No one really talks about the big problem in large scale search.

It’s calcification.

The document count is so high I can’t afford the time and resources to reindex and try changes.

The cluster is so large and tweaked we have significant cost and fear to try any new search systems.

The embeddings we need take so long to generate, insert, and test. They might not even be better anyway.

There are so many mature clusters using older Lucene based systems out there (solr, elastic, etc). Just delivering middling results.

Even if you know a newer system or approach, you can’t afford to run a second cluster for the duration of a migration.

This is a part of why I keep thinking about on-demand and ephemeral indexing.

Maybe it’s time to move past the era of pre-computing everything just to deliver 200ms search results a few times a day.