Skip to content

Latest commit

 

History

History
375 lines (320 loc) · 13.9 KB

CHANGELOG.md

File metadata and controls

375 lines (320 loc) · 13.9 KB

Changelog

Development

  • Added support for PEP 489 multi-phase initialisation and per-module state for our C extension, allowing us to support sub-interpreters with per-interpreter GIL.
  • Advertise support for free-threading python mode.
  • Enhanced generators so they yield all possible results to users before errors are raised (#123).
  • Added ijson.ALL_BACKENDS constant listing all supported backends (which might or not be available at runtime).
  • Added a capabilities constant to each backend describing which capabilities it supports.
  • Exposing backend's name under <backend>.backend_name, and default backend's name under ijson.backend_name. This is similar to the already existing name constant, only slightly better named to hopefully avoid confusion.
  • Restructured source code so all code lives under src/, and the ijson.backends._yajl2 extension under src/ijson/backends/ext/_yajl2. This allows C backend tests to actually run on cibuildwheel.
  • Fixed several potential stability issues in C backend around correct error handling.
  • Fixed potential issue with yajl and yajl2 backends where crashes could occur at interpreter shutdown.
  • Removed tox.
  • Removed support for Python 2.7 and 3.4, 3.5+ is still supported.
  • Distribute the existing benchmark.py script as ijson.benchmark. The module is an improved version of the script, supporting #iterations for a given function invocation, multiple input files, and more.
  • Fixed several issues in the yajl2_c backend and its async generators that were only made apparent when running it with PyPy and/or a CPython debug build (#101). As part of that, an issue was found and fixed in PyPy itself affecting all versions up to 7.3.12, so users will need to wait until the next version is released to be able to use async generators (https://foss.heptapod.net/pypy/pypy/-/issues/3956).
  • Adapted yajl2_c async generators to work against PyPy shortcomings (https://foss.heptapod.net/pypy/pypy/-/issues/3965).
  • Fixed compilation and async support of the yajl2_c backend in pyhthon 3.12 (#98).
  • Check IJSON_BUILD_YAJL2C environment variable when building ijson to force/skip building the yajl2_c backend (#102).
  • Added support for Python 3.12.
  • Fixed a memory leak in the yajl2_c backend triggered only when the underlying yajl functions reported a failure (#97).
  • Fixed minor README rendering issues that prevented upload of 3.2.0 distributions to PyPI.
  • New ijson.dump command-line utility for simple inspection of the ijson iteration process. This tool should be useful for new users who are usually confused with how to use the library, and the prefix in particular.
  • Fixed bug in yajl2_c backend introduced in 3.1.2 where random crashes could occur due to an unsafe reference decrement when constructing the parse/items/kvitems generators (#66).
  • Mark Python 3.10 and 3.11 as explicitly supported.
  • Fixed bug in yajl2_c backend introduced in 3.1.0 where ijson.items didn't work correctly against member names containing . (#41).
  • Python backend raises errors on incomplete JSON content that previously wasn't recognised as such, aligning itself with the rest of the backends (#42).
  • Python backed correctly raises errors when JSON numbers with leading zeros are found in the stream (#40).
  • Likewise, JSON numbers with fractions where the decimal point is not surrounded by at least one digit on both sides also produce an error now on the python backend.
  • Fixed detection of file objects with generator-based read coroutines (i.e., a read generator decorated with @types.coroutine) for the purpose of automatically routing user calls done through the main entry points. For example, when using aiofiles objects users could invoke async for item in ijson.parse_async(f) but not async for item in ijson.parse(f), while the latter has been possible since 3.1 for native coroutines.
  • Moved binary wheel generation from GitHub Actions to Travis. This gained us binary ARM wheels, which are becoming increasingly popular (#35)
  • Fixed minor memory leaks in the initialization methods of the generators of the yajl2_c backend. All generators (i.e., basic_parse, parse, kvitems and items) in both their sync and async versions, were affected.
  • Fixed two problems in the yajl2_c backend related to asyncio support, which prevented some objects like those from aiofiles from working properly (#32).
  • Ironing out and documenting some corner cases related to the use of use_float=True and its side-effect on integer number parsing.
  • Removed test package from binary distributions.
  • A new use_float option has been added to all backends to control whether float values should be returned for non-integer numbers instead of Decimal objects. Using this option trades loss of precision (which most applications probably don't care) for performance (which most application do care about). Historically ijson has returned Decimal objects, and therefore the option defaults to False for backwards compatibility, but in later releases this default could change to True.
  • Improved the performance of the items and kvitems methods of the yajl2_c backend (by internally avoiding unnecessary string concatenations). Local tests show a performance improvement of up to ~15%, but mileage might vary depending on your use case and system.
  • The "raw" functions basic_parse, parse, items and kvitems can now be used with different types of inputs. In particular they accept not only file-like objects, but also asynchronous file-like objects, behaving like their *_async counterparts. They also accept bytes and str objects directly (and unicode objects in python 2.7). Finally, they also accept iterables, in which case they behave like the ijson.common.* functions, allowing users to tap into the event pipeline.
