The other issue with using PyInstaller to build a program into a single binary is that every time the program runs, it needs to extract its own contents to a temporary directory. This causes slow start-up and, if the program exits abnormally, orphaned files left on disk.

I think other packaging tools have a workaround for this particular problem but, in my experience, have other issues. I’d love to use Python for building CLI tools, but building a binary is so fragile that I only use the language for prototyping, and rewrite in C++ (maybe Go in future?) for production.

What you are describing sounds like the onefile mode. Pyinstaller also has a onedir mode which is much faster to start. I packaged some python ml models into a pyinstaller built exe and a larger java application calls them. Onedir makes for a big install but they only take a second or two to run.

I've had similar false positives from McAfee when first installing pyinstaller, but not on the using of executables generated by pyinstaller. I hadn't known that pyinstaller was used by malware, and had assumed it was because the bundled interpreter could potentially make any system call, thus making it flag on lots of heuristics.

> assumed it was because the bundled interpreter could potentially make any system call, thus making it flag on lots of heuristics

You seriously overestimate AV “heuristics”.

Signing created binary with valid certificate usually prevents most of antivir problems, as they act way more lenient in regards to those binaries.

I've had this issue with recent versions of Pyinstaller and had to roll it back to version 3.4 to avoid it.