Most of the time, I'd assume all of them I can store (deduplication is fine, if I can recreate the raw data afterwards). Which is a lot, because they should compress well (in the limit, approaching the same size as deduplication solution).
Sometimes those data points don't matter - like if they're generated by some program stuck in an infinite loop. But in other cases, they do - like e.g. if each message is caused by some event, like another program doing some processing, or user pressing a key, etc. - then timestamps will be useful to identify the exact cause (e.g. logs only happen when process X is processing mouse input, or when user presses one of 20 specific keys on their keyboard, or only when my microwave oven is running).
