# Numba 2024 Community Survey
# DRAFT

The goal of this survey is to gather more information about how Numba is used, and what priorities the community feels the project should focus on. Most of this survey is focused on how people use Numba today, but there are a few questions at the end about potential new use cases for a Python compiler that could be developed.

In this survey, we will use the term “Numba” to refer to all Numba-affiliated projects (like llvmlite) unless we specifically say otherwise.

The survey will remain open for [INSERT DATE 1 MONTH FROM NOW] and then, the raw data and an interpretation will be posted to the Numba discourse at https://numba.discourse.group/. The raw data will be anonymised and published under a restrictive Creative Commons licence https://creativecommons.org/licenses/by-nc-sa/4.0/.

The survey should take between 5 and 10 minutes to complete.

**Who is this survey for**: This survey is intended for anyone who is familiar with Numba (and/or llvmlite) and uses it directly, contributes to the Numba project, or develops other software packages that depend on Numba in some way. All use cases and situations are relevant, whether your usage of Numba is for fun, school, research, or commercial activities. If you do not know what Numba is, or have not tried to use it yet, this survey is not for you (but please go learn more about Numba!)

**Q1**: How do you interact with the Numba project? (check all that apply)

- [ ] I regularly use Numba to speed up Python code I have personally written.
- [ ] I maintain an open source project that depends on Numba.
- [ ] I maintain a proprietary or non-public project that depends on Numba (might be internal to your organization).
- [ ] I have contributed a pull request to Numba.
- [ ] I am a regular reader of the Numba discussion forums.
- [ ] I answer other people’s questions about Numba on forums, StackOverflow, etc.
- [ ] I package Numba for distribution (conda, Linux distributions, homebrew, etc)

> Purpose of Q1: This question is not very interesting on its own, since the survey participants will not be a representative sample of the whole Numba community. However, it will be useful for segmenting responses for all the other questions and looking at the distributions of those answers for different groups.

**Q2**: What kinds of applications do you use Numba for? (check all that apply)

- [ ] Hobby
- [ ] Educational projects
- [ ] Visualization
- [ ] Financial analytics
- [ ] Scientific research
- [ ] Compiler development
- [ ] Web-based microservices
- [ ] High performance computing on clusters
- [ ] Numerical simulations / “Digital Twins”
- [ ] Machine learning / AI
- [ ] Time series analysis
- [ ] Spatial / GIS data set processing
- [ ] Other [write in]

> Purpose of Q2: We have some anecdotal evidence about areas where Numba is used, but almost always from projects that can be discussed in public. This survey might get us a little more information about internal projects if the users are willing to share this information anonymously. This is kind of a grab-bag of use cases which are in no way mutually exclusive. Not sure if there are better ways to organize this. A better idea of popular use cases for Numba can help us select and focus on “representative projects” that can show us how Numba is used in specific kinds of applications.

**Q3**: Which of the following Numba features have you found useful? (check all that apply)

- [ ] Support for raising exceptions (raise statement)
- [ ] Support for catching exceptions (try-except statement)
- [ ] Support for generators
- [ ] Typed containers (typed.List, typed.Dict)
- [ ] Using NumPy functions in compiled code
- [ ] Compiling Python classes as structs using @jitclass
- [ ] Automatic multithreading with @jit(parallel=True) and/or prange()
- [ ] CUDA kernel compilation (@cuda.jit)
- [ ] Creating C callbacks with @cfunc
- [ ] Creating NumPy ufuncs with @vectorize or @guvectorize
- [ ] Creating stencil functions with @stencil
- [ ] Calling back into the Python interpreter using “with objmode”
- [ ] Caching of compiled functions to disk
- [ ] Ahead of time compilation (with numba.pycc)
- [ ] Debugging tools (gdb support, LLVM IR and assembly dumps, etc)
- [ ] Autovectorization
- [ ] The extension API (@overload, etc)
- [ ] First class function types (types.FunctionType)

