All posts written without LLM assistance unless otherwise noted.
Python is fast enough. Python programmers tend to understand the Python Cost Model, Python’s strengths and weaknesses, libraries that give compiled performance, and when to use a compiled language from the start.
So why do I care? Why do I get obsessed enough to coerce Claude into running these benchmarks and writing these Plotly charts? I do not know.1
But! I do know what I care about (for now) - and today (and some of the past weekend, and perhaps some of the next one), it’s definitely the ** cost of defining (ideally immutable) record types (AKA structs) in Python**.
So let’s get this out of the way: this write up is about benchmarking “Python type speed” (informally: compile-time), it is NOT about benchmarking
- serialization
- instantiation
- attribute access
- validation
- memory
Right, so that’s what Python programmers often care about, because they are probably working on long running programs, like apps, servers or pipelines, where the cost of defining a type is paid upfront, one time, whereas the cost of allocation, instantiation, validation, and serialization is paid repeatedly. So yeah, if that’s what you care about, this post is not for you.
But I did include
[instance cost]benchmarks if you’re curious. 😻
If you already know you care about type definition speed, then jump straight to the analysis, otherwise keep reading for my motivation and context on this subject.
#”how fast to --help
”
I tend to work on CLIs for developers and tooling for build systems or test suites where the time from program start to end is what we’re measuring. Perhaps you’ve noticed that running a command from a CLI may be near instant in a compiled program, but in Python, it can easily be hundreds of milliseconds: perceptible for UX, noticeable in CI/CD, and amplified by repeated calls as part of build system tooling.
Unlike in a compiled language, Python type definitions are not free (free in the sense that they were paid for during compilation ahem, Rust). They are code to be executed on every startup. And that includes imports of libraries and their type trees and dependents trees. We’ll see in the benchmarks that (evil-runtime-) metaprogramming, like decorators, metaclasses, or worse, have more of an upfront runtime type generation cost than manual type definitions.
Can we get the best of everything: a Pythonic type definition style, complete static typing and match, with the speed of a hand-written C struct, and the startup time of a compiled extension? I think so. seriously, I’m not sure, need to do more work, but I have good preliminary data
But why not use a compiled language and framework like Rust + clap? I certainly do, but what can I say? I love the Python ecosystem, build tooling libraries, and the rapidly evolving type system. And I believe that the type system can continue evolving so that we can offload a lot of the correctness to the type checker, and reap runtime speed benefits. That’s what this post is about.
#OK, OK, whatever, but why “structs”?
I’ll confess that I am an advocate of functional programming (FP), with little compromise. But the tortured kind, that can’t be bothered to learn Haskell, or study Lisp, and seems to end up rewriting the same handful of patterns in every language. So, it’s not the structs alone that I am after. It’s the sum types and pattern matching.
Long story short, I use sum types and pattern matching everywhere, all the time, from Rust to embedded C, from Typescript to Python, from JSON to CBOR. Even if your not an FP…enthusiast, you’ve likely used them in Python without thinking of them as such, when reaching for MyType | None
(an Option
or Maybe
type).
This example imagines that some immutable device info burned onto a ROM is versioned V1 and V2. V1 guaranteed presence of the serial number, but not the manufactured date. V2 guarantees both and adds a boot SHA.
from typing import NamedTuple
class DeviceInfoV1(NamedTuple):
serial_number: str
manufactured_utc_ms: int | None
class DeviceInfoV2(NamedTuple):
serial_number: str
manufactured_utc_ms: int
boot_sha: int
type DeviceInfo = DeviceInfoV1 | DeviceInfoV2
DeviceInfo
is a sum type of two product types,DeviceInfoV1
andDeviceInfoV2
, and there are only two representable states, each validated by the type system, not at runtime. Here’s what the naive product type would look like:
class DeviceInfo(NamedTuple):
serial_number: str
manufactured_utc_ms: int | None
boot_sha: int | None
Invalid runtime states are now possible: DeviceInfo(serial_number="abc", manufactured_utc_ms=None, boot_sha=123)
is a valid instance of the naive product type, but it is not a valid DeviceInfoV1
or DeviceInfoV2
.
Using a product type instead of a sum type shifts the burden of correctness from the type system to the runtime.
#Aw, f&*#!
I promised myself I wouldn’t evangelize FP (Day 583). 💀
It’s not really about FP, that just happens to be my motivation. There are plenty of different ways to utilize Abstract Data Types (ADTs) in Python, and if you care about Python startup time, then I think you’ll enjoy these benchmark results.
