perf(bt_json): optimize _to_bt_safe and bt_safe_deep_copy hot paths by AbhiPrasad · Pull Request #139 · braintrustdata/braintrust-sdk-python

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) · 2026-03-25T00:16:50Z

Extracting just the json serialization changes from #101. Validated with benchmarking suite.

Add primitive fast-paths using type identity checks before expensive
isinstance calls against abstract classes and Pydantic model_dump.
Rewrite _deep_copy_object to use type(v) is dict/list instead of
isinstance(v, Mapping), inline primitive checks, and cache
visited set methods. Guard Pydantic model_dump/dict with hasattr
to avoid warnings overhead on non-Pydantic values.

Benchmark results (geometric mean: 8.49x faster):

_to_bt_safe:
primitive-int 2.02 us -> 42.2 ns 47.8x faster
primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster
str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster
dataclass 10.2 us -> 3.32 us 3.1x faster
pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise)
pydantic-v1-like 1.94 us -> 841 ns 2.3x faster
bt_safe_deep_copy:
small 19.7 us -> 2.30 us 8.6x faster
medium 179 us -> 17.3 us 10.4x faster
large 1.66 ms -> 138 us 12.1x faster
circular 178 us -> 17.9 us 9.9x faster
non-string-keys 16.7 us -> 2.12 us 7.9x faster

Add primitive fast-paths using type identity checks before expensive isinstance calls against abstract classes and Pydantic model_dump. Rewrite _deep_copy_object to use type(v) is dict/list instead of isinstance(v, Mapping), inline primitive checks, and cache visited set methods. Guard Pydantic model_dump/dict with hasattr to avoid warnings overhead on non-Pydantic values. Benchmark results (geometric mean: 8.49x faster): _to_bt_safe: primitive-int 2.02 us -> 42.2 ns 47.8x faster primitive-float-nan 2.01 us -> 67.8 ns 29.7x faster str-subclass-enum 2.02 us -> 97.5 ns 20.8x faster dataclass 10.2 us -> 3.32 us 3.1x faster pydantic-v2-like 1.68 us -> 1.78 us 1.1x slower (noise) pydantic-v1-like 1.94 us -> 841 ns 2.3x faster bt_safe_deep_copy: small 19.7 us -> 2.30 us 8.6x faster medium 179 us -> 17.3 us 10.4x faster large 1.66 ms -> 138 us 12.1x faster circular 178 us -> 17.9 us 9.9x faster non-string-keys 16.7 us -> 2.12 us 7.9x faster

Cedric / ViaDézo1er (viadezo1er)

Tested on a small Linux machine, the speedup ratio is the same

Abhijeet Prasad (AbhiPrasad) requested a review from Cedric / ViaDézo1er (viadezo1er)March 25, 2026 00:16

Abhijeet Prasad (AbhiPrasad) self-assigned this Mar 25, 2026

Abhijeet Prasad (AbhiPrasad)force-pushed the abhi-perf-optimize-bt-json-hot-paths branch from 858df3c to 7945ab6Compare March 25, 2026 00:25

Abhijeet Prasad (AbhiPrasad) marked this pull request as ready for review March 25, 2026 00:25

Cedric / ViaDézo1er (viadezo1er) approved these changes Mar 25, 2026
View reviewed changes

Abhijeet Prasad (AbhiPrasad) merged commit 01d0cea into mainMar 25, 2026
34 checks passed

Abhijeet Prasad (AbhiPrasad) deleted the abhi-perf-optimize-bt-json-hot-paths branch March 25, 2026 17:28

This was referenced Mar 25, 2026
perf: optimize tracing hot paths (5x faster user thread) #101
Closed
chore: Bump SDK version to 0.11.0 #148
Merged