	1.57

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Consumer (Raw Batch) (15 minutes, 1000B messages, 16,384B seed batches)

Client	CPU μs/msg	CPU μs/request	Messages/sec	Median msg/s	Drift	Slope %/min	MB/sec	Accepted msg/s	Errors	Standing cores
Dekaf	0.34	-	4,168,381	4,094,242	+0.6%	+0.14%	3975.28	-	0	1.44

Median msg/s is the median sampled client-side throughput interval; it shows steady-state throughput without letting a short late-run stall dominate the whole-run average.

Rows and Comparison Ratio use Median msg/s when available; older result files without interval samples fall back to Messages/sec.

Drift compares last-third with first-third average throughput. Slope is the normalized least-squares trend; steady-state below 85% of peak or slope below -1%/min fails the regression gate.

Memory & GC statistics — latest run

Client	Scenario	Gen0	Gen1	Gen2	Total Allocated	Alloc/msg
Confluent	Consumer	23021	93	1	2663.62 GB	2.38 KB
Confluent	Producer (Fire-and-Forget)	231796	35	1	1109.33 GB	1.26 KB
Confluent	Producer (Fire-and-Forget)	192267	1	1	1031.03 GB	1.26 KB
Confluent	Producer (Fire-and-Forget), 3 Brokers	149622	0	0	781.11 GB	1.26 KB
Confluent	Producer (Acks All)	260614	36	1	1254.39 GB	1.26 KB
Confluent	Producer (Acks All)	282816	1	1	1391.65 GB	1.26 KB
Confluent	Producer (Acks All), 3 Brokers	165567	1	1	811.72 GB	1.26 KB
Confluent	Producer (Fire-and-Forget, Idempotent)	267165	1	1	1308.82 GB	1.26 KB
Confluent	Producer (Fire-and-Forget, Idempotent)	270404	44	1	1320.72 GB	1.26 KB
Confluent	Producer (Fire-and-Forget, Idempotent), 3 Brokers	174194	0	0	865.04 GB	1.26 KB
Confluent	Producer → Consumer Round-Trip Steady State	6915	0	0	17.56 GB	953 B
Confluent	Producer (Transactional EOS), 3 Brokers	104	1	0	309.75 MB	2.10 KB
Dekaf	Consumer	26079	72	4	2959.98 GB	1.98 KB
Dekaf	Consumer (Batch)	25986	5	2	2952.68 GB	1.98 KB
Dekaf	Consumer (Raw Bytes)	5	2	1	492.08 MB	0 B
Dekaf	Consumer (Raw Batch)	9	2	1	989.48 MB	0 B
Dekaf	Producer (Fire-and-Forget)	225	4	2	788.32 MB	1 B
Dekaf	Producer (Fire-and-Forget)	222	3	2	192.89 MB	0 B
Dekaf	Producer (Fire-and-Forget), 3 Brokers	145	3	2	150.72 MB	0 B
Dekaf	Producer (Acks All)	224	3	1	791.56 MB	1 B
Dekaf	Producer (Acks All)	227	2	1	164.98 MB	0 B
Dekaf	Producer (Acks All), 3 Brokers	132	3	2	232.53 MB	0 B
Dekaf	Producer (Fire-and-Forget, Idempotent)	207	2	1	152.70 MB	0 B
Dekaf	Producer (Fire-and-Forget, Idempotent)	203	3	2	755.63 MB	1 B
Dekaf	Producer (Fire-and-Forget, Idempotent), 3 Brokers	151	3	2	135.65 MB	0 B
Dekaf	Producer → Consumer Round-Trip Steady State	1071	3	1	2.81 GB	153 B
Dekaf	Producer (Transactional EOS), 3 Brokers	84	1	1	175.07 MB	172 B
Dekaf (3conn)	Producer (Fire-and-Forget)	327	3	1	1.27 GB	1 B
Dekaf (3conn)	Producer (Fire-and-Forget), 3 Brokers	188	5	2	737.70 MB	1 B
Dekaf (3conn)	Producer (Fire-and-Forget, Idempotent)	276	2	1	1.14 GB	1 B
Dekaf (3conn)	Producer (Fire-and-Forget, Idempotent), 3 Brokers	157	3	2	670.81 MB	1 B
Dekaf (adaptive)	Producer (Fire-and-Forget)	307	3	2	1.13 GB	1 B
Dekaf (adaptive)	Producer (Fire-and-Forget), 3 Brokers	170	5	3	723.16 MB	1 B
Dekaf (adaptive)	Producer (Fire-and-Forget, Idempotent)	315	3	2	1.17 GB	1 B
Dekaf (adaptive)	Producer (Fire-and-Forget, Idempotent), 3 Brokers	148	3	2	675.10 MB	1 B

