Fix fastparquet Install and Import Errors

Three errors, three different causes

"fastparquet won't install" covers three unrelated failures. Reading the error carefully saves the afternoon.

1. pip tries to compile it

error: command 'gcc' failed: No such file or directory
error: Microsoft Visual C++ 14.0 or greater is required
ERROR: Failed building wheel for fastparquet

fastparquet contains compiled extension modules, so pip needs a wheel built for your exact Python version, OS and architecture. When none matches, pip downloads the source tarball and tries to compile — which fails on any machine without a C toolchain. The usual trigger is a Python release newer than the package: upgrade to 3.14 the week it ships and half your compiled dependencies have no wheels yet.

Confirm that is what is happening instead of guessing:

python -c "import sys, platform; print(sys.version, platform.machine())"
python -m pip install --only-binary=:all: fastparquet

With --only-binary=:all:, pip refuses to build from source. A clean "could not find a version that satisfies the requirement" means the problem is wheel availability, not your compiler. The fixes, in order of effort: upgrade pip first (python -m pip install -U pip, since old pip skips newer wheel tags), then install into a Python version the package actually supports, or use conda-forge, which publishes prebuilt binaries earlier for many platforms:

conda install -c conda-forge fastparquet

2. numpy ABI mismatch at import time

ValueError: numpy.dtype size changed, may indicate binary incompatibility.
Expected 96 from C header, got 88 from PyObject

Nothing is wrong with your Parquet file — this is a compiled package linked against a different numpy ABI than the numpy in the environment. It shows up when an old pinned fastparquet==0.x meets numpy 2.x, or when a wheel cached from a previous environment gets reused. Force a fresh resolution:

python -m pip install --force-reinstall --no-cache-dir fastparquet numpy

If the project pins fastparquet for a reason, pin numpy to the matching major version instead of fighting the rebuild.

3. Import fails inside pandas only

ImportError: Missing optional dependency 'fastparquet'.

This one is almost always an environment mismatch — the install landed in a different interpreter than the one running your code. Check sys.executable in the process that raised the error; the pandas engine error guide walks through the notebook-kernel version of the same trap.

Codec dependencies

fastparquet delegates compression to cramjam, a separate compiled package covering Snappy, Zstd, LZ4, Brotli and Gzip. On an unusual platform the failing wheel may be cramjam rather than fastparquet itself, so read which package pip actually named in the error before searching for the wrong one.

The faster path

If you are fixing this only to read one file, consider not fixing it. Installing pyarrow is normally a single working wheel away, and pyarrow is the better engine anyway:

python -m pip install pyarrow

And if the goal is just to see the schema, a sample or a row count, skip Python altogether: the Parquet Viewer reads the footer in your browser, and the SQL Workbench runs DuckDB queries against the same file — no environment, no wheels, and nothing uploaded.