Corrections

Several countries' daily price collection had been silently failing — in at least one case (Tunisia) for two days, with no record of the failure at all, not even an error.

Every scheduled collection job called the site through a helper that waited only 5 seconds for a response before abandoning the request. A real collection run — fetching a government feed and writing it to the database — routinely takes minutes. The connection was being cut before the job could even record that it had started.

The wait was raised to 90 seconds across all 138 scheduled jobs. The first affected job re-ran successfully immediately afterwards, taking 7 minutes 44 seconds — confirming the timeout was the cause rather than anything wrong with the data source.

Prices resumed collecting normally. Any price shown during the affected window carried its own real timestamp throughout, so nothing incorrect was ever displayed — the data was simply older than it should have been.

The system could spot and log when it had missed a window of price collection, and it saved a backup copy of the source data each time — but the final step, putting that backup copy back into the price history, had never been built. 37 gaps accumulated unrepaired.

The repair step was designed into the database from the start and documented as unbuilt. It simply never got written, and nothing flagged that omission.

The repair step was built and run against the whole backlog on 4 September 2026. 23 of the 37 gaps were successfully repaired from real archived copies of the original source data. It now runs automatically.

23 of 37 gaps repaired. The remainder had no usable archived copy and stay marked as gaps rather than being filled with guesses.

A review of the collection history surfaced 144 previously unrecorded gaps of more than 25 minutes, spread across the site's history — the largest being the six-day outage below.

The gaps had always existed in the run history, but nothing compared actual collection runs against the intended schedule. The data status page's 'gaps' figure had therefore always read zero, because no gap had ever been recorded — not because none existed.

Automatic gap detection now runs hourly and records every gap with its real start and end time. The data status page shows the actual list rather than a count that was always zero.

None of the 144 can be recovered. The UK scheme holds no history of its own, and backup copies of the source data weren't being kept until 27 August 2026. Charts spanning these windows show a visible break rather than drawing a line across it.

The government's official weekly UK national average figures weren't collected between the weeks ending 10 and 24 August.

The site fetched the source spreadsheet from a fixed government web address, on the assumption that the same address is reused each week. It isn't — part of the address changes on every weekly republication, and the old one began returning 'Gone'.

The site now reads the government's stable statistics page first and follows whatever the current link is, rather than assuming a fixed address. The missing weeks were then collected successfully.

Fully recovered. Unlike the live price feed, the government publishes its own historical series, so nothing was permanently lost.

Two collection attempts from 16 August never recorded finishing, and stayed marked as in-progress until they were cleared manually on 27 August.

The collector had no timeout. A run that crashed or hung without reaching its own error handling left a row marked in-progress indefinitely, which in turn made 'time since last successful run' checks read incorrectly.

Anything still marked in-progress after 10 minutes is now automatically marked failed, both at the start of every collection attempt and independently every 30 minutes.

No price data was affected — this distorted the site's own internal record of its health, not any price shown to a visitor.

No price history was recorded for five full days. There is data for 10 August, then nothing until 16 August.

There was no scheduler running the collector at all during this window — the automation that triggers collection wasn't added to the project until 16 August. Earlier data came from manual runs during initial development, not a working automatic pipeline.

Automatic scheduling was added on 16 August. It proved unreliable at short intervals and was later moved to run inside the database itself, which is how it works today.

Not recoverable. The UK scheme publishes only current prices and keeps no history, so a window we didn't record is gone for good. Charts covering these dates show a break.