Consider lazy durable entity state serialization semantics

Summary

Durable entity state is currently serialized eagerly when set_state() is called. This snapshots the value immediately, so mutations made to the same Python object after set_state() do not affect the state eventually persisted at the end of the operation.

state = Point(1)
self.set_state(state)
state.value = 99

The persisted state is Point(1), not Point(99).

This may surprise Python users who expect ordinary reference semantics and could introduce subtle bugs when entity code mutates an object after passing it to set_state().

Scope

The behavior is implemented in the core durabletask entity StateShim, so it affects core durabletask entities and providers built on it, including durabletask-azuremanaged and azure-functions-durable.

Current rationale

Eager serialization was introduced intentionally to:

• surface serialization errors inside the operation that called set_state();
• preserve per-operation rollback behavior in an entity batch;
• retain the original serialized wire payload when state is unmodified; and
• prevent mutations to values returned by get_state() from implicitly changing persisted state without a corresponding set_state().

Design question

Should set_state() retain the live value and defer serialization until successful operation completion, so mutations made after set_state() but before the operation returns are persisted?

If so, the design needs to define:

• whether get_state() after set_state() returns the same pending live object or a reconstructed copy;
• where serialization occurs so serialization failures still fail and roll back the current operation;
• behavior across multiple operations in one entity batch;
• compatibility implications of changing set_state() from snapshot-at-call semantics to retain-reference-until-commit semantics; and
• consistency with Durable Entity semantics in other language SDKs.

Suggested investigation

• Compare mutation semantics with other Durable Entity SDKs.
• Prototype deferred serialization at the operation commit boundary.
• Add tests for post-set_state() mutation, serialization failure, rollback, deletion, unmodified wire payloads, and multiple operations in one batch.
• Document the current requirement to perform mutations before the final set_state() call unless the behavior changes.

Bắt đầu với thực thể Durable Task cốt lõi StateShim và theo dõi cách set_state(), get_state(), việc hoàn tất operation, rollback và deletion hiện xử lý state đã serialize. So sánh ngữ nghĩa mutation với các Durable Entity SDKs khác trước khi quyết định liệu deferred serialization có phù hợp hay không. Bổ sung coverage cho mutation sau set_state, serialization failure, rollback, deletion, wire payload không thay đổi và nhiều operation trong một batch.