How do I export a whylogs report as HTML?

Build profiles with `why.log(df)`, then use `NotebookProfileVisualizer` — `viz.set_profiles(target_profile_view=..., reference_profile_view=...)` and `viz.write(rendered_html=viz.summary_drift_report(), html_file_name='drift')`. That writes a self-contained HTML file.

What does the drift report actually show?

Per-column distribution comparisons between a target and a reference profile, with a drift statistic and the histograms side by side — which is the part that makes drift arguable rather than just alarming.

Why publish instead of keeping it in the notebook?

Because drift is almost always somebody else's problem. The column changed because an upstream team changed a pipeline, and they do not have your notebook.

Can I publish this automatically?

Yes. Generate the report in your scheduled job, then PATCH one saved report id. Each run appends a revision at the same URL, so the drift trend is a scrubbable history.

Does whylogs send my data anywhere?

Profiles are statistical summaries, not rows — that is the design. Publishing the rendered report shares those summaries, so check that no column's sketches are themselves sensitive before widening access.

Share a whylogs report

whylogs profiles datasets instead of copying them: sketches, counts, distributions, null rates — a summary small enough to log every run. Comparing two profiles gives you a drift report, which is the closest thing an ML team has to an early warning.

The warning then fires inside a notebook or a job log, which is the one place the people who caused the drift will never look.

Render the drift report

import whylogs as why
from whylogs.viz import NotebookProfileVisualizer

target = why.log(df_today).view()
reference = why.log(df_baseline).view()

viz = NotebookProfileVisualizer()
viz.set_profiles(target_profile_view=target, reference_profile_view=reference)
viz.write(
    rendered_html=viz.summary_drift_report(),
    html_file_name="drift",          # → drift.html
)

Self-contained HTML: per-column distributions, the drift statistic, the two histograms side by side. The side-by-side is the part that matters — a drift score alone is a number people argue with, and two overlaid distributions is an observation people act on.

Publish it

Drop drift.html into Comma, or from the job:

import requests, os

requests.patch(
    f"https://commareports.com/api/v1/reports/{os.environ['REPORT_ID']}",
    headers={"Authorization": f"Bearer {os.environ['COMMA_API_TOKEN']}"},
    json={"title": f"Drift — features_v2 — {run_date}",
          "html": open("drift.html").read()},
).raise_for_status()

Scripts run inside a sandboxed iframe (allow-scripts, no allow-same-origin), so the report's interactive column views keep working. See the API reference and publishing from CI.

Drift is a cross-team message

The useful thing about a drift finding is almost never "retrain the model". It is "column plan_tier gained a category on Tuesday" — a fact about an upstream pipeline, owned by a team that does not use your notebook, your MLflow, or your scheduler.

A URL is the only form of that message which survives the trip. With anchored comments, the reply lands on the column:

"We added enterprise_trial in the billing migration — expected."
"Null rate on region jumped because the CDN header changed."

That exchange, on the evidence, is the entire loop. See commenting on HTML.

One URL, a revision per run

PATCH the same report id from a scheduled job and the URL becomes the standing drift page for that dataset. Each run appends a revision, so the question "has this been creeping for a month or did it break yesterday?" is answered by scrubbing the history rather than by digging out old notebooks.

Pair it with a routine if you want the publish scheduled rather than wired into an existing job.

Who can see it

Per report: private, your team, anyone signed in at your domain, or anyone with the link. Profiles are summaries rather than rows, but a sketch over a small segment can still be revealing — use domain or team access. See sharing & access control.

Limits

Entry HTML: 5 MB. A report over hundreds of columns can approach it — profile the feature set the model actually uses.
Assets: 25 MB per file, 250 MB and 500 files total.
60 requests/minute per token.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Publish a drift report →