Benchmarks

Every number carries the conditions that produced it.

Each entry states hardware, date, run count, run duration, delivery semantics, acknowledgement configuration, sink, payload size, source pacing, and what the latency figure actually measures. Current results and historical results are labelled separately, because a benchmark is evidence of what happened on a specific build and machine on a specific date.

Governed fraud/AML pre-ingest screening

99,102 records/sec Apple M5 Pro · July 30–31, 2026 · 3-run mean · 10-minute runs · range 98,287–99,927 records/sec · CV 0.827% At-least-once; mTLS + Keycloak OIDC; fail-closed OPA; SHA-256 provenance on every record Current

Governed Kafka egress path (mTLS + OIDC + fail-closed OPA)

2.184M events/sec Apple M5 Pro · August 15, 2026 · 3-run mean · 10-minute runs · range 2.086M–2.354M events/sec · CV 6.766% acks=all with idempotence; Kafka sink mTLS + Keycloak OIDC; fail-closed OPA as a cached admission gate Current · cite range

Kafka WireEvent, at-least-once

2.201M records/sec Apple M5 Pro · July 28, 2026 · 3-run mean · 10-minute runs · range 2.161M–2.241M records/sec · CV 1.833% At-least-once, acks=1, no governance path Current

Kafka WireEvent, exactly-once

1.947M–2.312M records/sec observed range Apple M5 Pro · July 29, 2026 · 8-run mean · 10-minute runs · range 1.947M–2.312M records/sec (mean 2.061M) · CV 5.338% Exactly-once: acks=all, idempotence, one transaction commit per 100,000-record pipeline batch Current · cite range

Kafka-sink NOOP, source-paced capacity validation

1.995M records/sec Apple M5 Pro · July 30, 2026 · 5-run mean · 10-minute runs · range 1.984M–2.000M records/sec · CV 0.334% At-least-once, acks=1, no governance path Sustained capacity · not a ceiling

Kafka → MiniLM-L6-v2 ONNX CPU → Kafka, batching enabled

1,386 records/sec Apple M5 Pro · August 15, 2026 · 3-run mean · 10-minute runs · range 1,362.40–1,408.69 records/sec · CV 1.674% At-least-once; end-to-end processed records/sec, not isolated model operations/sec Current

Benchmark matrix

Filter and sort the published baseline and evidence profiles.

Use these as evidence paths to reproduce, inspect, and compare against your hardware, payloads, sinks, policies, control-plane expectations, and enrichment complexity.