Besides, this can’t be about FP, because functional programmers don’t care about performance, memory, or know anything about compilers and instruction sets.
“Functional programming, strictly defined, is dumb…the way you manage mutable state is by making an entire copy of the data structure with the changes in the new copy of the data structure…here’s the problem: computers, they’re all bags of mutable state.”
Chris Lattner,[Creator Of Swift On Functional Programming (YouTube)]
Odd for the creator of LLVM, Clang, Swift, and Mojo to mischaracterize FP as anything other than an abstraction. I wasn’t aware of the “functional” instruction sets competing with x86 and ARM.
#WTF are we testing again?
I use NamedTuple
all the time, mostly because it means I don’t have to add @dataclass(frozen=true)
everywhere, but in the back of my mind I have always believed that NamedTuple
must be super efficient and compact, like const struct
in C or struct
in Rust. Once I realized that I’d been carrying on with this belief for years, I decided to setup this benchmark to understand how much I was truly paying for my types.
#THE CONTENDERS
*author’s commentary italicized to avoid bias
- manual python slotted class:
“Native Final Slots”* ewwwwwwwwwwww* - manual python slotted class (Brett Cannon’s manual
record-type
):“Manual Record Type”* oh god that’s even worse, this IS a waste of time, we’re going to turn Python into Java or something* - from Python’s standard library,
typing
module:NamedTuple
pewwwwww pew pew pewwwwwwoooo - also from the standard library,
dataclasses
:dataclass(frozen=True)
boooooooooo metaprogramming suuuuuuuuuuucks…unless it’s rust’s bs…or constexpr…at least it’s not C macros…but booooooooooooooooooo - from legendary Python core developer Br Br Bre Brett Cannon, iiiiiiiit’s
record-type
a new hope - from a 20 minute Claude hallucination that rips off
msgspec
andrecord-type
*WITH ATTRIBUTIONwait, I’m calling it!?!?!record-type (C)
hey, 20 minutes is not bad, it usually costs me $200 to get slop CPython! - weighing in at 11 years of development, the original
🎺attrs
medieval horns, but in tune🎺 - fast AF and only ~14.3% vowels iiiiiiiiiiiiiiit’s
msgspec
what is JSON for anymore?
#Structs Under Test (SUTs)
| Implementation | Description |
|---|---|
Final
-annotated fields, the closest thing to a naive native record.manual recordrecord-type
NamedTupletyping
moduledataclassdataclasses
modulefrozen dataclass@dataclass(frozen=true)
record-typerecord
type for Pythonrecord-type (C)record-type
and msgspec
attrsmsgspecEach of these implementations will be evaluated with and without mypyc compilation, and as a cold start (no bytecode cache) and warm start (bytecode cache present), when relevant. All of the implementations are tested on a struct type of three ints:
struct StructUnderTest {
a: int
b: int
c: int
}
Refer to the methodology section for details on how the benchmarks were run.
#Module Cost
When you import your base type or decorator, you also must pay a one time cost, regardless of how many types you define, for that module’s source tree. The cold import is roughly 6-8× a warm one, because the whole transitive source tree has to be recompiled to bytecode.
#Type Cost
So, how much does it cost to define a type? Remember that this cost is paid once on every program start, or at least when it is first imported.
Many of these benefit greatly from a warm start, which is the most common use of a Python program. Cold start is included because it’s the first impression that a user gets: “how fast to --help?”
Looking just at the warm start, we can start to see 3 performance tiers:
- ~7-12 µs:
native slots
,record-type (C)
,msgspec
, andmanual record
- ~76-96 µs (~8× slower):
NamedTuple
,record-type
- ~200-370 µs (~20-30× slower):
dataclass
,dataclass(frozen=True)
,attrs
The tiers come down to how many methods each implementation has to generate when the type is defined.
Use the table below to sort relative performance.
| implementation | |||
|---|---|---|---|
| 0.10× | 0.60× | 0.11× | |
| 0.11× | 0.35× | — | |
| 0.13× | 0.42× | — | |
| 0.15× | 2.1× | 0.18× | |
| 1.00× | 1.00× | 1.00× | |
| 1.3× | 1.2× | — | |
| 3.0× | 2.5× | 3.0× | |
| 4.0× | 3.2× | — | |
| 4.9× | 3.8× | 5.2× |
Per-type cost — each cell is × the baseline row (NamedTuple by default; click any row to re-base). Lower is faster to define. Click a column header to sort.
#So what’s the fastest startup?
Total startup () is calculated as the fixed dependency import (the module cost, ) plus the number of types () × the per-type definition cost ().
