Python Desktop App Development

PyQt and PySide expose the Qt toolkit, which gives tables, charts, docking panels and a native feel across Windows, macOS and Linux. The two libraries are close relatives with different licensing routes, and that difference matters for commercial distribution. Tkinter ships with Python, needs no extra dependency and suits a utility with a handful of forms. It looks dated, and that is sometimes the correct trade for a tool used by four people.

We choose by screen complexity and audience. A dashboard with filters, sortable tables and export buttons belongs in Qt. A converter that takes a folder and writes a summary is fine in Tkinter, and a command-line interface may be better still if the users are comfortable with one.

A Python program that runs on the developer machine is not a deliverable. PyInstaller freezes the interpreter and the imported packages into a folder or a single executable, which usually works and occasionally surprises. Hidden imports, data files, compiled extensions and plugin directories each need an explicit rule. We build on the target operating system, because a Windows executable cannot be produced reliably from a Linux machine.

The environment is pinned with exact versions and a lock file. Scientific packages move quickly, and an unpinned install can change numerical results between two machines. A frozen build also needs a test that launches the packaged program, not just the source, because the two are not the same artefact.

A pandas DataFrame with a million rows will freeze a naive table widget. The fix is a model that pages, filters on the server or database side, and shows aggregates before raw records. We design around the questions the analyst asks rather than around the columns in the file, which usually reduces both the data moved and the time to an answer.

Charts need the same care. Rendering every point is unnecessary when the screen has eight hundred pixels. Downsampling preserves the shape and keeps interaction responsive, and the export path can still produce the full-resolution figure when a report needs it.

The strongest case for this route is an existing codebase: a pricing model, a simulation, a parsing library or a set of validation rules that the business already trusts. Rewriting that logic in another language invites two versions of the truth. A desktop shell keeps one implementation and adds the controls, logging and file handling around it.

We read the existing code before quoting. Scripts that assume a notebook session, hard-coded paths or a specific working directory need work before they can sit behind a button, and that work is visible in the estimate rather than discovered later.

Qt has more than one licensing route, and the correct one depends on whether the application is distributed internally, sold, or bundled with a device. Tkinter carries a permissive licence that suits closed internal tools. Third-party charting and reporting libraries add their own terms. We list every dependency and its licence before the first installer is built, so the legal question is answered while the architecture can still change.

Distribution matters too. An internal tool can be copied to a shared folder; a customer-facing program needs an installer, a version number and an update path. The packaging decision follows from who receives the file.

Tests cover three layers: the logic, the interface and the frozen build. The logic is usually the easiest, with pytest around the functions that produce numbers. The interface needs a small set of scripted interactions for the screens that carry the workflow. The frozen build needs a smoke test on a clean machine, because missing data files and hidden imports appear only there.

We also test with realistic file sizes. A ten-megabyte spreadsheet hides problems that a two-gigabyte export exposes immediately, and the users will bring the larger file on the first day.

Python is a poor fit for a program that must start instantly on old hardware, hold a very large dataset in memory, or ship as a small download. A compiled language or a web application with a server-side data layer will serve better. It is also the wrong choice when the interface is the product and needs the polish of a design-led desktop application, because the toolkit will fight that ambition.