Financial Services	99,102 records/sec	Apple M5 Pro July 30–31, 2026	Current
Replicated claim-grade campaign on the current runtime, with a tight replicate distribution. Path Transaction-shaped synthetic records → deterministic AML/sanctions screening → Kafka Semantics At-least-once; mTLS + Keycloak OIDC; fail-closed OPA; SHA-256 provenance on every record Source pacing Declared 100,000 records/sec source rate Replicates 3 × 600s Observed range 98,287–99,927 records/sec Sample CV 0.827% Latency boundary End-to-end admission to Kafka batch acknowledgement, 4,000-record batches Latency E2E P50 114.8–126.8 ms; P99 196.6–787.0 ms Integrity in = processed = out = Kafka-sent-ok on all three runs; zero send failures, drops, or residual in-flight records Evidence benchmark-runs/fraud_aml_preingest_100k_post_priming_claim_grade_20260730/CLAIM_GRADE_REPORT.md Claim boundary: Sustained throughput at a declared 100K/s source rate, not an uncapped ceiling. Latency P99 varies materially across runs and must be cited as a range. Distinct from the earlier AML/sanctions triage scenario run.
Financial Services	142,080–156,427 records/sec observed range	Apple M5 Pro August 1, 2026	Current · cite range
Valid current evidence. The replicate spread is wide enough that the observed range must be cited instead of a single point. Path Transaction-shaped synthetic records → optimized deterministic AML/sanctions screening → Kafka Semantics At-least-once; mTLS + Keycloak OIDC; fail-closed OPA; SHA-256 provenance; pinned optimized JAR Replicates 3 × 600s Observed range 142,080–156,427 records/sec (mean 149,084) Sample CV 4.816% Latency boundary End-to-end admission to Kafka batch acknowledgement, 4,000-record batches Latency E2E P50 306.7–330.3 ms; P99 573.0–682.6 ms Evidence benchmark-runs/fraud_aml_preingest_optimized_uncapped_claim_grade_20260801/OPTIMIZATION_AND_CLAIM_REPORT.md Claim boundary: Cite the observed range, never the mean as a point. Between-run spread tracks young-GC pause share. The pre-optimization baseline (124,313–131,020 records/sec) is retained separately.
Security	2.184M events/sec	Apple M5 Pro August 15, 2026	Current · cite range
Valid current evidence. The replicate spread is wide enough that the observed range must be cited instead of a single point. Path 512-byte synthetic events → NOOP → Kafka Semantics acks=all with idempotence; Kafka sink mTLS + Keycloak OIDC; fail-closed OPA as a cached admission gate Payload 512 bytes Replicates 3 × 600s Observed range 2.086M–2.354M events/sec Sample CV 6.766% Latency boundary Admission to Kafka batch acknowledgement, 48,000-record batches (not per-record latency) Latency E2E P50 106.7–118.2 ms; P99 291.0–311.2 ms Integrity Exact input/processed/output/Kafka-success equality; zero send failures, drops, backlog discards, or residual in-flight records Evidence benchmark-runs/nsf_pitch_evidence_audit_20260815T2224Z/AUDIT_REPORT.md Claim boundary: OPA runs as a cached admission gate (20 remote calls per 600-second run), not per-event authorization. Provenance, checkpoints, and receipt signing are disabled in this profile. mTLS covers the Kafka sink only; the source is synthetic. Cite the range — the replicate spread is 6.8%.
Security	3.216M events/sec	Apple M5 Pro July 28, 2026	Historical · not reproduced
The original evidence set remains internally valid and is preserved. A later audit on the current build did not reproduce the rate, so it is not published as a current result. Path 512-byte synthetic events → NOOP → Kafka Semantics acks=all with idempotence; Kafka sink mTLS + Keycloak OIDC; fail-closed OPA as a cached admission gate Payload 512 bytes Replicates 3 × 600s Observed range 3.162M–3.284M events/sec Sample CV 1.945% Latency boundary Admission to Kafka batch acknowledgement, 48,000-record batches Latency E2E P50 81.1–85.3 ms; P99 153.7–182.1 ms Evidence benchmark-runs/m5_governed_e2e_clean_claim_set_20260728/CLAIM_GRADE_REPORT.md Claim boundary: The July evidence set is internally valid and preserved with its immutable JAR and effective-config digests. A 2026-08-15 audit re-ran the same effective configuration and full cold-volume protocol on the current JAR and measured 2.184M events/sec. StreamKernel therefore does not publish 3.22M as a current result.
Kafka	2.201M records/sec	Apple M5 Pro July 28, 2026	Current
Replicated claim-grade campaign on the current runtime, with a tight replicate distribution. Path 512-byte synthetic events → STRING_TO_WIREEVENT → Kafka Semantics At-least-once, acks=1, no governance path Payload 512 bytes Replicates 3 × 600s Observed range 2.161M–2.241M records/sec Sample CV 1.833% Latency boundary End-to-end batch acknowledgement, 100,000-record batches Latency E2E P50 mean 262.6 ms; P99 mean 417.9 ms Evidence benchmark-runs/kafka_delivery_claim_grade_20260728_221103/claim-grade-report.md Claim boundary: Residual monotonic warming across the three runs is disclosed in the source report. No governance, provenance, or authorization is active on this path.
