Performance Benchmark

At low load (linear scale) the performance differences between the high-efficiency persistence modes (NoPersistence, File) against JPA, Redis, and Valkey.

In Redis and Valkey, the latency cost per task remains manageable at 20 req/s but explodes as the load hits 150 req/s, especially for the 100-step workflow.

At medium load and high load (logarithmic scale on the Y-axis to make the chart readable) Redis and Valkey exhibit massive latency spikes (RTT-induced delay exceeds the time between incoming requests), while NoPersistence and File modes remain stable at sub-10ms ranges. JPA performs better than Redis/Valkey at 80 req/s (only 17 drops), due to efficient connection pooling.

CPU work (iterations over items and serialization) consumes the bulk of the time. Since NoPersistence, File, Redis, and Valkey all have extremely low overhead for state management, they are all waiting on the same CPU-bound task to finish.

JPA is the only one that stands out significantly because its overhead for managing the persistence context and database transactions is high enough to be visible even in a low load scenario: it is consistently around 2x slower than the others.

Zone 1 (1MB): All modes except JPA maintain very low latency (under 100ms). JPA shows significantly higher overhead even with 1MB files due to the serialization/deserialization into the database.

Zone 2 (5MB): As the output size grows to 5MB, the difference between NoPersistence/File and Redis/Valkey begins to widen. Redis and Valkey start showing the cost of network transfer for larger payloads.

Zone 3 (20MB): At 20MB, the latency for JPA spikes dramatically, reaching nearly 18 seconds at 5 req/s. Redis and Valkey also see substantial increases (approaching 1 second) while NoPersistence remains the most efficient, as it bypasses the write/read overhead entirely.

Dominant External Delay: In Zones 1 and 2, the latency is almost identical to the injected delay. For example, in the 1000ms delay scenario, all modes hover around 1000-1020ms.

Saturation in Zone 3: When reaching 150 req/s with a 1000ms delay, the system hits a significant bottleneck. Latencies jump from 1000ms to around 3660ms across all modes. This indicates that the worker pool or connection limits are saturated, causing requests to queue regardless of the persistence mode used.

Orchestration Overhead: At lower delays (50ms), we can see the relative overhead of the modes. JPA consistently adds about 10–20ms of additional latency compared to NoPersistence, while Redis and Valkey sit in the middle with around 10ms of overhead.

Zone 1: low load

Zone 2: medium load

Zone 3: high load / saturation

JVM mode

k6 constant-arrival-rate scenarios

30-second runs

http_req_duration avg as the main latency metric