The interactive chart below shows the startup time on the Y axis and the number of types defined on the X axis. The scales can be toggled together between log (Y log10, X log2 from 1 to 4,096 types) and linear (Y clipped to 0ms–1,000ms, X from 0 to 4,096 types). For each implementation, the solid line is the warm time, and the dotted line is the cold time. Click on a name in the legend to toggle, double click to isolate, and double click on a disabled name to reset.
#Conclusion
For my purposes, I can draw a few conclusions from this.
NamedTuple
(my goto) is sorta in the middle and is probably not dragging start times too much. But, it’s per-type cost is ~8× the native/C implementations, so as the program grows, it will start to add up.msgspec
is faster thanNamedTuple
above ~256 (warm) type definitions. But this assumes absolutedependency discipline that negates some of the upsides of Python’s ecosystem. If you importmsgspec
, ordataclass
, anywhere, or if any of your dependencies have a high module or type cost, thenNamedTuple
’s low module cost is dwarfed and you may as well have started with a cheaper struct implementation.- The decorator-based implementations (
dataclass
,record-type
, andattrs
) all have a high type cost, but with that comes (evil-runtime-) metaprogamming capabilities. - The C implementation of record-type is good enough (wins by every metric) that I’ll be rewriting it and getting it under a test suite.
- I will update this article once I have a tested implementation!It may be too good to be true - I will definitely be trying out
msgspec
in the future. I wasn’t familiar with it before working on this report, but it’s very exciting to see these numbers, not to mention that it has de/serialization on top of being a basic struct. I’d love to seeCDDL/CBOR 🔥andpostcard ✉️de/serializers!
#Appendix
Here lives more stuff that wasn’t directly relevant to my goal of assessing startup time, but is still fun.
#Instance Cost
What can I say, since the benchmark suite was setup, I couldn’t resist. The instance costs are relevant to the program speed once it’s begun, and you’ll see that they are quite a bit tighter than the module and type cost comparisons. There’s a total spread of under 4x, from ~60ns up to ~220ns per instance.
#Construction
| implementation | ||
|---|---|---|
| 0.44× | — | |
| 0.45× | — | |
| 0.63× | 0.53× | |
| 0.63× | 0.77× | |
| 1.00× | 1.00× | |
| 1.5× | — | |
| 1.6× | 0.55× | |
| 1.6× | 1.6× | |
| 1.6× | — |
Per-instance construction cost — each cell is × the baseline row (NamedTuple by default; click any row to re-base). Lower is faster. Click a column header to sort.
#Memory
Memory is driven by object layout. Freezing a type never changes its footprint —
frozen=True
only changes the write path, not the storage. mypyc trades a few bytes per instance (one pointer to its method table, akin to a C++ vtable) for speed, 2 and gives every compiled class a fixed layout even without
__slots__
.
Immutability sometimes costs time or space and is never more efficient.
#native slots
A plain slotted class with Final
fields.
from typing import Final
class NativeFinal:
__slots__ = ("a", "b", "c")
def __init__(self, a: int, b: int, c: int) -> None:
self.a: Final = a
self.b: Final = b
self.c: Final = c
The Final
is for the static checker, meaning that it has zero runtime cost.4mypy
rejects o.a = 99
, but the assignment succeeds anyway, on the
interpreted class and the compiled .so
. So this is the closest thing to a
native record mypyc can produce — a compact slotted object (64 bytes; 72
compiled) whose __init__
it lowers to C-level slot stores, but it is not actually immutable at runtime (zero cost abstraction).
