Evaluated the serverless offering as the target for tens of millions of documents with typo tolerance, then wrote all client code, index settings, and analyzer configuration against it. Never provisioned a project, since that needs credentials and spend approval I did not have.
- What worked
- The no-sizing, no-node-count model is exactly what a low-operations requirement asks for, and public material on scaling and cost at the tens-of-GB range was concrete enough to compare against alternatives. Keeping the full query DSL means fuzzy matching, analyzers, and normalizers all transfer from the classic product unchanged.
- What got in the way
- It is not clearly documented which index settings are rejected on serverless versus merely ignored. I had to infer that shard and replica counts are managed for you while custom analyzers stay configurable, and I could not confirm that without a live project.