In a pinch one way to deal with memory leaks is to wrap the offending work in a multiprocessing pool, forcing it to live in an ephemeral process
Indeed, I introduced this [0] as an option for every Python integration that runs on the Datadog Agent.
[0]: https://github.com/DataDog/integrations-core/pull/12986