**Q4**: Which of the following Numba features did not work for you or were too limited for your use case? (check all that apply)

**[Same feature list as above]**

**Q5**: Which of the following Numba features were usable for you, but you wish could be improved? (check all that apply)

**[Same feature list as above]**

> Purpose of Q3-Q5: These questions should help us narrow down which features in Numba we should focus on, and which we should consider freezing at current functionality, or possibly removing. The exact logic will depend on circumstances, but I can imagine various scenarios like:

> - Features which score low in Q3: Consider deprecating or at least not improving further because they are relatively unused.
> - Features which score high in Q4 and low in Q5: Consider deprecating for removal because they are a trap. (i.e. they don’t work for users and they don’t want to see them improved)
> - Features which score high in Q3 and Q5: Popular features that people wish could be better. Improvements here will benefit the existing audience, so maybe good to focus more on.
> - Features which score low in Q3 and high in Q5: It is tempting to want to do these, but they might be things we need to declare out of scope for Numba because they address additional use cases that don’t have current users.

**Q6**: Please rank these Numba development activities in order of how important they are to you. (From highest priority to lowest priority)

* Response to bug reports
* Review of pull requests
* Support for new Python minor releases (3.11, 3.12, etc)
* Support for new NumPy releases
* Support for new LLVM releases
* Improved reference documentation
* Creating examples of Numba usage and performance best practices
* Feature development
* Answering user questions in the forums

> Purpose of Q6: The goal of this question is to check our assumptions about what different user groups care about. In particular, we can check hypotheses like:
> - Prompt support for new Python minor releases matters more to OSS developers who depend on Numba than end users.
> - For many Numba users, Numba is essentially feature complete and they would rather that we focus on maintaining functionality than adding new functionality.
> - No one actually cares what LLVM version we use, except maybe llvmlite users and Linux distributions.

**Q7**: Please rank these potential new Numba features in order of how important they are to you. (From highest priority to lowest priority)

* Ahead-of-time compilation with support for type-based dispatching and multiple CPU target specialization
* Calling C++ (not C) code directly from nopython functions
* WebAssembly (WASM) support for standalone applications
* WebAssembly (WASM) support for Numba use in Pyodide
* Support for non-NVIDIA GPUs
* Better support for struct-like data types and arrays of struct data on CPU and GPU
* Context managers/with statement
* async features
* typed.Set
* yield from
* Python typing ↔ Numba type integration (i.e. get a Numba signature from Python type annotations)
* More NumPy functions
* Performance profiling
* Other [write in]

> Purpose of Q7: We’ve received anecdotal feedback about various new features like this that would be of interest to users and could be part of the Numba++ roadmap. It would be good to know which of these are the most compelling.

**Q8**: Do you use llvmlite separately from Numba? (Yes/No)

**Q9**: If yes, why do you use llvmlite? (check all that apply)

- [ ] Learning about compilers
- [ ] Compiler research
- [ ] Building Python compilers (other than Numba)
- [ ] Building non-Python or domain specific compilers
- [ ] Other [write in]

> Purpose of Q8 and Q9: We tend to develop llvmlite as “Numba-first”, with other use cases as a secondary concern. As we are starting to decouple llvmlite releases from Numba, I am curious if some of our other compiler components might have an audience with this group.

**Q10**: How would you prefer to get Numba project-related announcements and updates? (check all that apply)
- [ ] Mastodon
- [ ] X (formerly Twitter)
- [ ] youTube
- [ ] Discourse forums
- [ ] Changelog (in docs)
- [ ] Email newsletter
- [ ] Blog
- [ ] Other [write in]

> Purpose of Q10: We have a lot of different communication channels, and we need a better idea of which are being used, and if there are some channels (like Youtube) we need to pay attention to.

## To discuss
* Are there questions we can delete from this list (like the llvmlite questions, Q8 and Q9)?
* Are the choices reasonable for all the questions? Not too many and not too few.
* When should we roll this out?