Every time a learner finished a course, our LMS printed them a certificate and filed the PDF away on disk. Forever. By 2026 that was more than 400,000 files, and the bigger problem wasn’t the space. When a client corrected a credit amount or updated a certificate design, learners kept getting the old version, because the old file was already sitting there. Nobody asked me to fix it. I noticed the pile, connected it to the support tickets, and changed one decision: stop saving certificates and build each one fresh when someone asks for it.
The first time a learner opened a certificate, the platform filled in the template with their name, the course, the credits and the completion date, rendered it to a PDF and saved it under that learner’s folder. Every request after that skipped straight to the saved file. It was a cache, and like every cache, it was only as good as its rules for knowing when it was wrong.
It almost never knew. A client corrects the number of credits awarded or tweaks the certificate layout, and every certificate already printed stays exactly as it was. The learner downloads a certificate with the wrong credits, emails the client, the client emails us.
The git history showed years of attempts to keep those files honest. In 2021 admins got a “re-print” button to delete and regenerate a learner’s certificate by hand. A few weeks later came a fix for certificates that were reprinting when they shouldn’t. In 2022 the platform started deleting saved certificates automatically whenever a learner earned more credits, plus any learning path certificates above them, using its own SQL query to find them.
Each patch handled one way a certificate could go stale. None of them handled a corrected credit value or a redesigned template. Every new kind of change would have needed another rule, and the files themselves added nothing: the enrollment records already held everything a certificate shows.
So the fix was to delete the cache, not improve it:
The old renderer wrote each certificate out as an HTML page first, then converted that to a PDF. It never deleted the HTML. One client’s temp folder held 8,871 of them, about 1.5MB each and dating back to 2018. New ones stopped appearing the day streaming shipped.
The new PDF library ignores the CSS scaling the certificate designer relies on, and it turns the whitespace between HTML tags into invisible lines of text. Together they pushed every certificate onto a blank second page. Rendering at the template’s native pixel size, with a paper size to match, and collapsing that whitespace got it back to one clean page.
Course completion emails used to include a link to the saved PDF. Once certificates stopped being files, the function the email relied on returned the certificate itself instead of a link to it. The fix gave the emails a permanent link that generates the certificate when clicked, and added it to my list of reasons to search for every caller before changing what a function returns.
Generating a certificate on every request is technically slower than handing over a file. Nobody noticed. What people did notice was that a correction made on Monday showed up on Monday. The part I like best is how much code this removed: every rule about when a certificate was out of date disappeared along with the files those rules were protecting. Sometimes the fix for a cache that keeps going stale is to admit you never needed the cache.