#manual record
native slots
is cheap precisely because it
does less. It has no __eq__
, __hash__
, or __repr__
, and — as we saw — it
isn’t even immutable. Every other record here gives you all of that. So here is
Brett Cannon’s record-type pattern: a complete, genuinely-immutable hand-written record with
__slots__
, __match_args__
, a real __setattr__
guard, and
__eq__
/__hash__
/__repr__
:
class ManualRecord:
__slots__ = ("a", "b", "c")
__match_args__ = ("a", "b", "c")
def __init__(self, a: int, b: int, c: int) -> None:
object.__setattr__(self, "a", a)
object.__setattr__(self, "b", b)
object.__setattr__(self, "c", c)
def __setattr__(self, _attr, _val):
raise TypeError("immutable")
def __eq__(self, other):
if not isinstance(other, type(self)):
return NotImplemented
return self.a == other.a and self.b == other.b and self.c == other.c
def __hash__(self):
return hash((self.a, self.b, self.c))
#record-type (C)
The manual record marks the pure-Python performance ceiling: complete and immutable, with
near-zero import, but either slow to construct (222 ns) or — once mypyc lowers its
object.__setattr__
init — fast (78 ns) yet larger (96 bytes). msgspec.Struct
shows C clears that ceiling: compact (64 bytes), immutable, ~62 ns construction,
~10 µs/type. Its one catch is the module cost. import msgspec
runs ~19 ms, because it’s a serialization library and you can’t get just the struct without importing the whole kitchen sink.5
Can you get msgspec’s record qualities without its import tax? A research prototype (read: LLM slop) on a branch of Brett Cannon’s record-type answers yes. It’s a ~600-(slop)-line C extension: an inheritable
Record
base you
subclass (subtype) exactly like NamedTuple
:
from native_record import Record
class Point(Record):
a: int
b: int
c: int
A C metaclass reads the class-body annotations directly (no inspect
, no
exec
) and builds a frozen, slotted type whose constructor is a C-level vectorcall, borrowing msgspec’s type-creation trick, with none of its codec machinery. And you saw in the charts above that it wins in every category.
#buuuuuuuuuuuuut…
It’s a research prototype, not a release. It lives on a
PR branch, not PyPI. And there’s one real semantic limit: a class body can’t
express Python’s full parameter grammar (positional-only, keyword-only, *args
,
**kwargs
) the way @record
’s function signature can — fine for the
record-shaped common case, but not literally 1:1 with the decorator. (Per-type
here is measured exactly like every other construct — module self-time ÷ K, which
includes the ~7 µs the bare class
statement costs regardless — so it is directly comparable to the figures above.)
#Why three type-cost tiers?
- fastest:
native slots
,record-type (C)
,msgspec
, andmanual record
- ~8× slower:
NamedTuple
,record-type
- ~20-30× slower:
dataclass
,dataclass(frozen=True)
,attrs
The single best predictor turned out to be how many methods each construct has to generate at class-creation: zero, one, or several. (Trace it yourself with codegen_probe.py, which captures every
exec
/ eval
/ compile
a single definition triggers.)#Tier 1 — nothing generated.
native slots
and manual record
are
hand-written, so their methods compile once into the .pyc
and the class
statement only has to build the type. msgspec
and record-type (C)
generate no Python either. A C metaclass assembles the type directly.
#Tier 2 — one generated method.
collections.namedtuple builds a
tuple
subclass — a descriptor per field and a single eval
’d __new__
:
lambda _cls, a, b, c: _tuple_new(_cls, (a, b, c))
with typing.NamedTuple adding
PEP 649annotation handling on top.
record-type
@record
inspect.signature
to read the fields, then one exec
’d
class whose only generated logic is the __init__
(__eq__
/ __hash__
/
__repr__
come from a Record
base):
class C(Record):
__slots__ = ('a', 'b', 'c')
def __init__(self, /, a, b, c) -> None:
object.__setattr__(self, 'a', a)
object.__setattr__(self, 'b', b)
object.__setattr__(self, 'c', c)
A metaclass-plus-factory and a decorator-plus-inspect
: different machinery, the same one-method’s-worth of work, the same tier.
#Tier 3 — several generated methods, plus field work and a rebuild.
dataclass turns the annotations into
Field
objects and generates __init__
, __repr__
,
and __eq__
in one shot (a factory that returns the three):
def __create_fn__(
__dataclass_type_a__,
__dataclass_type_b__,
__dataclass_type_c__,
__dataclass_HAS_DEFAULT_FACTORY__,
__dataclass_builtins_object__,
__dataclass___init___return_type__,
__dataclasses_recursive_repr
):
def __init__(
self,
a:__dataclass_type_a__,
b:__dataclass_type_b__,
c:__dataclass_type_c__
) -> __dataclass___init___return_type__:
self.a=a
self.b=b
self.c=c
@__dataclasses_recursive_repr()
def __repr__(self):
return f"{self.__class__.__qualname__}(a={self.a!r}, b={self.b!r}, c={self.c!r})"
def __eq__(self,other):
if self is other:
return True
if other.__class__ is self.__class__:
return self.a==other.a and self.b==other.b and self.c==other.c
return NotImplemented
return (__init__,__repr__,__eq__,)
frozen=True
adds three more: __setattr__
, __delattr__
, __hash__
— and
slots=True
creates the class a second time, since slots can’t be added in place. attrs is a more layered version of the same idea.
