Product: WooCommerce store on WordPress. Client: anonymous. Engagement: site cleanup and maintenance. Status: diagnosed, reported to the plugin vendor; stopped after their next release.
How it came up
I was running a performance baseline on a client's online store before a cleanup. The server's PHP error log had grown from about 19 MB to 860 MB in two weeks, and it never rotated. Nearly all of the growth was one pair of messages, repeated hundreds of thousands of times a day: a warning about a missing database table name, followed by a failed query.
The errors had a clean starting point, one evening, and nothing before it. At the same moment, the store's background job runner went from about 20 requests an hour to about 230, and the server started hitting its limit on worker processes more often than it had in the previous ten days.
Finding the cause
The start time lined up with the store's automatic plugin updates, which run in batches every six hours. The inventory plugin had shipped a new version the day before, and its changelog said it had reworked a nightly job "to avoid timeouts." I ruled out the other likely causes in that window (an AI connector, admin logins, and the plugin's next version, which arrived hours after the errors started), then compared the new release's code with the previous one.
The rewritten job had two separate bugs, and each one made the other worse.
Bug one: work moved somewhere without permission
The new version moved part of a recalculation into a background request. That request runs anonymously: no logged-in user, no permissions. But the database table the recalculation needed was only registered when a logged-in user with purchase-order permissions loaded the plugin.
So in the background, the table name was empty. Every product produced two warnings and a failed query. And the failed query returned nothing, which the plugin then saved as zero. Every product's incoming-stock figure was quietly overwritten with 0.
The failed query returned nothing, and the plugin saved nothing as zero.
Bug two: two clocks
The job picked products whose last update was more than three hours old, measured in UTC. But it wrote each product's update time shifted by the store's time zone. The store was seven hours behind UTC.
So a product updated one second ago looked seven hours old, always past the three-hour cutoff. Every product was "stale" at all times. Every batch was full, every full batch queued another, and the job re-processed the whole catalog, forever. A store at UTC, or within three hours of it, would have caught up and stopped. This one never could.
Fixing only the first bug would have stopped the log flood but left the loop running. Fixing only the second would have stopped the loop but left the errors on every pass.
Checking my work
Before telling the client anything, I had a second, independent review of the plugin's source done against my explanation. It confirmed the mechanism, corrected some details, and found the zeroed stock field, which I'd missed. It also showed that the plugin's own setting to disable this job wasn't a safe workaround, because the replacement job made the same failing call.
What happened next
I wrote it up for the client with the options and their tradeoffs: roll the plugin back and pin it (which also gives up the release's security fixes), switch off the module behind the error flood (which stops the errors but not the loop), or wait for a fix. Nothing was changed without his approval, and the diagnosis wasn't billed.
I also posted a public bug report on the plugin's support forum describing both bugs and their causes, with no client or site details. Two days later the vendor released a new version. Once it auto-installed, the errors stopped and stayed gone in the days after, and the rollback was no longer needed. The one leftover was a 2 GB error log that needed truncating and a rotation schedule.
Time spent: an evening of diagnosis, unbilled. Status: resolved by the vendor's next release.
What this says about software you didn't write
Most of a store's code belongs to someone else, and automatic updates change it while nobody is watching. Two patterns here show up in every kind of codebase, AI-written or not. Work moved into a background process quietly loses the permissions it used to have. And when one part of a system stores time in UTC and another in local time, the bug only appears in some time zones, which is why it can ship and nobody else notices for days.