I host and support websolr.com and bonsai.io and have seen a lot of search implementations.

The main thing for good stability and performance is to be very good at batching your updates. You don't want to sling a ton of highly-parallel single-document updates at Lucene, lest you thrash the JVM and start garbage collecting like crazy.

From there, on the query side, you'll want to get a good working knowledge of the different tokenization and analysis options. There are a lot of subtle and interesting combinations to be had in there that influence performance and relevance of your search results.