#NamedTuple in mypyc
I was really hoping that mypyc was going to compile NamedTuple to a native struct. Compiling the module changes almost nothing about the NamedTuple
, while it transforms native slots
:
| metric | NamedTuple interpreted | NamedTuple mypyc | native slots interpreted | native slots mypyc |
|---|---|---|---|---|
isinstance(_, tuple) |
yes | yes | no | no |
| bytes / instance | 88 | 88 | 64 | 72 |
__new__ (type) instructions |
7 bytecodes | 7 bytecodes | C | C |
__init__ (instance) instructions |
C | C | 9 bytecodes | C |
| instance (ns) | 138 | 142 | 87.5 | 75.7 |
The NamedTuple columns are identical: same footprint, same construct time. Its __new__
is still seven interpreted bytecodes inside the compiled extension module, building a tuple and handing it to tuple.__new__
:
1 RESUME 0
LOAD_GLOBAL 1 (_tuple_new + NULL)
LOAD_FAST_BORROW_LOAD_FAST_BORROW 1 (_cls, a)
LOAD_FAST_BORROW_LOAD_FAST_BORROW 35 (b, c)
BUILD_TUPLE 3
CALL 2
RETURN_VALUE
Contrast the native record. It has no __new__
at all; its __init__
writes
the three fields straight into their slots with STORE_ATTR
— no tuple, no
length field, no boxed item array. (The Final
annotations add zero bytecode; they’re a pure type-checker hint, so this is byte-for-byte a plain slotted class.)
11 RESUME 0
12 LOAD_FAST_BORROW_LOAD_FAST_BORROW 16 (a, self)
STORE_ATTR 0 (a)
13 LOAD_FAST_BORROW_LOAD_FAST_BORROW 32 (b, self)
STORE_ATTR 1 (b)
14 LOAD_FAST_BORROW_LOAD_FAST_BORROW 48 (c, self)
STORE_ATTR 2 (c)
LOAD_CONST 0 (None)
RETURN_VALUE
mypyc does lower this __init__
to C — recall its 9 bytecodes became C-level in
the compiled column. But for this record you barely see it in the construction
numbers (87 → 76 ns, within run-to-run noise): the __init__
is only three STORE_ATTR
s, and the interpreted timeit
loop crosses the
interpreter↔native boundary on every call, which caps any gain. Where compiling a
hand-written __init__
does pay off is when it does real interpreted work —
manual record routes every
field through object.__setattr__
and drops from 222 to 78 ns once compiled, a
speedup a frozen dataclass
can’t get. NamedTuple’s __new__
, by contrast, stays interpreted even when compiled and there’s nothing for mypyc to lower at all without breaking the tuple contract.
So, I’ve been right to reach for NamedTuple
as a cheaper immutable type than dataclass(frozen=True)
, but I was wrong to think that it was perfectly efficient and compact like a C struct.
#Further reading
A first-class record type for Python. Brett Cannon’srecord-type proposal(and a terserstruct Point(x: int, y: int)
spelling), with the proof-of-conceptrecord
decorator already on PyPI. As proposed it standardizes the boilerplate — a concise frozen, slotted dataclass — rather than adding a performance primitive: a decorator’s generated__init__
stays interpreted, so it can’t push past the pure-Python floor themanual recordmaps out.Unboxed value types in mypyc.mypyc#841tracks the performance angle these benchmarks can’t reach: user-definedunboxedvalue types (≈16 bytes vs 40 for a heap object), passed around in native code and boxed only when they enter a Python container. mypyc already does this for native integers (i64
/i32
) — just not yet for user-defined records. Open since 2021 with no implementation: a direction, not a date, and nothing to benchmark yet.
#Methodology
All measurements were taken on a single machine: CPython 3.14.0 (installed and
managed with uv), mypy/mypyc 2.1.0, attrs 26.1.0,
msgspec 0.21.1, and record-type 2023.1.post1, on x86_64 Linux (WSL2) with gcc
13.3. The C-backed record-type (C)
is built from the
branch linked above (a
research prototype, not a release). Absolute numbers will differ on your hardware
and Python build; the relative shape is the takeaway. Every struct carries the
same three int
fields.
#Interpreted vs compiled
The standard-library constructs (plain classes, slotted, Final
-slotted,
NamedTuple
, and the dataclass variants) live in one module that is the unit of
compilation: mypyc containers.py
produces a containers.*.so
. An interpreted
driver imports that module and detects which form it got by testing whether
__file__
ends in .so
. This mirrors how mypyc is actually used — you compile
the definitions and call into them from ordinary interpreted code. The attrs
,
msgspec
, and both record-type
classes are defined in the driver itself, not in
the compiled module, so there is no mypyc form to measure — the charts and tables
leave their mypyc column empty rather than copy in the interpreted value.
