Managed Kafka Options Compared

The fullest managed Kafka: Confluent runs the cluster in its account in your chosen cloud and region, and sells the surrounding platform with it: Schema Registry, a connector catalogue, ksqlDB, stream governance, and tiered storage. Support follows the product and is authoritative on it. Cost is usage-based, which is easy to start and hard to forecast; throughput, storage, connectors and partitions are billed on separate lines. It wins when you want the platform, not just the broker, and are not bound by where the data lives.

Kafka as an AWS service inside your own account: brokers on a supported version list, IAM authentication, CloudWatch metrics, and billing on the AWS invoice. AWS supports the service; your topics, consumers, connectors and the incidents that come from them are yours, which is why MSK estates so often carry an independent support contract at the Kafka layer. It wins for an all-AWS estate with a standard workload and a procurement team that prefers one vendor.

You run the cluster; a senior engineer is on the other end of the contract when it breaks, and alongside you for upgrades, KRaft migration and capacity. Open-source Apache Kafka is fully supportable this way, and so is Confluent Platform. It keeps version choice, connector freedom and data residency entirely yours, at a cost well below a managed service. It fits a platform team that can hold the pager and wants expertise rather than outsourcing.

The cluster stays in your account, on your VMs or Kubernetes, and a specialist takes over monitoring, patching, upgrades, capacity and on-call. Nothing about data residency changes. It costs more than a support contract, because it is engineer time with no infrastructure leverage, and it exists for estates where nobody on the team wants to own Kafka and the data cannot move to a hosted platform.

The self-hosted versus cloud question is usually framed as cost, and cost is the least reliable way to answer it. Four things decide it more cleanly. Where the data must live: a residency or air-gap obligation rules out a vendor-account service before price is discussed. Who holds the pager: self-hosting is cheap until the first multi-hour incident with nobody who has seen that failure before. How predictable the load is: usage-billed cloud Kafka is cheapest for spiky or small workloads and gets expensive on steady high throughput, which is where self-hosted hardware is at its best. How much of the platform you use: if you need Schema Registry, managed connectors and governance, you are buying a platform and cloud wins; if you need brokers, you are renting machines at a markup.

There is a middle that the question hides: self-hosted Kafka with someone else operating it, or with a support contract behind your own team. Both keep the data and the version choice in your account and remove the 3am problem.

1. Can the data leave your account? If not, Confluent Cloud is out; MSK stays in if you are on AWS.
2. Do you need Confluent's platform components, or just Kafka? If the platform, Confluent Cloud or Confluent Platform. If just Kafka, any of the other three.
3. Is there a platform team that can hold the pager? If yes, self-managed with a support contract. If no, a hosted service or operated-for-you.
4. Is the estate customised, or on a version a platform will not run? If yes, self-managed or operated-for-you.