Results from repeatable container benchmarks and sustained validation across the supported CDC sources and realtime transports.
During the July 22-23, 2026 run, the harness interrupted source and broker availability, restarted engine processes, and compared expected output with observed sink documents or client deliveries.
Five CDC workloads and three realtime workloads completed their full validation blocks. A ninth Redis raw-WebSocket workload completed 230 rounds. Every recorded cycle matched the expected document or delivery count; realtime rounds also reported zero gaps and zero duplicates.
Each direct CDC, projection, and 200-client realtime workload was isolated and verified independently. Reproduce these tests with production data shapes before assigning capacity limits.
Realtime rows each delivered 20,000 events to 200 clients: 4 million exact deliveries with zero gaps and zero duplicates.
The adaptive controller continuously reduces intake and dispatch pressure as memory approaches its budget, then restores concurrency as pressure recedes.
Peak RSS fell from 658.71 MiB to 490.27 MiB while sustained throughput increased from 19,527 to 19,976 events per second.
Docker Desktop with 7 available CPUs and an 8 GiB VM. Every engine case received 2 vCPU.
CDC cases ran with a 512 MiB hard limit. Realtime cases used a 1 GiB hard limit.
CDC used 1 KiB payloads. OpenSearch 2.17.1 ran with 2 vCPU, a 1 GiB heap, no replicas, and refresh disabled during ingestion.
All expected documents or deliveries had to arrive. Realtime additionally required zero gaps and zero duplicates.
