The underlying model of what tags are for depends on their purpose. On a social media site like Instagram, they can be allowed to expand indefinitely, but if you apply the same model to a news site, they can become distractors rather than enhancing user experience.

My client was an alternative news organization that had allowed contributors to enter free-form tags for over a decade without any editorial oversight. As a result, the tags database had ballooned to more than 30,000 tags, over 20,000 of which had only been used a handful of times. Because they wanted to use the tags to enhance discoverability, I suggested that they reduce to a usable number, treating them more like categories than the #twitterlike hashtags that had proliferated. 

I helped them engage a consultant with a background in database design and library classification who helped them rationalize their tags to a usable number (just over 1000) that captured the majority of their uses and standardized spelling, punctuation, and taxonomy.

I then created a series of custom Drupal/Drush modules to clean the database back to the remaining tags, mapping them to the remaining values.