Blog
Serverless, AWS architecture and AI agents: what I'd pick, why, and what it costs. My topics usually come from a technical problem I had to solve or something I had to explain. Serverless and event-driven in depth, architecture patterns in general, and AI agents running on serverless.
Filter by series or tag
5 posts
No posts match this filter.
Data Streaming on AWS: Kinesis, Firehose, Flink, or MSK?
Rule of thumb: Kinesis when there's no commitment to Kafka; MSK when the team already lives on Kafka.
Context and trade-off
Four streaming services on AWS, and “which one is best” is the wrong question: the choice comes down to context and ecosystem.
What you gainKinesis is simpler and native, and fits better with Lambda and the rest of serverless.
What you payOrdering is per shard, not global, and at-least-once delivery needs an idempotent consumer.
Is multi-cloud more resilient? The math says no
Rule of thumb: Multi-AZ and multi-region on the same cloud first. Multi-cloud only for a business reason, and then active-passive, documented in an ADR.
Context and trade-off
After every big cloud outage, the pitch comes back: “Let's go multi-cloud, so if AWS goes down we stay up on Google.”
What you gainNative managed services, the same IAM and the same tooling, for a fraction of the complexity.
What you payYou stay exposed to the very rare event of the whole provider going dark.
Would change if: Regulation, a contract or data sovereignty requires a second cloud.
AWS Lambda MicroVMs: run untrusted code with VM-level isolation (no infra to manage)
Rule of thumb: Lambda Functions for the event-driven backbone; MicroVMs only for the steps that run untrusted code.
Context and trade-off
Container, VM or Lambda Function: none of them gave VM isolation, fast launch and per-session state at the same time.
What you gainVM isolation with near-instant launch and state kept for up to 8 hours per session.
What you payNo sa-east-1 yet, ARM64 only, capped at 16 vCPU and 32 GB.
High Availability Has a Price: Resilience Is a Decision, Not a Stack
Rule of thumb: Choose between HA, multi-region and on-premise by the cost of each nine and by RTO/RPO, and record the accepted risk in an ADR.
Context and trade-off
Resilience is a business decision before it's a technical one, and the context behind each choice gets lost if nobody writes it down.
What you gainThe risk is mapped, owned and justified. Failure doesn't catch the team by surprise.
What you payNot everything gets HA: batch jobs, internal systems and dashboards can tolerate minutes or hours down.
AWS MCP Server: which one to use, when, and how to set it up (now that AWS recommends just one)
Rule of thumb: One server only, the managed AWS MCP Server. Knowledge is for those who just want to query docs, without an AWS account.
Context and trade-off
When the post came out, the answer was “it depends on the pain”. Three weeks later, the official setup guide started recommending a single server: duplicated tools confuse the agent.
What you gainLess config, fewer conflicts, the same IAM leash.
What you payAn agent that operates the account: start in a sandbox, with scoped IAM and a human in the loop for irreversible actions.
New posts, straight to your inbox
One serverless, architecture or AI agents topic on AWS per issue: what I'd pick, why, and what it costs. No spam.