(record-type (C)
is already a compiled C extension, so mypyc has nothing to add — it is the native form.)
Even inside the compiled .so
, the @dataclass
decorator and the NamedTuple
metaclass run as interpreted CPython, and the __init__
/ __new__
they generate stay interpreted bytecode: mypyc compiles the module’s own code, not the code those tools synthesize at runtime.
#Memory footprint
sys.getsizeof
reports one object’s size but doesn’t follow the __dict__
pointer, so it understates classes that carry one. 6 The headline figures instead come from a bulk
tracemalloc
measurement — allocate 200,000 instances and
subtract a same-length [None] * n
list measured the same way, so the list’s own backing storage cancels and what remains is the instances’ allocation (GC header included):
import gc, tracemalloc
def mem_per_instance(ctor, args, n=200_000):
gc.collect()
tracemalloc.start()
base = [None] * n
base_cur, _ = tracemalloc.get_traced_memory()
objs = [ctor(*args) for _ in range(n)]
cur, _ = tracemalloc.get_traced_memory()
tracemalloc.stop()
return (cur - base_cur) / n
Treat the per-instance figure as ±one allocator alignment word.
#Bytecode
Allocation bytecode is counted with dis.get_instructions
on __new__
and
__init__
(unwrapping the staticmethod
that wraps a NamedTuple’s __new__
),
and disassembled with dis.dis
for the listings above. Deallocation has no
Python bytecode to count: teardown is C-level tp_dealloc
/ tp_free
unless a
class defines a Python __del__
, which none of these do.7
#Per-instance timing
Construction and attribute access are timed with timeit — the minimum of seven repeats of 1,000,000 iterations for construction, 5,000,000 for access, reported as nanoseconds per operation.
The
8timeit
loop is interpreted, so every iteration crosses the interpreter↔native boundary. mypyc’s attribute-access and call speedups land on the compiled→compiledpath, so an interpreted loop reaching into a compiled class won’t see them (and can read slightly slower) — which is why the compiled instantiation numbers sit on top of the interpreted ones rather than below.
#Import / type-construction time
The obvious approach — timeit
on make_dataclass()
or namedtuple()
—
measures the wrong thing. The dynamic factory forms differ from the @dataclass
and class C(NamedTuple)
forms you actually write (the functional NamedTuple(...)
call understates the class-statement form by roughly 3×), and timeit
is blind to both mypyc and the one-time cost of importing the supporting library, since those happen before the loop starts.
So every import number comes from a fresh interpreter under
python -X importtime
, reading the self time attributed to the module — self time excludes child imports, so the supporting library isn’t double-counted:
Per-type cost. Generate a module of K = 200 identical-shape classes in the real class-statement form, import it, and read its self-time; the per-type figure is that self-time ÷ 200, the median of five fresh interpreters (this is what the committedimporttime_sweep.py
reports). Dividing by K folds a small fixed per-module overhead into each figure. What that per-type costconsistsof — the methods each construct generates at class-creation — is dissected inWhy three type-cost tiers.Cold vs warm.“Warm” imports with the__pycache__/*.pyc
already written; “cold” deletes__pycache__
first, so the source is recompiled to bytecode in-process. Their difference is the source→bytecode compile cost (tens of µs/type — ~25–55 here, scaling with each class’s source size).Dependency import.python -X importtime -c "import LIB"
in a fresh interpreter gives the cumulative cost of first-importing a library. The cold variant pointsPYTHONPYCACHEPREFIX
at an empty directory so the whole transitive source tree must recompile.mypyc axis. The generated module is compiled withmypyc
and the resulting.so
imported under the same harness. A compiled extension has no Python source to recompile, so there’s no cold/warm gap — yet its per-type creation cost is barely lower than interpreted, likely because type creation is dominated by CPython’sPyType_Ready
, which runs either way.
#The crossover model
The startup chart is a model, not a direct measurement: total startup is taken as
a fixed dependency import plus N × the measured per-type construction cost,
evaluated for cold and warm. The crossover is where two such lines meet —
N = (dep_b - dep_a) / (per_type_a - per_type_b)
. It assumes a single dependency imported once and a linear per-type cost (both hold well here); the cold curves roll up shared sub-dependencies, so several of these libraries imported together cost less than the sum of their individual lines.
#Reporting
Bytes and counts are integers; timing data is quoted to three significant figures.
