Skip to content

Speed up float encoding in json #159702

Open
Open
Speed up float encoding in json#159702
Assignees
corona10
Labels
extension-modulesC modules in the Modules dirperformancePerformance or resource usagetype-featureA feature request or enhancement

Description

@nineteendo
nineteendo
opened on Oct 19, 2026

Feature or enhancement

I ran a throughput benchmark for different json libraries.

Float encoding is a particularly weak area for the json library, especially for large-magnitude floats. Would it be possible to optimize this by adding a special case to PyUnicodeWriter_WriteRepr(), similar to _PyLong_FormatWriter() for integers?

encode json jsonyx fastsimdjson msgspec orjson fastest time
0.0 12.68x 14.10x 1.32x 1.00x 1.11x 7.32 ns
1e-05 13.05x 13.47x 1.12x 1.81x 1.00x 29.43 ns
5e-324 25.71x 26.07x 1.07x 1.00x 1.33x 24.83 ns
1.7976931348623157e+308 79.89x 80.30x 1.00x 1.28x 1.03x 30.29 ns

cc @vstinner

Has this already been discussed elsewhere?

This is a minor feature, which does not need previous discussion elsewhere

Links to previous discussion of this feature:

No response

Linked PRs

Activity

  1. nineteendo
    added
    type-featureA feature request or enhancement
    on Oct 19, 2026
  2. StanFromIreland
    added
    performancePerformance or resource usage
    extension-modulesC modules in the Modules dir
    on Oct 19, 2026
  3. StanFromIreland
    added this to JSON issueson Oct 19, 2026

  4. eendebakpt commented on Oct 19, 2026

    @eendebakpt
    eendebakpt
    on Oct 19, 2026
    Contributor

    A fast path in PyUnicodeWriter_WriteRepr seems reasonable, but would not help a lot (I did a quick benchmark, only about 10% depending on the type of floats). A faster float-to-str would help more. There are algorithms for it, but not readily available in C. See https://discuss.python.org/t/faster-float-to-string-conversions/107397

  5. corona10
    self-assigned this
    on Oct 19, 2026

  6. corona10 commented on Oct 19, 2026

    @corona10
    corona10
    on Oct 19, 2026
    Member

    @eendebakpt @nineteendo

    @vstinner and I discussed this at the sprint that we are going to measure general benchmark how much this call paths are executed and adding the fast path.

  7. vstinner commented on Oct 19, 2026

    @vstinner
    vstinner
    on Oct 19, 2026
    Member

    I added the int fast-path in #125214 which contains benchmark results: it was between 1.54x and 1.93x faster on repr(list) depending on the list size (between 3 and 50 items).

    FYI I added the str fast-path in #133969 which also contains benchmark results (1.18x faster in average, up to 1.61x faster in the best case).

  8. vstinner commented on Oct 19, 2026

    @vstinner
    vstinner
    on Oct 19, 2026
    Member

    In the past, I considered to add a type slot (like PyTypeObject.tp_str_writer and/or PyTypeObject.tp_repr_writer) instead of using hardcoded fast paths. If we add multiple fast paths in the future, we might reconsider adding a type slot?

  9. corona10 commented on Oct 19, 2026

    @corona10
    corona10
    on Oct 19, 2026
    Member
    • pyperformance would not be a good benchmark for adding this fast path: 0% of calls.
    • But even in the standard library test suite:
    Argument type Calls Share Note
    exact int 2,924,088 86.7% existing fast path (_PyLong_FormatWriter)
    None 181,066 5.4%
    exact list 69,501 2.1%
    other types 58,946 1.7% user classes, type objects, etc.
    exact str 58,158 1.7%
    exact dict 40,537 1.2%
    bool 20,759 0.6%
    exact float 10,657 0.32% proposed fast path
    exact bytes 5,700 0.2%
    exact tuple 2,244 0.07%
    total 3,371,663 100%

    So I am -1 unless we can provide a benchmark that justifies this fast path, and even then float would not be the first candidate, since it is lower than str and dict.

  10. vstinner commented on Oct 20, 2026

    @vstinner
    vstinner
    on Oct 20, 2026
    Member

    pyperformance would not be a good benchmark for adding this fast path: 0% of calls.

    If floats are more common in json than in general in Python, json can add a fast-path for float using the internal C API.

    The _json extension already has a code path for float which gets a PyUnicodeWriter:

        else if (PyFloat_Check(obj)) {
        PyObject *encoded = encoder_encode_float(s, obj);
        if (encoded == NULL)
            return -1;
        return _steal_accumulate(writer, encoded);
    }

  11. vstinner commented on Oct 20, 2026

    @vstinner
    vstinner
    on Oct 20, 2026
    Member

    PyUnicode_Format() calls formatfloat() for %f (and other similar formats): it calls PyOS_double_to_string() + _PyUnicodeWriter_WriteASCIIString().

    float.__float__() calls _PyFloat_FormatAdvancedWriter() which is implemented with PyOS_double_to_string() + _PyUnicodeWriter_WriteASCIIString() for the common case.

    PyUnicode_FromFormat() doesn't support %f format.

  12. vstinner
    added a commit that references this issue on Oct 20, 2026

    pythongh-159702: Add fast-path for float in _json encoder

    2edcecb

  13. vstinner commented on Oct 20, 2026

    @vstinner
    vstinner
    on Oct 20, 2026
    Member

    If floats are more common in json than in general in Python, json can add a fast-path for float using the internal C API.

    I wrote PR gh-159826 to do that. It calls PyOS_double_to_string() + PyUnicodeWriter_WriteASCII() (functions of the public C API).

  14. vstinner commented on Oct 20, 2026

    @vstinner
    vstinner
    on Oct 20, 2026
    Member

    @nineteendo: Would it be possible for you to re-run your benchmark on my PR?

    Oh crap, I wrote #156897 (JSON float optimization for json.loads()) last month, but I used git push --force and so lost my work :-(

    See also issue gh-150638.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

extension-modulesC modules in the Modules dirperformancePerformance or resource usagetype-featureA feature request or enhancement

Projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions