pg_tokenizer will cache text analyzer, model and tokenizer object in memory for each connection. The cache will always be updated when calling create_..., drop_... functions. And it doesn't follow transaction isolation level.
The cache may not be cleared when you rollback a transaction. You may need to call drop_... functions manually or reconnect to the database to clear the cache.
Example:
BEGIN;
SELECT create_text_analyzer('text_analyzer1', $$
pre_tokenizer = "unicode_segmentation"
$$);
-- The text analyzer is created and cached in memory
ROLLBACK;
-- The text analyzer is still cached in memory, but no effect for other connections
SELECT drop_text_analyzer('text_analyzer1'); -- extra call to clear the cache