Redis Cache Strategy Benchmark: Cache-Aside vs. Framework-Managed Read-Through in Spring Boot under Concurrent Load

This dataset compares explicit Cache-Aside with Spring Cache-managed application-layer Read-Through (@Cacheable) in a Spring Boot application backed by Redis and MySQL. It concerns verified Hot Cache operation, not cold-cache behavior or a Redis-native read-through service. The two implementations share the application, repository/SQL, Redis, TTL, JSON response format and request ID pool within each experimental stage. Version 2.0.0 retains the original Environment A archive and adds 70 formal runs from Environment B, giving 100 formal runs in total: (1) Environment A: 2 strategies × 3 concurrency levels (25, 50, 100) × 5 repetitions = 30 runs; (2) Environment B Phase 1: the same 2 × 3 × 5 design = 30 replication runs; (3) Environment B Phase 2 v4: 2 strategies × 4 additional concurrency levels (5, 10, 150, 200) × 5 repetitions = 40 runs. Preliminary tests and the paused v3 runs are excluded. Each formal run uses a 15 s ramp-up followed by a 60 s measurement window, with a fixed 20-ID pool and prewarmed cache. The primary metrics are mean response time (AvgRT, ms), P95 response time (ms) and throughput (requests/s); HTTP error rate is auxiliary. A run, not an individual HTTP request, is the statistical unit. Concurrency is treated as categorical. The 70 supplemental runs passed their validity checks with zero HTTP errors, Redis misses and application database loads in the recorded validation intervals. Files: formal-30runs-20260928.tar.gz is the unchanged original Environment A release. redis-cache-supplement-v2.0-20261022.tar.gz contains the 70 supplemental runs, complete request-level JTL files, original run-level summaries, randomized plans, per-run validation evidence, 140 resource-monitoring files, selected logs, public provenance/checksums, documentation, a release-integrity verifier and 15 final analytical CSV tables, including a harmonized 100-run table. The supplemental JTL files contain 55,780,199 request records in total, of which 45,553,411 fall in the defined measurement windows; full JTL files also retain ramp-up observations. The accompanying .sha256 file verifies the supplemental archive. Resource coverage differs by stage. Environment B includes application, Redis and load-generator resource measurements; complete JVM GC/heap measurements are available only for Phase 2, and missing GC data for Environment A or B Phase 1 are not imputed. GC rates use actual sampling coverage. Redis counter intervals may include ramp-up. TCP monitoring scope also differs between B phases. The final analytical tables use separate stage-specific models g(Y) ~ Strategy * C(Concurrency) + C(Repetition), with log-transformed AvgRT and P95, untransformed throughput, HC3 robust standard errors and 95% confidence intervals. Environment comparisons use an explicit environment-interaction model with repetition blocks nested within environment. The original archive is retained for provenance; its older analytical outputs are not the final 100-run analysis. The new archive contains a file-integrity verifier and final result tables, not a complete one-command regression reproduction pipeline. Interpretation limits: Environment A and B are not a hardware-only intervention; software/image identity is not fully established. B Phase 1/2 were collected in different batches with different monitoring coverage, and the added concurrency levels do not overlap between phases, so phase and concurrency cannot be disentangled by a phase indicator. P95 has integer-millisecond resolution and many groups have zero within-group variation. Resource associations do not establish causal mechanisms, and conclusions are limited to these implementations and measured Hot Cache environments. The supplemental release has been sanitized for public sharing: host addresses, cloud/resource identifiers, private paths and sensitive configuration have been removed or anonymized. Raw JTL bytes and numerical experimental values are preserved; timestamps remain. Original and public file hashes document provenance, while SHA256SUMS verifies the sanitized files. Repository and versioned release: https://github.com/rys888/redis-cache-strategy-benchmark/releases/tag/v2.0.0Documentation and final analytical tables: https://github.com/rys888/redis-cache-strategy-benchmarkPrevious dataset version: https://doi.org/10.5281/zenodo.23011850