SQS backlog keeps growing: compare arrival and drain rates

A queue turns an outage into a delay

In a QuantumSketch practice pipeline, submissions enter a rendering queue and workers complete jobs. If 12 jobs arrive each second and only 8 finish, the backlog grows by about 4 jobs each second in this simplified constant-rate model.

More workers help only when they increase real completion capacity. A shared database, downstream limit or slow job can remain the bottleneck. Retries can also add load without increasing useful output.

Measure useful completion

Track age of work, arrival rate, successful completions, failures and downstream health. Queue length alone cannot distinguish a traffic spike from stuck processing. Standard-queue approximate metrics are evidence, not a transaction ledger.

Visibility timeout hides an in-flight message temporarily; it is not a throughput control. If processing outlasts that timeout, another worker can see the message again. Idempotent processing and a suitable timeout or extension strategy matter.

Experiment with capacity

Change arrivals, workers and per-worker capacity below. Step the model and inspect backlog and estimated drain time. If arrivals equal service capacity, new work can stop growth while the old backlog remains. Draining it requires spare completion capacity.

This model excludes message size, polling, batch effects, scheduling and real SQS quotas. It is useful for explaining conservation of work, not predicting production performance from a single slider.

Handle poison messages separately

A repeatedly failing message can consume worker effort indefinitely. A dead-letter queue and an appropriate redrive policy make failure inspectable. Choose thresholds for the application and alert on the resulting failures.

When restoring a failed downstream system, consider the backlog as additional future load. Controlled drain and backpressure are often safer than immediately maximizing every concurrency setting.

Sources and scope

Reviewed against these official references. The model’s supported scope appears alongside its controls.

• AWS: SQS visibility timeout
• AWS: SQS dead-letter queues