Likely when it's small enough to keep in RAM and you want to do some sort of on-the fly aggregation/correlation.

Then you can usually just store it in the memory of your application. No need to complicate your stack by running another service.

When you need to be able to retrieve the timeseries data for some period of time, storing it in the application memory doesn't really work since the application will restart whenever updates are made.

Also, redis timeseries offers the ability to downsample to some defined period which is really handy (and afaik isn't really provided by other timeseries databases) as well as set a retention policy.

Some large IoT systems need ephemeral timeseries.

Which you can store just fine in-memory in a normal data structure. And if you need advanced query capabilities or a query planner there is DuckDB. Using Redis seems like you get most of the disadvantages of having to run a whole database with few of the advantages.

your application can consist of multiple processes.

Or DuckDB