Pydantic drops Python 3.10, then layers serialization fast paths

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

Wednesday's 3.10 removal is the week's one breaking change; the rest of the period is serialization and schema-generation speedup work.

99 commits 78 PRs merged 3 releases 7 briefings covered

all fastapi reviews →

← previous week fastapi wire →

Improve performance of model serialization pydantic/pydantic

by Viicos

Three changes to the fields serializer used by models, dataclasses and TypedDicts, pulled from maxisbey/pydantic#16: no second __dict__ lookup for models serialized outside unions, and fewer dict-path checks. This is throughput on the hot path every response and cache write traverses.

Improve performance of dict serialization pydantic/pydantic

by Viicos

Stacked on the fields serializer work, this skips the include/exclude filter logic, and the key hashing that goes with it, when there is nothing to filter by. Models and dicts serialized without field filters stop paying for a decision that always resolves the same way.

Improve performance of type inference during serialization pydantic/pydantic

by Viicos

For values whose type is inferred during serialization (Any, dict[str, Any]), the recursion guard is dropped for None, bool, int and str, which cannot be part of a reference cycle. Loose-typed payloads, the ones most code defaults to at boundaries, get the benefit.

Improve schema generation for fixed type handlers pydantic/pydantic

by alexmojaki

Fixed no-argument core schemas for builtin scalars, object, date/time types, Decimal, UUID, Url, Fraction, MultiHostUrl and the sentinels now build once and return a shallow copy. Schema construction for plain field types gets cheaper on every model definition.

Skip general schema dispatch for builtin scalars pydantic/pydantic

by alexmojaki

Builds on that, skipping origin lookup, hook inspection and general dispatch for exact immutable types in generate_schema() and the annotation inner handler. Custom hooks still work; scalar fields just stop paying the dispatch tax first.

Fix `ByteSize.human_readable()` showing `1024.0KiB` instead of `1.0MiB` just below a unit boundary pydantic/pydantic

by Viicos

human_readable() returned 1024.0KiB for values just under a unit boundary instead of 1.0MiB. Cosmetic unless byte sizes reach logs or metrics, where the off-by-scale string is the thing humans read.

Fix JSON Schema for `arguments` schema pydantic/pydantic

by Viicos

Corrects the generated JSON Schema for arguments, tracked in issue #13938. If you publish tool or function schemas built from Pydantic models, regenerate them: the old output was wrong, not merely different.

Benchmark `TypedDict` schema generation across builds pydantic/pydantic

by alexmojaki

Alongside: a CodSpeed benchmark for two independently built child TypedDicts sharing a leaf type, benchmarks for serializing large strings to JSON, a Ruff bump to 0.17.3, entry points API in the plugin loader, and FastAPI bumping pillow to 13.0.0 and dnspython to 2.9.2.

$ ls fastapi/week/ # the briefings behind this review

Mon Nov 2 Serialization hot paths trimmed, ByteSize human_readable fixed at boundaries Tue Nov 3 Pydantic caches fixed core schemas, and its schema path skips dispatch for builtin scalars Wed Nov 4 Pydantic targets schema-build fast paths, adds large-string JSON benchmarks Thu Nov 5 Pydantic measures duplicate TypedDict schema cost, FastAPI drops conference rail Fri Nov 6 FastAPI bumps pillow and dnspython, Pydantic corrects arguments JSON Schema Sat Nov 7 Pydantic reworks `TypeAdapter` call arguments, tightens `Decimal` schemas Sun Nov 8 Pydantic drops Python 3.10, accepts list and DecimalTuple for Decimal

all fastapi reviews →

← previous week fastapi wire →