Any solution is quite fine until it has to scale. But that depends on the dataset and on the goals. You were talking of simple search and replace operations; now it's about finding interesting patterns via search operation through a hugh dataset. This usually requires indexing of tokens / database-like operations / vectorizing terms.

I begin to suspect an XY Problem... maybe you should use a search engine like Swish-E or Lucene.

What are you really trying to do?

--shmem