Your lab's Streamlit or Shiny app works. Can another scientist depend on it?

The app works until its author goes on holiday

A scientist builds a Streamlit app that turns an assay export into summary tables and plots. Colleagues rely on it. Then the author takes two weeks off, a long run fails, nobody knows how to restart it, and a research partner asks for access to their own projects only.

The science is fine. What is missing is the operating layer: who can see what, where runs live, how failures recover and who ships the next release. An R Shiny app on a shared server reaches the same point.

Northbank Bio, used throughout this guide, is a fictional example. It wants two internal groups and one partner to use its app, each seeing only their own projects. Start by recording the accepted inputs, method version, expected outputs and intended users. Changing the operating layer and changing the science need separate approvals.

Use the platform you already have

Check what your hosting already provides before anyone proposes a rebuild. Streamlit supports OpenID Connect sign-in. For Shiny, Posit Connect content settings limit an app to logged-in users or to specific users and groups.

Those controls decide who can open the app, not which project, run or file a signed-in partner may see inside it. Write a small access table with the workflow owner, listing groups, projects and actions, then enforce it in the app.

Make a submitted run survive the browser

Keep short interactions simple. When computation must outlive a closed tab, use your existing job service or add the smallest durable queue that works. Streamlit session state keeps values during a session; it is not storage for a run that must survive a server restart.

Give each request a stable ID. Retrying after a timeout must not launch a second copy, while a deliberate rerun gets a new run ID linked to the first. Save parameters, input references, outputs and failure details outside the session, so a user who refreshes returns to the same run and sees whether it is queued, running, finished, failed or cancelled.

A corrected assay export becomes a new input version even when its filename is unchanged. Reusing a result because a filename matches can serve an old calculation after the scientist believes the correction ran.

Test permissions on the result, not the menu

Hiding another project in a dropdown is interface design, not access control. Check permission when accepting a request, reporting its status and serving every result or download link, including links generated from backend storage.

Caches leak quietly. Streamlit’s caching documentation says cached values are available to all users by default. Shiny’s bindCache() is scoped to the application by default, so every user session shares it. Choose the scope deliberately and keep an authorization check in front of cached results.

Sign-in decides who arrives. The app still has to decide what each person may see.

In the acceptance test, the partner signs in, copies an internal run ID and requests its result directly. Access is denied. Repeat with a warm cache and an old download link, then revoke the partner and confirm that sessions and links expire as agreed.

Prove the calculation and the recovery

Keep a small scientist-approved reference dataset with expected outputs. Compare the deployed calculation against it under an agreed rule: exact values where possible, justified tolerances where not. A good-looking chart is no evidence that the numbers stayed right.

Also submit an unsupported file, double-click submit and drop the connection mid-run. Errors should say what to fix without exposing another project, and a failed run must never leave half-written outputs that look complete.

Hand over an operation, not just a repository

Name who owns deployment, access changes, failed jobs and scientific updates. Before sign-off, someone other than the author should deploy an approved revision, find a failed run, restore retained data and roll back a release.

Measure completed tasks, manual interventions, support time and running cost against today. Hosting, sign-in and tests do not confer scientific or regulatory approval: the scientific owner decides whether outputs stay valid for their intended use.

Scope the first team release

Request a call from this article and tell us in a few lines what the app does and what broke last. The first conversation needs no data; confidential code and files can wait for an NDA or your approved access process. See how we work with biotech teams. If the app wraps a multi-step pipeline, the bioinformatics handover guide covers its run evidence.