Kafka	1.947M–2.312M records/sec observed range	Apple M5 Pro July 29, 2026	Current · cite range
Valid current evidence. The replicate spread is wide enough that the observed range must be cited instead of a single point. Path 512-byte synthetic events → STRING_TO_WIREEVENT → Kafka Semantics Exactly-once: acks=all, idempotence, one transaction commit per 100,000-record pipeline batch Payload 512 bytes Replicates 8 × 600s Observed range 1.947M–2.312M records/sec (mean 2.061M) Sample CV 5.338% Latency boundary End-to-end transaction acknowledgement, 100,000-record batches Latency Run-summary means 282.5 ms P50 and 491.7 ms P99 Evidence benchmark-runs/kafka_eos_claim_grade_5x_20260729_110832/diagnosis-and-eight-run-report.md Claim boundary: The eight-run distribution is not tight enough to publish 2.061M as a deterministic point result. Always show the range alongside the average.
Kafka	1.995M records/sec	Apple M5 Pro July 30, 2026	Sustained capacity · not a ceiling
A validated sustained operating point at a declared source rate. It measures capacity at that rate, not the uncapped ceiling. Path 64-byte synthetic events → NOOP → Kafka Semantics At-least-once, acks=1, no governance path Payload 64 bytes Source pacing Declared 2,000,000 records/sec source rate Replicates 5 × 600s Observed range 1.984M–2.000M records/sec Sample CV 0.334% Latency boundary End-to-end batch acknowledgement, 100,000-record batches Latency E2E P50 mean 51.6 ms (47.5–55.5); P99 mean 313.3 ms (118.6–418.9) Evidence benchmark-runs/noop_kafka_m5_sustained2m_claim_grade_20260730/CLAIM_GRADE_REPORT.md Claim boundary: A rigorously measured source-paced operating point, not an uncapped producer ceiling. It does not replace the historical Intel 956K ceiling; no current claim-grade M5 producer ceiling exists.
Runtime	9.595M–12.408M ops/sec observed range	Apple M5 Pro July 31, 2026	Current · cite range
Valid current evidence. The replicate spread is wide enough that the observed range must be cited instead of a single point. Path 64-byte synthetic events → NOOP → DEVNULL Semantics No Kafka, no external I/O. Orchestration-only measurement. Payload 64 bytes Replicates 8 × 600s Observed range 9.595M–12.408M ops/sec (mean 10.720M) Sample CV 10.106% Latency boundary DEVNULL completion, 5,000-record batches Latency P50 4.7–6.1 ms; P99 6.2–7.6 ms Evidence benchmark-runs/noop_injvm_uncapped_extension_20260731_1811/EIGHT_RUN_DISTRIBUTION_REPORT.md Claim boundary: Publish the range as a distribution. Never cite the mean or the best run as a ceiling, and never compare a DEVNULL result against a Kafka-sink result.
Runtime	8.498M ops/sec	Apple M5 Pro July 31, 2026	Sustained capacity · not a ceiling
A validated sustained operating point at a declared source rate. It measures capacity at that rate, not the uncapped ceiling. Path 64-byte synthetic events → NOOP → DEVNULL Semantics No Kafka, no external I/O Payload 64 bytes Source pacing Declared 8,500,000 records/sec source rate Replicates 3 × 600s Observed range 8.4974M–8.4979M ops/sec Sample CV 0.0031% Latency boundary 5,000-record batch completion Latency P50 0.31–1.51 ms; P99 3.76–5.10 ms Evidence benchmark-runs/noop_injvm_sustained_8_5m_20260731/DIAGNOSIS_AND_CLAIM_REPORT.md Claim boundary: A stable connector-comparison reference at a declared rate, not an uncapped ceiling. It neither replaces nor is replaced by the uncapped distribution.
AI	1,386 records/sec	Apple M5 Pro August 15, 2026	Current
Replicated claim-grade campaign on the current runtime, with a tight replicate distribution. Path Kafka → DJL_EMBEDDING, EMBEDDING_TO_WIREEVENT → Kafka; CPU ONNX Runtime 1.20 Semantics At-least-once; end-to-end processed records/sec, not isolated model operations/sec Payload Fixed 96-byte synthetic records, tokenizer max length 16 Replicates 3 × 600s Observed range 1,362.40–1,408.69 records/sec Sample CV 1.674% Latency boundary Median of five-second window transform P50/P99, not an isolated model microbenchmark Latency Window P50 360.6–381.2 ms Integrity Exact processed/output/Kafka-success equality; source backlog discards at shutdown reported separately (0, 224, 96 records) Evidence benchmark-runs/nsf_pitch_evidence_audit_20260815T2224Z/AUDIT_REPORT.md Claim boundary: The 96-byte synthetic input is universally truncated at 16 tokens. This is not a representative natural-language short-text corpus, and the number is not isolated embedding operations/sec. The historical July result was 1,401.65 records/sec.
AI	1,402 records/sec	Apple M5 Pro July 28, 2026	Historical
