Python: Are Kernel and Chat Services thread-safe?

[abdhx]

Hi team 👋 ,

I'm integrating Semantic Kernel into a Python project that involves multi-threaded environments (e.g., FastAPI server, background workers). I need to know if the following components are thread-safe and suitable to be shared across threads:

• semantic_kernel.Kernel
• Chat services (e.g., OpenAIChatCompletion, AzureChatCompletion)
• Any related memory/planner/skill components

I couldn't find definitive info in the docs or issues. Any clarification or best practice guidance would be appreciated.

Thanks!

[moonbox3]

Hi @abdhx, thanks for your question.

Here are some thoughts:

• Kernel: safe to share read-only. It holds dicts of plugins/services/filters without locks, so don’t mutate those (add/remove) while other threads/tasks are using it. For per-request tweaks, use kernel.clone(); note that it deep-copies filters/plugins but reuses the same service clients.
• Chat services: generally fine for concurrent, non-streaming calls. Don’t mutate shared config while in use; copy execution settings per call if you modify them. Keep streaming iterators within the same task/thread that created them. Don’t pass streams across threads.
• Memory/Planner/Skills: (Note: we do not recommend using any built-in planner. We've updated our docs to reflect this and you won't find planners in our code base ay more. Skills are an old terms and have been replaced with Plugin. If one asks an LLM about SK it may still refer to things as Skills, FYI.) Stateless pieces are fine to share. Anything with in-memory, mutable state isn’t thread-safe without your own synchronization. If backed by external stores (Redis/Postgres/Vector DB), rely on the driver’s thread-safety and prefer connection pooling.

Recommended pattern (FastAPI/background workers)

• Create and configure a Kernel and service clients once at process startup; treat them as read-only singletons.
• Per request: create fresh ChatHistory and KernelArguments. If you need request-scoped filters or minor config changes, clone the Kernel and modify the clone’s filters/settings (or copy settings dicts).
• Keep streaming consumption in the same task/thread.

[moonbox3]

I will close this issue as I don't see any further action required. Please ping back on this if you have further questions, or open a new issue. Thanks.

[abdhx]

Thank you for the detailed and well-explained answer