Import timings vary run to run, so each is reported as the median of five fresh
processes; instantiation is the minimum of seven timeit
repeats (the conventional low-noise estimator). Treat the per-instance nanosecond figures as ±10% — the construct-to-construct shape is what’s robust, not the third digit.
#Limitations and cross-validation
One machine, no isolation. Everything ran on a single WSL2 host — which sits on Hyper-V, as does the Windows install beside it, so there’s no bare-metal baseline on this box (and no WSL2-specific virtualization penalty to factor out either)— with no CPU pinning or frequency-scaling control. Repeating on separate hardware, several Python versions, and a second OS would confirm the shape; pinning the CPU steadies the absolute numbers.9Compiled construction is timed from an interpreted loop. That measures the common interpreted-caller-into-compiled-class case, not compiled→compiled throughput. A benchmark loop itself compiled with mypyc would show whether its call and attribute speedups close the gap.“Cold” is a cold The source stays in the OS page cache between runs, so the cold figures isolate source→bytecode compilation, not first-read I/O.bytecodecache, not a cold disk.Per-type cost is self-time ÷ K. That folds a small fixed per-module overhead into each figure; a regression over several values of K would separate the fixed cost from the per-type slope (the correction is sub-microsecond for the cheap constructs).There is no struct-only import to isolate — the codec comes with it — so it’s a fair number to report but not a pure struct-definition cost.msgspec
’s ~19 ms import is library-wide.5Its numbers may shift once it’s hardened and packaged.record-type (C)
is a research prototype.Five runs is modest. More repeats, and reporting dispersion alongside the median, would tighten the import figures.
#Reproducing
Everything here is reproducible from the python-struct-profiling repository — the data in this post was produced at commit
b2f2eb7
int
-field shape:bench.py
— memory (tracemalloc
), bytecode (dis
), and instantiation (timeit
), run once against the interpreted module and once against themypyc
-compiledcontainers.so
.importtime_sweep.py
— the import / type-creation axis: it generates a module of K real class-statement / decorator forms per construct, imports it underpython -X importtime
in a fresh interpreter, and divides the module self-time by K. The figures here are--k 200 --runs 5
.codegen_probe.py
(added at) — the mechanism behind thed8acfd5
three type-cost tiers: it traces theexec
/eval
/compile
each construct runs at class-creation and counts how many methods each one generates (zero, one, or several).
#Raw data
Every figure above is derived from this one table set (the charts and these tables read the same array, so they cannot disagree):
Table 1 — Import / type-creation cost, µs per class (median of 5 fresh
-X importtime
runs, K = 200). mypyc is the compiled .so
;
“—” means the construct is off the compiled axis (attrs, msgspec, and both record-types are
defined outside the compiled module; record-type (C)
is already a C extension).
| construct | variant | warm | cold | mypyc |
|---|---|---|---|---|
| native slots | mutable | 7.3 | 59.3 | 6.9 |
| native slots | frozen | 7.4 | 62.2 | 6.9 |
| manual record | frozen | 11.5 | 214.5 | 11.1 |
| NamedTuple | frozen | 76.2 | 104.3 | 63.3 |
| dataclass | mutable | 228.4 | 261.0 | 190.3 |
| dataclass | frozen | 373.4 | 401.2 | 328.5 |
| record-type | frozen | 96.4 | 122.4 | — |
| record-type (C) | frozen | 8.6 | 36.0 | — |
| attrs | mutable | 264.6 | 288.7 | — |
| attrs | frozen | 301.4 | 332.2 | — |
| msgspec | mutable | 10.5 | 40.1 | — |
| msgspec | frozen | 10.2 | 44.0 | — |
Table 2 — One-time dependency import, milliseconds cumulative in a fresh interpreter. Paid once per process regardless of how many types you define. The native record imports no library.
| library | warm | cold |
|---|---|---|
| native (none) | 0.0 | 0.0 |
| manual (none) | 0.0 | 0.0 |
| typing | 4.0 | 33.9 |
| dataclasses | 11.5 | 81.9 |
| record-type | 12.5 | 91.3 |
| record-type (C) | 0.2 | 0.2 |
| attrs | 22.2 | 128.5 |
| msgspec | 19.1 | 131.7 |
Table 3 — Per-instance memory, bytes (tracemalloc, GC header included). Freezing never changes the footprint; mypyc adds one 8-byte vtable word to the native classes it compiles.