Preserved exactly as measured, on the hardware, build, and date that produced it. Path Kafka → DJL MiniLM-L6-v2 CPU → WireEvent → Kafka Semantics At-least-once; end-to-end processed records/sec Payload 96-byte input, tokenizer max length 16 Replicates 3 × 600s Observed range 1,367.89–1,442.51 records/sec Sample CV 2.697% Evidence benchmark-runs/m5_validation_20260727/M5_VALIDATION_REPORT.md Claim boundary: Closely reproduced by the August audit at 1,386.49 records/sec. Retained as the original three-run result; the matched batching-disabled control measured 1,351.80 records/sec.
MongoDB	366,543 docs/sec	Apple M5 Pro July 23, 2026	Scenario proof run
A single completed scenario run, kept as use-case evidence. It is not a replicated, CV-qualified throughput campaign. Path StreamKernel → MongoDB insert, local 3-node replica set Semantics Write concern w:1 Replicates 1 × 300s Evidence benchmark-runs/mongodb-write-concern-mac-proof/20260723_185046 Claim boundary: Single proof run, not a replicated CV-qualified campaign. Matched runs on the same replica set measured 307,451 docs/sec at majority and 290,422 docs/sec at w:2.
Kafka	956K ops/sec	Intel i9-8950HK, 6 cores / 12 threads, 32 GB Spring 2026 (pre-M5)	Historical
Preserved exactly as measured, on the hardware, build, and date that produced it. Path Synthetic events → NOOP → Kafka, local single broker Semantics At-least-once, acks=1 Integrity 563M records, zero record loss Evidence StreamKernel-AI README.md historical Intel baseline table Claim boundary: Dated local Intel result, preserved as originally measured. No current claim-grade Apple M5 Pro producer-ceiling replacement exists: the M5-tuned profile reached a 2.384M mean but a 10.189% CV, which is not claim-grade.
Kafka	525K ops/sec	Intel i9-8950HK Spring 2026 (pre-M5)	Historical
Preserved exactly as measured, on the hardware, build, and date that produced it. Path 512-byte events → transform → Kafka Semantics At-least-once Payload 512 bytes Integrity 313M records Evidence StreamKernel-AI README.md historical Intel baseline table Claim boundary: Historical Intel baseline. The current Apple M5 Pro at-least-once result is 2.201M records/sec.
Kafka	507K ops/sec	Intel i9-8950HK Spring 2026 (pre-M5)	Historical
Preserved exactly as measured, on the hardware, build, and date that produced it. Path 512-byte events → transform → Kafka Semantics Exactly-once Payload 512 bytes Integrity 301M records Evidence StreamKernel-AI README.md historical Intel baseline table Claim boundary: Historical Intel baseline, measured at −3.5% against the matched Intel at-least-once run. The current Apple M5 Pro exactly-once distribution is 1.947M–2.312M records/sec.
Security	366K ops/sec	Intel i9-8950HK Spring 2026 (pre-M5)	Historical
Preserved exactly as measured, on the hardware, build, and date that produced it. Path Synthetic events → NOOP → Kafka with TLSv1.3 and fail-closed OPA Semantics At-least-once; TLSv1.3 client authentication; fail-closed OPA Integrity 217M records Evidence StreamKernel-AI README.md historical Intel baseline table Claim boundary: Historical Intel security-path baseline, preserved as measured. It predates the Keycloak OIDC profile and is not directly comparable to the current governed Apple M5 Pro path.
MongoDB	163K docs/sec	Intel i9-8950HK Spring 2026 (pre-M5)	Historical
Preserved exactly as measured, on the hardware, build, and date that produced it. Path StreamKernel → MongoDB insertMany Semantics At-least-once Integrity 95.5M documents Evidence StreamKernel-AI README.md historical Intel baseline table Claim boundary: Historical Intel insertMany baseline. The current Apple M5 Pro single-proof run measured 366,543 docs/sec at write concern w:1 on a local 3-node replica set.
Runtime	Retired no longer published	Not recoverable Pre-M5	Retired
Withdrawn. The supporting raw artifact is missing and the value is not republished. Path Synthetic events → NOOP → DEVNULL Semantics Not recoverable Evidence benchmark-runs/M5_BENCHMARK_INDEX.md Claim boundary: The raw artifact behind the exact 9.98M ops/sec and 3.47B record figures is missing, so the claim is retired rather than silently replaced. The value is statistically plausible inside the current 9.595M–12.408M distribution, but plausibility does not restore the artifact.
AI	336.8 records/sec	Not recorded in the current evidence tree Pre-M5	Unverified
Published in earlier collateral. No matching raw artifact was located during the current audit, so it is not used in headline claims. Path Kafka → in-process ONNX embedding → MongoDB Vector Semantics At-least-once; zero record loss reported in the source run Evidence Earlier StreamKernel collateral; no raw artifact located Claim boundary: Carried forward from earlier collateral. No matching raw artifact was located in the current benchmark evidence tree, so this value is excluded from headline claims pending a re-run.

