Ask the questions in order; most platforms drop out at the first two, before price comes up.
| Question | What rules a platform out | See |
|---|---|---|
| Where do producers and consumers run? | Another cloud or region: egress and latency | 6.17.6 |
| Which Kafka 129 features do you rely on? | No transactions, old version, short retention | 6.16.1, 6.17 |
| Who is on call for the brokers? | Self-managed with no one able to run 6.13-6.15 | 6.19.1 |
| What does it cost at today's and 10x traffic? | Floors that dwarf a small stream | 6.16.6, 6.18.8 |
| How do clients authenticate? | Only proprietary identity, if you must stay portable | 6.14, 6.17.1 |
| What runs around Kafka? | Registry, Connect or Flink 129 missing or vendor-only | 6.8, 6.9, 6.11 |
| How would you leave? | No way to replicate out with group offsets | 6.19.2-6.19.4 |
Test features rather than reading datasheets: run your clients' tests, transactions and consumer protocol included, against a trial cluster, since "Kafka compatible" covers different subsets. Price the traffic expected in two years, every group's reads included. A platform you can replicate out of is a decision you can revisit.
BookNest's answers: one shop in one region, at most 60 events an hour (Capacity Planning), clients in three languages using idempotence and transactions, no one on call at night. That points to managed, real Kafka near the shop, SCRAM or certificate logins, a Confluent-compatible registry and MirrorMaker 2 as the rehearsed exit: the grounds on which Choosing a Platform chose Aiven's Startup plan. If the shop moves into one hyperscaler, ask again.