sqs

On 9 Sep 2026 AWS raised LMI async and ESM timeout from 15 to 90 minutes (5,400 s). A 90-minute SQS consumer needs at least 9 hours of visibility. Sync and classic on-demand Lambda stay at 15 minutes.

FIFO guarantees shrink throughput—and unbounded queues only move backpressure to your AWS bill. Ordering, flow control, and Amazon MQ dead-letter patterns vs Kinesis resharding.

The worst AWS bills still come from small systems with invisible failure modes. July 2026 refresh — SQS/Lambda retry storms, zombie resources, public AI endpoints, and Bedrock/AgentCore spend loops — plus the detection radar.

A B2B SaaS stack that costs $500/month at launch does not need to cost $50,000/month at 100,000 users if the architecture decisions at each stage are deliberate. This is the end-to-end reference architecture with real cost numbers.

A technical guide to hybrid compute architectures that combine EC2, Lambda, Fargate, and Step Functions — with worked cost calculations, SQS buffering patterns, and decision frameworks based on invocation pattern rather than unit cost.

SQS is cheap until retry storms: short polling × uncapped workers. July 2026 — long poll 20s, visibility math, DLQ alarms, concurrency caps.

SQS, MSK Kafka, and Redis queues are not interchangeable. Each has different cost models, ordering guarantees, and failure modes. This guide covers when to use each, how to autoscale workers on queue depth, and how to build idempotent consumers.

SQS is "reliable" only if you understand visibility timeout, dead-letter queues, and the silent failure mode where messages get processed twice. Standard vs FIFO, DLQ tuning, Lambda integration, and the patterns that turn SQS from "queue with default settings" into reliable backbone messaging.