Report PDFs in Python

Why reports become PDFs

A report is a snapshot of numbers at a moment in time. Its whole value is that it does not change when the underlying dashboard does. That is the difference between a document and a view of a database, and it is why analytics, operations and account teams keep asking for this.

What a report has to carry

Before any of the code below matters, the template has to produce a document that is actually a report. These are the fields that make it one, and the ones a reader or an auditor will look for first.

• the reporting period as an explicit start and end date
• the generation timestamp, distinct from the period
• the data source or filter set, so a reader can tell what was excluded
• each metric with its unit, since a bare number is not a figure
• comparison against the prior period where the report claims a trend

The Python implementation

httpx is the client most projects here already have. The render happens in an async endpoint that returns a streaming response, and you stream the response with httpx.stream and yield chunks.

import os, httpx
from fastapi import FastAPI
from fastapi.responses import StreamingResponse

app = FastAPI()

@app.get("/report/{report_id}.pdf")
async def report_pdf(report_id: str):
    html = render_report(report_id)

    client = httpx.AsyncClient(timeout=60.0)
    request = client.build_request(
        "POST",
        "https://api.pdfpipe.xyz/v1/pdf",
        headers={"Authorization": f"Bearer {os.environ['PDFPIPE_KEY']}"},
        json={"html": html, "options": {"format": "A4", "printBackground": True}},
    )
    response = await client.send(request, stream=True)

    # stream=True keeps a long report off the worker's heap.
    return StreamingResponse(
        response.aiter_bytes(),
        media_type="application/pdf",
        background=response.aclose,
    )

The CSS that makes a report page correctly

The layout problem specific to this document is that charts and tables must not split mid-element, and a report long enough to matter will page-break somewhere awkward unless break rules are set. These rules handle it.

/* Charts and tables must not split. Sections start on a fresh
   page so a heading never ends up alone at the foot of one. */
.chart, figure, table { break-inside: avoid; }
h2 { break-after: avoid; }
section { break-before: page; }
section:first-of-type { break-before: auto; }

@page { size: A4; margin: 20mm 18mm; }

What goes wrong in Python

requests holds the entire body in memory before you touch it. httpx.stream does not, and on documents past a few megabytes that difference decides whether your worker survives a burst.

What people try first

Most Python projects reach for WeasyPrint, which is excellent at CSS Paged Media but needs Cairo, Pango and GDK-PixBuf present in the image, and diverges from a browser on modern layout. That works until it is running on more than one machine, at which point the browser becomes the thing you operate rather than the thing you use.

Where the report lives afterwards

Rendering is the short part. Reports pile up faster than any other document here because they are produced on a schedule whether or not anyone reads them. Most teams keep twelve months and expire the rest, which makes storage lifecycle a design decision rather than an afterthought.

Getting the document right

• Check the period the figures cover, because a report without an unambiguous date range is worse than no report.
• Generation is triggered by a schedule, usually month end or week end, so size the timeout for that path rather than for a health check.
• This is a batch shape: render the run through the batch endpoint rather than firing thousands of individual requests.
• Backgrounds are painted by default here, so a design that uses colour needs nothing set. Only an explicit print_background of false turns them off.

Frequently asked

Related

Other Python documents, and the same report in other stacks.