| construct | variant | interpreted | mypyc |
|---|---|---|---|
| native slots | mutable | 64 | 72 |
| native slots | frozen | 64 | 72 |
| manual record | frozen | 64 | 96 |
| NamedTuple | frozen | 88 | 88 |
| dataclass | mutable | 64 | 72 |
| dataclass | frozen | 64 | 72 |
| record-type | frozen | 64 | — |
| record-type (C) | frozen | 64 | — |
| attrs | mutable | 80 | — |
| attrs | frozen | 80 | — |
| msgspec | mutable | 64 | — |
| msgspec | frozen | 64 | — |
Table 4 — Instantiation, nanoseconds (min of 7 timeit repeats of 1e6 iterations). The timeit loop is interpreted, so a compiled class called from it shows no mypyc speedup — and can read noticeably slower from the per-call interpreter↔native boundary (e.g. mutable dataclass 87.5→109.5). Treat these as ±10%; the construct-to-construct shape is the robust signal, not small interpreted-vs-mypyc deltas.
| construct | variant | interpreted | mypyc |
|---|---|---|---|
| native slots | mutable | 87.3 | 75.2 |
| native slots | frozen | 87.5 | 75.7 |
| manual record | frozen | 222.5 | 78.4 |
| NamedTuple | frozen | 138.3 | 141.5 |
| dataclass | mutable | 87.5 | 109.5 |
| dataclass | frozen | 224.3 | 226.0 |
| record-type | frozen | 227.0 | — |
| record-type (C) | frozen | 61.2 | — |
| attrs | mutable | 88.5 | — |
| attrs | frozen | 209.1 | — |
| msgspec | mutable | 63.0 | — |
| msgspec | frozen | 62.5 | — |
Table 5 — Construction bytecode, instruction counts from dis
.
“C” = no Python bytecode (C-level). Freezing is what turns the 9-instruction
__init__
into 25 (every field routed through object.__setattr__
);
these counts are unchanged inside the compiled module except the native
__init__
, which mypyc lowers to C.
| construct | __new__ |
__init__ (mutable) |
__init__ (frozen) |
|---|---|---|---|
| native slots | C | 9 | 9 |
| manual record | C | — | 24 |
| NamedTuple | 7 | — | C |
| dataclass | C | 9 | 25 |
| record-type | C | — | 24 |
| record-type (C) | C | — | C |
| attrs | C | 9 | 25 |
| msgspec | C | C | C |
Derived: the NamedTuple ↔ msgspec startup crossover sits at 229 types (warm) and 1,622 types (cold), computed from Tables 1 and 2.
#Footnotes
If you have any ideas, please LMK so I can explain it to my family.
↩ -
“Introduction”. mypyc.readthedocs.io. Retrieved 2026-06-21. “Classes are compiled toC extension classes. They use vtables for fast method calls and attribute access.”↩ -
“Native classes”. mypyc.readthedocs.io. Retrieved 2026-06-21. “Only attributes defined within a class definition (or in a base class) can be assigned to (similar to using__slots__
).”↩ -
“typing.Final”. docs.python.org. Retrieved 2026-06-21. “There is no runtime checking of these properties.” (See alsoPEP 591.)↩ -
. github.com/jcrist/msgspec. Retrieved 2026-06-21.src/msgspec/__init__.py
Struct
is imported from the compiled._core
extension, and importing the package eagerly runsfrom . import inspect, json, msgpack, structs, toml, yaml
; the codecs injson.py
/msgpack.py
re-export from that same_core
, so there is no struct-only import to isolate.↩↩2 -
“sys.getsizeof”. docs.python.org. Retrieved 2026-06-21. “Only the memory consumption directly attributed to the object is accounted for, not the memory consumption of objects it refers to.”↩ -
“tp_dealloc”. docs.python.org. Retrieved 2026-06-21. “A pointer to the instance destructor function. […] free all memory buffers owned by the instance, and call the type’stp_free
function to free the object itself.”↩ -
“timeit”. docs.python.org. Retrieved 2026-06-21. The module “provides a simple way to time small bits of Python code”; the minimum is reported because “the lowest value gives a lower bound for how fast your machine can run the given code snippet; higher values in the result vector are typically not caused by variability in Python’s speed, but by other processes interfering with your timing accuracy. So themin()
of the result is probably the only number you should be interested in.”↩ - “Comparing WSL Versions”. learn.microsoft.com. Retrieved 2026-06-21. “WSL 2 is running as a Hyper-V virtual machine.” The Windows host beside it is itself a partition on that same hypervisor —“Hyper-V Architecture”: “The Microsoft hypervisor must have at least one parent, or root, partition, running Windows … [which] has direct access to hardware devices.”↩
© 2026 by JP Hutchins. Published under a Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.