Deciding that it's fast enough with the cold cache (likely what would happen in an emergency) and no longer require manual intervention

It's not that it's slow, it's that we end up rejecting lots of requests while the cache fills back up. To prevent this one use case (more_like requests, primarily via related articles and fetched on significant % of mobile page views) from dominating the cluster it has it's own PoolCounter, once that fills we reject a large stream of requests. If we have too much load the remedy would instead be to reduce the concurrency allowed by this PoolCounter until the cluster is stable and typical autocomplete/fulltext requests continue. If my memory serves the cache fills to a reasonable level in a few hours, but the initial spike will include significant numbers of rejected requests.

For reference on the non-automated approach: https://gerrit.wikimedia.org/r/c/operations/mediawiki-config/+/588754