  • ijson.common routines parse, items and kvitems are marked as deprecated. Users should use the ijson.* routines instead, which now accept event iterables.
  • New ijson.get_backend function for users to import a backend programmatically (without having to manually use importlib).
  • New IJSON_BACKEND environment variable can be used to choose the default backend to be exposed by ijson.
  • Unicode decoding errors are now reported more clearly to users. In the past there was a mix of empty messages and error types. Now the error type is always the same and there should always be an error messages indicating the offending byte sequence.
  • ijson.common.number is marked as deprecated, and will be removed on some later release.
  • Fixed errors triggered by JSON documents where the top-level value is an object containing an empty-named member (e.g., {"": 1}). Although such documents are valid JSON, they broke basic assumptions made by the kvitems and items functions (and all their variants) in all backends, producing different types of unexpected failures, including segmentation faults, raising unexpected exceptions, and producing wrong results.
  • Fixed segmentation fault in yajl2_c backend's parse caused by the previous fix introduced in 3.0.2 (#29).
  • Fixed memory leak in yajl2_c backend's parse functionality (#28).
  • Adding back the parse, kvitems and items functions under the ijson.common module (#27). These functions take an events iterable instead of a file and are backend-independent (which is not great for performance). They were accidentally removed in the redesign of ijson 3.0, which is why they are coming back. In the future they will slowly transition into being backend-specific rather than independent.
  • Exposing backend's name under <backend>.backend, and default backend's name under ijson.backend.
  • Exposing ijson.sendable_list to users in case it comes in handy.
  • Implemented all asynchronous iterables (i.e., *_async functions) in C for the yajl2_c backend for increased performance.
  • Adding Windows builds via AppVeyor, generating binary wheels for Python 3.5+.
  • Fixed known problem with 3.0rc1, namely checking that asynchronous files are opened in the correct mode (i.e., binary).
  • Improved the protocol for user-facing coroutines, where instead of having to send a final, empty bytes string to finish the parsing process users can simply call .close() on the coroutine.
  • Greatly increased testing of user-facing coroutines, which in turn uncovered problems that were fixed.
  • Adding ability to benchmark coroutines with benchmark.py.
  • Including C code in coverage measurements, and increased overall code coverage up to 99%.
  • Full re-design of ijson: instead of working with generators on a "pull" model, it now uses coroutines on a "push" model. The current set of generators (basic_parse, parse, kvitems and items) are implemented on top of these coroutines, and are fully backward compatible. Some text comparing the old a new designs can be found here.
  • Initial support for asyncio in python 3.5+ in the for of async for-enabled asynchronous iterators. These are named *_async, and take a file-like object whose read() method can be awaited on.
  • Exposure of underlying infrastructure implementing the push model. These are named *_coro, and take a coroutine-like object (i.e., implementing a send method) instead of file-like objects. In this scheme, users are in charge of sending chunks of data into the coroutines using coro.send(chunk).
  • C backend performance improved by avoiding memory copies when possible when reading data off a file (i.e., using readinto when possible) and by avoiding tuple packing/unpacking in certain situations.
  • C extension broken down into separate source files for easier understanding and maintenance.
  • Fixed a deprecation warning in the C backend present in python 3.8 when parsing Decimal values.
  • New kvitems method in all backends. Like items, it takes a prefix, and iterates over the key/value pairs of matching objects (instead of iterating over objects themselves, like in items). This is useful for iterating over big objects that would otherwise consume too much memory.
  • When using python 2, all backends now return map_key values as unicode objects, not str (until now only the Python backend did so). This is what the json built-in module does, and allows for correctly handling non-ascii key names. Comparison between unicode and str objects is possible, so most client code should be unaffected.
  • Improving error handling in yajl2 backend (ctypes-based) so exceptions caught in callbacks interrupt the parsing process.
  • Including more files in source distributions (#14).
  • Adjusting python backend to avoid reading off the input stream too eagerly (#15).
  • Fixing backwards compatibility, allowing string readers in all backends (#12, #13).
  • Default backend changed (#5). Instead of using the python backend, now the fastest available backend is selected by default.
  • Added support for new map_type option (#7).
  • Fixed bug in multiple_values support in C backend (#8).
  • Added support for multiple_values flag in python backend (#9).
  • Forwarding **kwargs from ijson.items to ijson.parse and ijson.basic_parse (#10).
  • Fixing support for yajl versions < 1.0.12.
  • Improving common.number implementation.
  • Documenting how events and the prefix work (#4).
  • New ijson.backends.yajl2_c backend written in C and based on the yajl2 library. It performs ~10x faster than cffi backend.
  • Adding more builds to Travis matrix.
  • Preventing memory leaks in ijson.items
  • Parse numbers consistent with stdlib json
  • Correct JSON string parsing in python backend
  • Publishing package version in init.py
  • Various small fixes in cffi backend