Fix vector type native adapters #1263
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
`Vector.from_native` and `Vector.to_native` both used to have alternative implementations using NumPy or the Rust extensions when available to speed up the process. It turns out, that both didn't follow the reference Python implementation (internally relying on `struct.pack`/`struct.unpack`) for 32 bit float values. * Python 3.14+ preserves the signaling/quieting bit of `NaN` values whereas NumPy doesn't and Rust doesn't make strong guarantees on how f64s are cast to f32s: https://github.com/rust-lang/rfcs/blob/master/text/3514-float-semantics.md * Both NumPy and Rust turn large f64s into `inf` when cast to f32, while Python raises an `OverflowError`. To harmonize the behavior, only the Python implementation remains. It has been significantly sped up by not iterating over the elements and calling `struct.(un)pack` multiple times, but instead making use of the fact that the format string allows to specify a number of values to (un)pack. With this the performance of the Python implementation was never slower than 5 times of what Rust/NumPy offered and for low dimension vectors even faster. It's deemed not worth the effort of maintaining multiple implementations assuring their behavioral parity for such a small performance gain. The same multi-value (un)pack approach has been applied to the pure Python implementation of the byte swapping to improve it's performance by a factor of roughly 30 (still a factor of ~10 slower than using NumPy/Rust).
RichardIrons-neo4j
approved these changes
Nov 5, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Vector.from_nativeandVector.to_nativeboth used to have alternative implementations using NumPy or the Rust extensions when available to speed up the process. It turns out, that both didn't follow the reference Python implementation (internally relying onstruct.pack/struct.unpack) for 32 bit float values.NaNvalues whereas NumPy doesn't and Rust doesn't make strong guarantees on how f64s are cast to f32s: https://github.com/rust-lang/rfcs/blob/master/text/3514-float-semantics.mdinfwhen cast to f32, while Python raises anOverflowError.To harmonize the behavior, only the Python implementation remains. It has been significantly sped up by not iterating over the elements and calling
struct.(un)packmultiple times, but instead making use of the fact that the format string allows to specify a number of values to (un)pack. With this the performance of the Python implementation was never slower than 5 times of what Rust/NumPy offered and for low dimension vectors even faster. It's deemed not worth the effort of maintaining multiple implementations assuring their behavioral parity for such a small performance gain.The same multi-value (un)pack approach has been applied to the pure Python implementation of the byte swapping to improve it's performance by a factor of roughly 30 (still a factor of ~10 slower than using NumPy/Rust).
Fixes: DRIVERS-162