Notes on Caching

Numba supports caching of compiled functions into the filesystem for future use of the same functions.

The Implementation

Caching is done by saving the compiled object code, the ELF object of the executable code. By using the object code, cached functions have minimal overhead because no compilation is needed. The cached data is saved under the cache directory (see NUMBA_CACHE_DIR). The index of the cache is stored in a .nbi file, with one index per function, and it lists all the overloaded signatures compiled for the function. The object code is stored in files with an .nbc extension, one file per overload. The data in both files is serialized with pickle.


On Python <=3.7, Numba extends pickle using the pure-Python pickler. To use the faster C Pickler, install pickle5 from pip. pickle5 backports Python 3.8 pickler features.

Requirements for Cacheability

Developers should note the requirements of a function to permit it to be cached to ensure that the features they are working on are compatible with caching.

Requirements for cacheable function:

  • The LLVM module must be self-contained, meaning that it cannot rely on other compiled units without linking to them.
  • The only allowed external symbols are from the NRT or other common symbols from system libraries (i.e. libc and libm).

Debugging note:

  • Look for the usage of inttoptr in the LLVM IR or target_context.add_dynamic_add() in the lowering code in Python. They indicate potential usage of runtime address. Not all uses are problematic and some are necessary. Only the conversion of constant integers into pointers will affect caching.
  • Misuse of dynamic address or dynamic symbols will likely result in a segfault.
  • Linking order matters because unused symbols are dropped after linking. Linking should start from the leaf nodes of the dependency graph.

Features Compatible with Caching

The following features are explicitly verified to work with caching.

  • ufuncs and gufuncs for the cpu and parallel target
  • parallel accelerator features (i.e. parallel=True)

Caching Limitations

This is a list of known limitation of the cache:

  • Cache invalidation fails to recognize changes in symbols defined in a different file.
  • Global variables are treated as constants. The cache will remember the value in the global variable used at compilation. On cache load, the cached function will not rebind to the new value of the global variable.

Cache Sharing

It is safe to share and reuse the contents in the cache directory on a different machine. The cache remembers the CPU model and the available CPU features during compilation. If the CPU model and the CPU features do not match exactly, the cache contents will not be considered. (Also see NUMBA_CPU_NAME)

If the cache directory is shared on a network filesystem, concurrent read/write of the cache is safe only if file replacement operation is atomic for the filesystem. Numba always writes to a unique temporary file first, it then replaces the target cache file path with the temporary file. Numba is tolerant against lost cache files and lost cache entries.

Cache Clearing

The cache is invalidated when the corresponding source file is modified. However, it is necessary sometimes to clear the cache directory manually. For instance, changes in the compiler will not be recognized because the source files are not modified.

To clear the cache, the cache directory can be simply removed.

Removing the cache directory when a Numba application is running may cause an OSError exception to be raised at the compilation site.