An Empirical Evaluation of Message Delivery Reliability and Recovery Characteristics in Redis Streams and NATS JetStream

Consumer-progress acknowledgments in Redis Streams and NATS JetStream have different persistence implications even under always-synchronize append policies. We evaluate that distinction alongside recovery timing, consumer concurrency and publication cost using single-replica deployments on a two-node Kubernetes homelab. The study compares Redis 7.4.2/8.10.2 and NATS Server 2.10.24/2.15.0, preserving an original campaign and adding three deployments with fresh broker processes and stores, matched publication durations, CPU sensitivity and storage diagnostics. All 240 active follow-up recovery episodes completed, including live, deliberately unacknowledging consumers; twelve plain Redis read controls remained pending. Receipt timing followed residual eligibility plus policy-dependent excess. Redis's integrated CLAIM path exhibited greater excess than the tested explicit reclamation loop in this quiet, one-message workload. Sixteen consumers increased current memory/periodic drain throughput by 10.43 to 14.49 times across deployment-specific comparisons, while cycle p99 increased and driver capacity affected rates. At sixteen publishers, current periodic/memory publication ratios averaged 0.8257 for Redis and 0.8207 for JetStream. Ten of 72 planned always-profile publication trials failed (13.9%), including four warm-up failures, compared with one of 72 periodic trials and none of 72 memory trials. Timed failures returned five-second client timeouts with unknown confirmation outcomes. A compact analytical model and selected Lean-checked statements delimit the claims; full proofs and detailed results are included in the appendices. The licensed artifact retains timings, failures, configurations and reproducible analysis. These observations characterize specified policies and acknowledgment costs; they do not establish equal durability contracts, power-loss survival or hardware-independent rankings. Research artifact. Version 1.3.0 contains the complete 43-page article, including all proofs, detailed results, and diagnostics in its appendices. The complete research artifact is available from the immutable GitHub v1.3.0 release linked below. Its sealed archive contains the source, original and follow-up observations, configurations, analysis, formal-model receipts, licenses, and validation records. The release also provides RELEASE.json, VERIFICATION.log, and SHA256SUMS for verification. This Zenodo record deposits the complete article; the research archive is hosted in that linked release. Historical preparation metadata inside the unchanged archive records its status before this Zenodo deposit. This is a preprint and has not been peer reviewed. License scopes. Original manuscript, research data, figures, derived results, and narrative documentation: Creative Commons Attribution 4.0 International (CC BY 4.0). Original software, formalization, build and deployment code: MIT. These licenses apply to their respective components, not as alternatives for every file. Third-party material retains its applicable terms; LICENSE.md in the archive details the scope. Source and versioned release: GitHub artifact v1.3.0.