These are reproducible local evidence lanes, not universal deployment guarantees. Hardware differs by row: current results were captured on an Apple M5 Pro (Mac17,9, 18 cores, 24 GB), while older baselines were captured on an Intel i9-8950HK and are preserved as dated evidence rather than restated on newer hardware. Measurement boundaries differ too: “in-JVM NOOP” means SYNTHETIC → NOOP → DEVNULL with no external I/O, “Kafka-sink NOOP” means SYNTHETIC → NOOP → KAFKA with producer acknowledgements, and batch-ack latency is not per-record latency. Source-paced rows measure sustained capacity at a declared rate and are not ceilings. Production results depend on hardware, payloads, sink behavior, network topology, JVM configuration, control-plane settings, and policy or enrichment complexity.

Inspect Sample Receipt

Reconciliation

What changed, and what we chose not to restate.

Benchmark provenance is a product asset. When newer evidence disagrees with an older published number, the older run is labelled rather than rewritten, and a result the current build does not reproduce is not published as a current result.

Governed Kafka egress path, 3.22M events/sec

Historical · not reproduced

The July 28, 2026 evidence set measured 3,216,213 events/sec across three 10-minute runs (CV 1.945%) and remains internally valid with its pinned JAR and effective-config digests. A August 15, 2026 reproducibility audit re-ran the same effective configuration and the full cold-volume protocol on the current build and measured 2,183,840 events/sec (CV 6.766%). StreamKernel publishes the current figure and preserves the July set as dated evidence. The difference is not attributed to a specific cause; the negative result is retained rather than discarded.

In-JVM NOOP orchestration ceiling, 9.98M ops/sec over 3.47B records

Retired

The raw artifact behind the exact figures could not be located, so the claim is withdrawn rather than quietly replaced with a newer number. The current eight-run uncapped distribution is 9.595M–12.408M ops/sec. The retired value is statistically plausible inside that distribution, but plausibility does not restore a missing artifact.

Kafka-sink NOOP producer ceiling, 956K ops/sec

Historical · preserved

Measured on an Intel i9-8950HK across 563M records at acks=1 with zero record loss. It is kept as dated Intel provenance and is not restated as an Apple M5 Pro result. No current claim-grade M5 producer-ceiling replacement exists: the M5-tuned profile reached a 2.384M mean but a 10.189% CV, which does not meet the bar. The separate 1.995M source-paced result answers a capacity question, not a ceiling question.

ONNX short-text inference, 1,401.65 records/sec

Historical · closely reproduced

The August 2026 audit measured 1,386.49 records/sec on the current build, 1.08% below the July figure, and added model and tokenizer SHA-256 digests that the original run metadata lacked. Both are end-to-end pipeline records/sec on fixed 96-byte synthetic input truncated at 16 tokens — not isolated embedding operations/sec, and not a representative natural-language corpus.

ONNX → MongoDB Vector, 336.8 records/sec

Unverified

Carried forward from earlier collateral. No matching raw artifact was located in the current evidence tree during this audit, so it is excluded from headline claims and marked unverified pending a re-run.

The authoritative evidence for every current claim is the timestamped campaign directory recorded in each row, including run logs, GC logs, effective properties, Prometheus snapshots, integrity reconciliation, and health-gate output. Where a report and a marketing figure disagree, the report wins.

Commercial path

Want a benchmark review for your pipeline shape?

Compare published profiles against your runtime envelope, sink behavior, payloads, policy path, and AI enrichment requirements.

Request Commercial Demo Discuss OEM Licensing