All four alternatives in the report meet the server-side API p99 target of 50 ms. RabbitMQ accumulates a larger backlog and takes longer to drain it after the peak. Fast acceptance does not imply fast task completion.
Conditions
Load
Target 500 RPS · slow main database · worker interruption
Docker Compose · the original run does not record exact CPU or Node.js version
Methodology
Four implementations expose the same POST /tasks contract. The main database is intentionally slowed down; ingestion and task transfer are separated.
Load and a worker interruption create a backlog. Server-side API latency, durable ingestion, and commits to the main database are measured separately.
The values below come from the saved comparison report. Drain time runs from the backlog peak to an empty backlog.
Results
The table scrolls horizontally.
Task ingestion and backlog · values from the comparison report
Approach
API p99, ms
Durable p99, ms
Peak backlog, tasks
Peak to zero, s
WAL
14
14
46,939
133
Ingest DB
44
44
49,988
138
Valkey Streams
46
45
49,695
152
RabbitMQ
23
22
88,256
353
API p99 is measured inside the server. Durable ingestion records confirmation at the ingestion boundary; completion in the main database is a separate measurement.
Limits of the conclusion
Server-side API p99 is not client end-to-end p99 and does not establish a client-side SLA.
Zero loss is a requirement. Aggregate counters without an ID-level reconciliation cannot establish it; the generator and final database counters differ for WAL.
The result depends on durability settings, load, and the failure model. It does not compare every broker feature or establish that queues are unnecessary.
Reproduce the experiment
Requires Docker, Docker Compose, and GNU Make. This command runs a new worker-interruption campaign; it does not guarantee the original numbers. The repository documents environment variables and durability settings.