Images and files stop hanging after a deploy

On a portal running more than one server, an image or file could load instantly on one request and then hang for a full minute on the next — the same file, the same page, alternating. Course and documentation pages rendered with broken images, and the pattern came back after every update.

The two servers were talking past each other. A request always reached the server that owned the file, but the answer travelled back over an internal channel whose routing table was kept only in memory — and every deployment retires a server, which quietly erased it. The answer was then thrown away with nothing reported, so the request simply waited out its whole budget.

That routing table now lives in the database, so it survives a server being replaced. Requests keep being answered across a deployment instead of only until the next one.