Confluent.Kafka uses native librdkafka; .NET GC allocation counters exclude unmanaged allocations.

---

About These Tests

Stress tests measure sustained performance over extended periods against real Kafka brokers, with both clients paired on the same VM for a fair comparison.

Methodology — how these numbers are produced

• Real Kafka: Tests run against actual Apache Kafka instances
• CPU Isolation: Brokers are pinned to dedicated cores and the client under test to its own cores, so the client — not the broker — is the measured bottleneck
• RAM-backed Broker Logs: Kafka log dirs are mounted on tmpfs so disk I/O never caps broker ingestion
• Delivered Throughput: producer tables report broker-confirmed throughput, measured as the end-offset delta across all partitions — not the client-side append rate, which can run far ahead of what the broker ever accepts
• Median Interval Throughput: table order and comparison ratios use median sampled client-side msg/s when available, which is less sensitive to short late-run stalls than the whole-run mean
• Same-VM Pairing: comparable Dekaf and Confluent scenarios run sequentially inside one job/VM; 1-broker producer acceptance lanes run twice in opposite client orders and publish a geometric-mean aggregate, while other lanes alternate order by workflow run number
• Backpressure Parity: both producers are bounded to the same 512 MB local buffer (Dekaf BufferMemory, librdkafka queue.buffering.max) and block on a full buffer, so neither client can absorb an unbounded backlog into RAM
• Consumer Loop Replay: Consumer tests re-read a pre-seeded topic (seek to beginning when drained) instead of racing a live feeder, so the consumer itself is measured; table headings report the 16KB seed batch size because it amplifies per-batch costs relative to well-batched workloads
• Delivery Latency Sampling: 1 in 1000 produced messages is awaited end-to-end to record true broker round-trip latency
• Adaptive-Connections Row: the four paired fire-and-forget/idempotent producer lanes also run one Dekaf pass with the library default (adaptive connection scaling enabled, one connection to start), labelled Dekaf (adaptive); like the 3-connection control it is excluded from the headline comparison, which stays pinned to one connection to match Confluent
• Round-Trip Correctness: Bounded sequenced payloads are consumed back and checked for corruption, wrong partitions, gaps, duplicates, and reordering
• Round-Trip CPU Scope: CPU time covers both bulk production and consumer validation; it is not a producer-only metric
• Round-Trip Alloc Scope: the GC/alloc window likewise spans production plus consume-side validation; values are deliberately consumed as byte[] on both clients for parity, so each consumed payload is materialized as a fresh array (~152 B at 128 B messages) — the expected allocation floor for this lane, not a leak
• CPU Efficiency: CPU time per message differentiates client efficiency even at equal throughput
• Noise-Aware Trends: each scenario's throughput, CPU per message and Dekaf delivery-latency p50/p95/p99 are compared with its last 10 matching runs using a median ± 2×MAD band; one adverse excursion warns and two consecutive regressions fail the workflow, and paired lanes fail only when the same-run Dekaf/Confluent ratio regressed too
• Latency Product Bars: p95 within 3× the configured delivery-latency target and p50/p99 within 2× the same-run Confluent control are reported as product goals, not gates; a lane that misses a bar shows a warning until the trend band moves it
• Parallel Execution: Each scenario runs in its own isolated environment
• Both Clients: Direct comparison between Dekaf and Confluent.Kafka
• Memory Monitoring: Tracks GC behavior and memory usage over time
• Error Rates: Ensures stability under load