perf(http-client-python): precompute model class state to speed up list operations

[l0lawrence]

Summary

Profiling azure-storage-blob's list_blobs against a 500-blob container on the in-progress TypeSpec migration (Azure/azure-sdk-for-python#45133) identified the generated _utils/model_base.py deserializer as the dominant hot path – roughly half the wall time was spent there, causing a ~-48% throughput regression vs. the current msrest-generated package. Download/upload were unaffected because they deserialize exactly one model per HTTP call regardless of payload size; list_blobs is the only operation where one response produces N model instances, so per-instance overhead is what matters.

This PR addresses the most mechanical of those hot spots in the shared model_base.py.jinja2 template so every TypeSpec-emitted Python SDK benefits.

Changes

1. Precompute cls._defaults once in Model.__new__ – Model.__init__ no longer walks _attr_to_rest_field on every instance build; it just copies the pre-built mapping.
2. Promote _RestField._rest_name from @property to plain attribute – set once in __new__; removes a descriptor lookup that was hit once per field per model (~73k calls for a 500-blob list).
3. Replace the string-keyed _calculated set with a _calculated_done: bool flag – checked via cls.__dict__.get("_calculated_done", False) so subclasses re-run the per-class setup correctly without inheriting the flag.
4. Narrow _deserialize_default's except Exception to except DeserializationError – the caller _deserialize_with_callable already wraps everything into DeserializationError, so this is semantically equivalent for the success-and-expected-failure paths while surfacing real coding bugs (AttributeError/TypeError) instead of masking them as deserialization failures.

No public API change; no generated-code change for consumers beyond the _utils/model_base.py body itself.

Validation

• npm run build – clean.
• npm run regenerate – unbranded fixtures regenerate successfully. Regenerated _utils/model_base.py contains the new code paths (_calculated_done, _defaults =, except DeserializationError).
• Unit tests: tests/unit/test_model_base_serialization.py + test_model_base_xml_serialization.py – 156/157 pass. The one failure (test_null_serialization) is a pre-existing cross-package isinstance mismatch (azure.core.serialization._Null vs. corehttp.serialization._Null) in the unchanged _deserialize_with_callable dispatcher and is unrelated to this change – it reproduces on main when both azure.core and corehttp are importable simultaneously in the test env.

Follow-ups (not in this PR)

There are larger wins still on the table that I'd like to address in a second PR if this one is well-received:

• Cache per-class XML metadata (xml_name, wrapped, attribute) so _init_from_xml stops recomputing it once per field per instance.
• Cache the typing.get_args / annotation analysis in _deserialize that currently re-runs for every instance.
• Flatten the 13-branch _deserialize_with_callable dispatcher into a precomputed callable per field.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@typespec/http-client-python@10475

commit: 021e589

[github-actions]

All changed packages have been documented.

• ✅ @typespec/http-client-python

Show changes

@typespec/http-client-python - fix ✏️

Speed up model deserialization for list-style operations by precomputing per-class defaults and rest-name metadata once in __new__, turning _RestField._rest_name and Model._calculated_done into plain attribute reads instead of a @property and a string-keyed set lookup. Also narrows a broad except Exception in _deserialize_default to except DeserializationError so real coding bugs are no longer silently swallowed.

[azure-sdk]

You can try these changes here

🛝 Playground	🌐 Website	🛝 VSCode Extension

[microsoft-github-policy-service]

Hi @@l0lawrence. Your PR has had no update for 30 days and it is marked as a stale PR. If it is not updated within 30 days, the PR will automatically be closed. If you want to refresh the PR, please remove the stale label.