memray runnow supports--aggregateto produceaggregated capture files <aggregated capture files>, which can be much smaller but aren't able to be used for generating every type of report. (#277)- Add integration with debuginfod to automatically download debug information for binaries if it is available. (#308)
- Flame graphs produced by
memray flamegraphare now around 85% smaller. (#314)
memray run --liveandmemray run --live-remotesilently dropped the--trace-python-allocatorsflag. This has been fixed, and the flag is now properly propagated from the CLI to the tracker. (#283)- Fix a bug that was causing Memray to crash when the Tracker is being destroyed and some other thread is still registering allocations or deallocations (#289)
- Work around a bug in GDB versions before 10.1 that could cause
memray attachto fail. (#310) - Work around a bug in LLDB on Linux that could cause
memray attachto hang. (#311)
- Speed up native allocation tracking by up to 45% (#294)
memray run --liveandmemray run --live-remotesilently dropped the--trace-python-allocatorsflag. This has been fixed, and the flag is now properly propagated from the CLI to the tracker. (#283)- Fix a bug that was causing Memray to crash when the Tracker is being destroyed and some other thread is still registering allocations or deallocations (#289)
- Memray is now fully supported on macOS, and the warnings that macOS support is experimental have been dropped. (#194)
- Add a checkbox to flamegraphs that allows hiding frames from the import system (#261)
memray attachcan be used toattach to a running process <attach>(#266)- Consider frames from the import system as "irrelevant" in the generated flamegraphs. (#268)
- Fix a crash that can happen when two different threads try to register frames at the same time without the GIL held. (#251)
- Add a new
transformsubcomand that allows transforming Memray capture files into output files compatible with other tools. We're starting by supporting conversions to the gprof2dot format, which allows producing graph-like reports when combined with graphviz. (#200) - Added a new
--temporary-allocationsoption to theflamegraph,table,tree, andsummaryreporters for showing thetemporary allocations </temporary_allocations>instead of the high water mark ones. (#201) - When the
greenletmodule is in use, also assign a distinct thread ID to each greenlet. Greenlets aren't threads, but they are distinct threads of execution within a single process, with distinct stacks, so assigning different thread IDs to each makes it easier to interpret reports wheregreenletwas used. (#209) - Use a monotonic counter to generate thread IDs, rather than using the pthread ID. Those pthread IDs can be reused, making it difficult to tell what thread performed an allocation. (#209)
- Print a warning when we detect that the Python interpreter was built without debug information or without symbols, letting the user know in advance that these conditions may result in incorrect stack traces or missing filenames and line numbers. (#211)
- A new
%%memray_flamegraphJupyter cell magic is provided by%load_ext memray, and can be used to memory profile code directly in a Jupyter notebook. (#237) - Add
csvas a possible target format formemray transform, producing a report of all of the allocations that made up the process's high water mark of allocated memory. This CSV file can then be loaded and analyzed using libraries likepandas. (#241)
- Up until now, if the program being profiled included a Cython module built with profiling support enabled, those Cython functions would show up in our Python call stacks. This was rarely useful in practice, as most Cython libraries aren't distributed with profiling support enabled, and supporting this had a surprisingly high maintenance cost. We've removed this integration, so you'll need to use
--nativemode to see inside of Cython modules. We are not considering this a backwards-incompatible change, since it does not affect any of our public interfaces (though it could affect the contents of reports generated by Memray). (#206)
- Fix a bug that caused incorrect
--nativemode stacks on Python 3.11 for allocations performed directly by the interpreter's eval loop. (#209) - Fix a crash when an extension module terminates the program using non-Python APIs under tracking. (#228)
- Prevent a crash that could occur when forked processes that have been under tracking without
follow_fork=Trueremove the profiling function with pending frames needed to be flushed to the results file. (#196)
- We now capture Python stacks for allocations made by threads that existed before the Memray tracker was started. (#130)
- Add support for Python 3.11 (#138)
- Add support for MacOS. (#174)
- Add experimental support for Greenlet. (#185)
- Prevent a crash that could occur if the Memray API was used to stop and later restart tracking while another thread was running Python code. (#152)
- Prevent a use-after-free bug that could result in a crash if
sys.setprofile()was called while Memray was tracking. Now ifsys.setprofile()is called, all future allocations on that thread will report unknown Python stacks, instead of potentially incorrect stacks. (#176)
