[ltbarcly3]

I first used pg full text in around 2008. I've also used SOLR and ElasticSearch to power search and recommendation in substantial products.

The issue I have had with postgres full text search isn't that it's too slow, it's that it's too inflexible. It's a nice way to add simple search to fields but poor if you want to tune the search at all. Even allowing for general substrings is too much to ask, even allowing for custom tokenization is too much to ask. There's no tokenization pipeline to speak of unless you want to write c extensions (which of course you can't do for hosted databases anyway). Solr and Elasticsearch let you set up very complex indexes and search processing via configuration. There's absolutely nothing that would prevent postgres from adopting a lot of this capability, but instead postgres offers literally NOTHING. I get the impression that most of the developers for postgres full text haven't spent much time with other solutions as from previous discussions they don't really understand what I mean when I talk about tokenization and filter setup, and they don't really understand why this is a deal-breaker even for very simple applications. Postgres just splits on whitespace (and lets you basically manually use stopwords and stemming, which is crap). There is really no way to concatenate fields in a clean way into a single index, which again makes it extremely annoying to work with. There's no way to score searches based on field weighting or really any other kind of weighting beyond BM. Compared to the alternatives it's a toy system.