python: encrypt converts each value twice, once to check its range and once to write it #1640

Background

The Python SDK, stackencrypt in languages/python, encrypts a list of models with Cipher.encrypt. For each field of each model, its Rust extension reads the Python value, checks it against the field's type (for example, that an int fits a 16-bit Smallint), and writes it into the input that the engine stack-encrypt encrypts.

Problem

Each value is converted into the engine's scalar type twice.

• languages/python/crates/engine/src/encrypt.rs:224 calls values::scalar to check the range rules first, so that a value that does not fit is refused with its own error code and the field's name.
• encrypt.rs:280 then calls values::write, which converts the same value again to write it (languages/python/crates/core/src/values.rs:137).

The result is correct. The cost is a second conversion of every value in every encrypt call. Nobody has measured that cost. In a large batch it may be visible next to the request to ZeroKMS (CipherStash's key service), or it may not.

Proposal

1. Measure first: time encrypt on a large batch, such as 10,000 models of mixed field types, with and without the second conversion.
2. If the cost is worth removing, have values::write take the scalar that values::scalar already made, or have write return the misfit itself. Either changes values::write's signature, and every caller of it.
3. If the cost is not worth removing, close this issue with the numbers.

Relationship to other work

• Left open on purpose in review, because the cost was not measured: feat(python): encrypt models into EQL v4 rows #1499 (comment)
• Part of the Python SDK work in Python SDK: core call over psycopg, on EQL v4 and Stack Encrypt #1341.