It does not unpickle every time you get something from cache. It maintains a first-level in-memory LRU per-thread object cache which is the actual Python object (not its pickle representation). In practice, that means you don't need to maintain your own results cache (with e.g. redis or memcached) as you do with most SQL-based systems. If it's slow for you, it's probably not due to its cache.

Write conflicts are indeed hairy to deal with. Using a BTree in places where you might instead use a dictionary usually makes this a lot better.

It sounds like you got burned by using it without understanding it very well.