- Add a progress bar indicator to the record processing phases in the different reporters so users can have an approximate idea of how much time processing the result files will take. (#111)
- The
memray statsreporter is now up to 50% faster, and its output is easier to interpret because it now processes all allocations by default. (#136) - Add a line showing the heap size over time to the memory plot in the html-based reporters (which already showed the resident size over time). (#142)
- Remove the
--include-all-allocations/-aargument to thememray statsreporter. Previously this was too slow to be used by default, but now that it has been sped up, it doesn't make sense to use anything else. The old default behavior of only processing allocations that made up the high water mark of the application's memory usage was confusing and misleading. (#136)
- Fix a crash with SIGBUS when the file system fills up while
memray runis writing a capture file. (#117) - Recognize when a capture file has been truncated (most likely because the tracked process was killed unexpectedly) and ignore any incomplete record at the end of the file. (#129)
- Fix the histogram used by the
memray statsreporter to choose sane bin sizes when all captured allocations are the same size. (#133) - Fix the aggregation by location at the bottom of the
memray statsreport when the--include-all-allocationsoption is used. (#134) - Fix a bug causing deallocations with
freeandmunmapto be included in the reported "Total allocations" count ofmemray stats --include-all-allocations. (#136) - Fix the two "largest allocating locations" sections in the
memray statsreport to actually aggregate by location. Previously they were aggregating by distinct stacks, so if two different paths hit the same line of code, it would be counted separately instead of together. (#136) - Fix a bug causing memory freed by
munmapto be incorrectly added into the reported "Total memory allocated" ofmemray stats --include-all-allocations. (#136) - Exclude
PYMALLOC_FREEfrom the allocator type distribution (other deallocators were already being ignored, but this recently added one was missed). (#136) - Fix the
memray statshistogram to be based on the actual sizes of all allocations. Previously it only saw the sizes after a rollup by stack had already been performed, so it was binning allocation sizes that had already been summed. (#136) - Fixed a bug where aggregating native call stacks could give misleading results on aarch64 under some circumstances. (#141)
- Fix a bug that made
memray run --live -cfail if the command to run contained double quotes. (#147) - Ensure our TUI isn't displaying stale data by periodically flushing the latest available data from the tracker (rather than only flushing when a buffer fills up). (#147)
- Fix the handling of the thread switch commands in the live mode TUI before the first allocation has been seen. (#147)
- Finalize and document the Memray
tracking API <api>. (#42) - Ensure that wheels built by
make distare reproducible (so that running the build twice produces identical artifacts). (#47) - Reduce the size of the
memray runcapture file by around 20% by using a more efficient encoding for which allocator was used to perform a given allocation and whether wecaptured a native stack <native tracking>for that allocation. (#52) - Support
memray run -c "..."to profile an in-line script provided on the command line. (#61) - The capture files produced by
memray runare now around 90% smaller thanks to a more efficient encoding scheme for the binary files. (#67) - Add support for Alpine Linux and musl libc. (#75)
- Capture allocations made through the C99
aligned_allocfunction. (#79) - By default the capture file will now be compressed using LZ4 after tracking completes. This temporarily requires extra disk space while the compression runs, but results in roughly 75% less disk space required in the end. Compression can be disabled with
--no-compress. (#82) - Speed up tracking by around 5% by building with link-time optimization (LTO). (#91)
- Add a new
--trace-python-allocatorsoption tomemray runthat allows tracking all allocations made using the Python allocators. This will result in bigger output files and slower profiling but it allows getting insights about all of the interpreter's memory allocations. (#92)
- Previously we attempted to read all allocation records into memory when processing a capture file in our reporters. This could fail on large files, so now we process the file in a streaming fashion instead. (#62)
- Make
memray runperform the same modifications to sys.path as the interpreter itself would when running a script. (#86) - Fixed a bug in the
stats reporter <stats>that could result in the largest allocations being omitted from the histogram. (#95) - Fixed a bug that caused Memray reporters to display incorrect stacks when
native trackingwas enabled and native allocations from different locations occurred underneath the same Python stack. (#96)
- Support the latest versions of Rich (previously we pinned to an old version due to some formatting changes in more recent versions). (#98)
- Add
memrayas a command line entry point. (#20)
- Add publishing of ManyLinux2010 wheels for 64 and 32 bit systems. (#2)
- Fix 32 bit builds. (#2)
- Initial release.