Design principles

What DynaSharp holds itself to, and what follows from it for code that uses it.

• Stay true to DynamoDB. Every verb is an operation or a clause. There is no Save, no identity map, no lazy load and no implicit read before a write. Limit is per page, Update is an upsert, and batches are not atomic, as in DynamoDB.
• Nothing hidden. What the client adds, such as a version condition or a discriminator filter, shows up in ToRequestJson().
• Mistakes belong to the compiler. Types rule out what they can, analyzers check expression grammar and constant paths, and most of the rest fails before a request is sent. A broken expression is DYNA212, a warning, so it stops only a build that treats warnings as errors. The diagnostics list every rule.
• Pay only for what you use. Requests are written straight to pooled UTF-8 buffers, and known attribute names cost no string. A cost that shows up in a benchmark is a bug until it is explained.
• No reflection. The source generator produces the mapping, so Native AOT works and nothing is discovered at run time.
• Keys are data on disk. Encodings are specified and tested against an independent oracle, because changing one is a migration.
• Retries are honest. A retry can apply a write twice when a response is lost. The docs say where that matters, and MaxAttempts(1) is one call away.
• Tests answer to someone else. Wire formats are checked against botocore and real AWS responses, behaviour against DynamoDB Local and Dynoxide.

How it is built says what follows from these in the code: why a command is a ref struct, why a condition is a struct and a key holds values instead of text.