Should I store JWT tokens in redis?

You can use Redis for storing jwt label. Redis is much faster and convenient for storing such data. The request to Redis should not greatly affect the performance. You can try the library jwt-redis

TLDR: If you want the capability to revoke the token at some point, yes, store it in something fast like Redis.

One of the well documented drawbacks of using JWT is that there's no simple way to revoke a token if for example a user needs to be logged out or the token has been compromised. Revoking a token would mean to look it up in some storage and then deciding what to do next. Since one of the points of JWT is to avoid round trips to the db, a good compromise would be to store it in something less taxing than an rdbms. That's a perfect job for Redis.

As suggested in the comments a good approach is to make the list a blacklist (i.e. a list of invalidated tokens). Upon each request you look up the list to ensure the token is not present in it. You can further improve on memory space and performance during the lookup step by using a probabilistic algorithm to store the token. A simple approach is to have layered lookups. For instance, you could have a small in-app store that only tracks the first few (e.g 1 to 4) bytes of your blacklisted tokens. Then the redis cache would track a slightly more complete version of the same tokens (e.g. first 2 to 8 bytes). You can then store a full version of the blacklisted tokens using a more persistent solution (filesystem, rdbms, etc). This is an optimistic lookup strategy that will quickly confirm that a token is absent from the blacklist (which would be the more common case). If a token being looked up happens to match an item in the in-app blacklist (because its first few bytes match), then move on to an extra lookup on the redis store, then the persistent store if need be. Some (or all) of the stores may be implemented as tries or hash tables. Another efficient and relatively simple to implement data structure to consider is something called a Bloom filter.

Obviously, you'd have to adapt the above approach if you routinely blacklist millions of long-lasting tokens (which may also indicate that you have a different problem).