Out of interest, what would you say the weaknesses are of the full-text search in PostgreSQL?
NB I've been using PostgreSQL for a few months on a side project and I've been hugely impressed. I wanted to add full text searching at some point and rather than using Lucene or Solr (or similar) I thought I would use PostgreSQL's own search capabilities - which certainly makes some things a lot simpler than using a separate search engine.
I can't speak for the original poster, but I have three issues with the postgres offering:
1) it has suboptimal support for handling compound words (like finding the "wurst" in "bratwurst"). If the body you're searching is in a language that uses compounds (like german), then you have to use ispell dictionaries which have rudimentary support for compounds and which aren't maintained any more in many cases because ispell has been more or less replaced by hunspell which has far superior compound support which in turn is not supported by postgres.
2) If you use a dictionary for FTS (which you have to if you need to support compounds), the dictionary has to be loaded once per connection. Loading a 20MB dictionary takes about 0.5 seconds, so if you use Postgres FTS, you practically have to use persistent connections or some kind of proxy (like pgbouncer). Not a huge issue, but more infrastructure to keep in mind.
3) It's really hard to do google-suggest like query suggestions. In the end I had to resort to a bad hack in my case.
Nothing unsolvable, but not-quite-elastic search either.
Thank you very much for 2) - this will allow me to considerably lessen the amount of infrastructure I'm having to take care of (pgbouncer can go now).
About 3: That's what I was using before moving to tsearch, but pg_trgm based suggestions sometimes were really senseless and confusing to users. Also, using pg_trgm for suggestion will cause suggestions to be shown that then don't return results when I later run the real search using tsearch.
I'm happy with my current hack though, so I just wanted to give a heads-up.
Thanks - I'd be delighted to get to the point where I have those kinds of problems!
What I will do is bear those limitations in mind and if I ever do have those problems there is, I guess, a reasonable chance that postgres development will have addressed them by them or I will redesign things to use a separate search engine.
If you use a dictionary for FTS (which you have to if
you need to support compounds), the dictionary has to
be loaded once per connection. Loading a 20MB dictionary
takes about 0.5
Does this happen for every connection, or just those that use FTS?
I'm thinking of a case where only maybe ~0.5% of queries will use FTS. I'm working on discussion software, and I'd like the discussions to be searchable, but realistically only a small percentage of queries will actually involve full text searching.
Only happens for connections that use FTS and only if you have configured it to use a dictionary (I believe there's no need to in case of an english body). But keep in mind that this means that you'll have to pay at least 0.5 seconds per search. If you do some google suggest style incremental searching on key press, you'll need to be faster than that.
So either use a persistent connection for full text searching (I'm connecting via pgbouncer for FTS-using connections) or, as karavelov recommended below, use http://pgxn.org/dist/shared_ispell/ which will load the dictionary once into shared memory (much less infrastructure needed for this one)
This isn't a problem of the index type (I'm using GIN in this case), but of the tsearch lexizer which only has limited support for hunspell.
The step that causes the trouble is ts_lexize(), not storing and consequently looking it up.
Again, I made it work for my case, but it was some hassle and involved running a home-grown script over an ispell dictionary I've created by converting a hunspell one into an ispell one.
Because using a real full text solution goes much farther than using regexes and LIKE. It also allows the use of indexes which many regexes and many LIKE patterns would not allow.
For example, I can now find the "wurst" in "Weisswürste" which, yes, I could do with a regex, but I can also find the "haus" in "Krankenhäuser" and all other special cases in the language I'm working with without having to write special regexes for every special term I might come up with.
Exactly. The only reason I have to bother with Solr at the moment is to get efficient ngram indexing for sequence data; which consists of lists of 7-8 character strings from various sources. What I have works, but feels like overkill for my case.
These advances within the GIN inverted index infrastructure will also greatly benefit jsonb, since it has two GIN operator classes (this is more or less the compelling way to query jsonb).
2. Solr uses TF-IDF out of the box, and PG doesn't.
3. PG is good enough for 90% of cases, but Solr has some advanced stuff that PG doesn't. Like integration with OpenNLP, things like the WordDelimiterFilter. (Andre3000 = Andre 3000)
4. PG is kinda annoying in that it will parse "B.C." as a hostname, even though I want it to be a province.
5. Solr is faster than PG, but PG has everything in one server.
6. Solr handles character-grams and word-grams better.
Another weakness not mentioned is document similarity search. You have one document and you want to find the N most similar in your collection to offer suggestions like Amazon does. You can do it with fts in pg, but it is way to slow to be practical. Custom fts engines are much better at that task.
If search is your application's main purpose, you'll want something more flexible.
Otherwise, it's fairly easy to implement, and you get full SQL support so joins, transactions, etc. So you can always prototype it and see what limitations you run into.
NB I've been using PostgreSQL for a few months on a side project and I've been hugely impressed. I wanted to add full text searching at some point and rather than using Lucene or Solr (or similar) I thought I would use PostgreSQL's own search capabilities - which certainly makes some things a lot simpler than using a separate search engine.