conda-index

conda index, formerly part of conda-build. Create repodata.json for collections of conda packages.

The conda_index command operates on a channel directory. A channel directory contains a noarch subdirectory at a minimum and will almost always contain other subdirectories named for conda's supported platforms linux-64, win-64, osx-64, etc. A channel directory cannot have the same name as a supported platform. Place packages into the same platform subdirectory each archive was built for. Conda-index extracts metadata from these packages to generate index.html, repodata.json etc. with summaries of the packages' metadata. Then conda uses the metadata to solve dependencies before doing an install.

By default, the metadata is output to the same directory tree as the channel directory, but it can be output to a separate tree with the --output <output> parameter. The metadata cache is always placed with the packages, in .cache folders under each platform subdirectory.

After conda-index has finished, its output can be used as a channel conda install -c file:///path/to/output ... or it would typically be placed on a web server.

Run normally

python -m conda_index <path to channel directory>

Note conda index (instead of python -m conda_index) may find legacy conda-build index.

Run for debugging

python -m conda_index --verbose --threads=1 <path to channel directory>

Contributing

conda create -n conda-index "python >=3.9" conda conda-build "pip >=22"

git clone https://github.com/conda/conda-index.git
pip install -e conda-index[test]

cd conda-index
pytest

Summary of changes from the previous `conda-build index` version

Approximately 2.2x faster conda package extraction, by extracting just the metadata to streams instead of extracting packages to a temporary directory; closes the package early if all metadata has been found.
No longer read existing repodata.json. Always load from cache.
Uses a sqlite metadata cache that is orders of magnitude faster than the old many-tiny-files cache.
The first time conda index runs, it will convert the existing file-based .cache to a sqlite3 database .cache/cache.db. This takes about ten minutes per subdir for conda-forge. (If this is interrupted, delete cache.db to start over, or packages will be re-extracted into the cache.) sqlite3 must be compiled with the JSON1 extension. JSON1 is built into SQLite by default as of SQLite version 3.38.0 (2022-02-22).
Each subdir osx-64, linux-64 etc. has its own cache.db; conda-forge’s 1.2T osx-64 subdir has a single 2.4GB cache.db. Storing the cache in fewer files saves time since there is a per-file wait to open each of the many tiny .json files in old-style .cache/.
cache.db is highly compressible, like the text metadata. 2.4G → zstd → 88M
No longer cache paths.json (only used to create post_install.json and not referenced later in the indexing process). Saves 90% disk space in .cache.
Updated Python and dependency requirements.
Mercilessly cull less-used features.
Format with black

Parallelism

This version of conda-index continues indexing packages from other subdirs while the main thread is writing a repodata.json.

All current_repodata.json are generated in parallel. This may use a lot of ram if repodata.json has tens of thousands of entries.