:warning: This project is archived: This was an experiment, and most of the findings make up their way into the Tvix Store protocol. Check my blog and Tvix for more updates!
A more efficient way to store and substitute Nix store paths.
Docs are a bit sparse right now, please refer to https://flokli.de/posts/2021-12-10-nix-casync-intro/ for a description on how this works.
$ go build ./cmd/nix_casync/
./nix_casync serve --cache-path=path/to/local
nix copy \
--extra-experimental-features nix-command \
--to "http://localhost:9000?compression=none" $storePath
As of now, nix-casync
can be used as a space-efficient binary cache.
You probably want to put some reverse proxy doing SSL in front of it, and add
some protection on the PUT
endpoints.
The following section describes some internal behaviour of nix-casync
, and
how it treats Narfiles and Narinfo files.
Narfiles can be uploaded with most of the compression mechanisms Nix supports.
The path it's uploaded at HTTP PUT /nar/….nar[.$suffix]
doesn't really matter.
Files will be decompressed, chunked, and put in a content-addressed store.
Subsequently uploaded .narinfo
files can refer to that file via the NarHash
attribute, and downloads can happen via HTTP GET /nar/$narhash.nar[.$suffix]
.
For downloads, only a subset of compression algorithms (fast ones) are supported, as those are assembled on the fly and should really only be considered a poor-man's Content-Encoding.
While it's possible to upload with narfile compression, as written above, this is only used to compress uploads into the binary cache.
This will have some unintuitive implications - if you upload a to
/nar/$filehash.nar.zst
, the upload won't be available on that location (but
at /nar/$narhash.nar
).
This also means HTTP HEAD
requests to "compressed locations" will 404
, and
as a result, Nix clients might end up uploading the same .nar
files multiple
times. [^1]
If the binary cache is remote, it is preferable to use compression during
upload to reduce bandwidth usage. In that case, using a fast compression
algorithm, such as zstd
is recommended.
By default, downloads are served with ZSTD Compression. This can be tweaked via
the --nar-compression
command line parameter.
Narinfo files describe information about a store path, as well as some (redundant) information about the referred .nar file.
Internally, nix-casync
splits data from .narinfo
into NarMeta
and
PathInfo
models.
When a Narfile is uploaded, the following checks are made:
nix-casync
.References
field matches with what nix-casync
's internal bookkeeping
of References in NarMeta
matches.
Right now, that field is populated on the first .narinfo
upload , but as
it's possible to determine the references just by looking at the .nar
file
itself, a reference scanner could be added to nix-casync
directly.References
in the uploaded .narinfo
refer to PathInfo
(aka.narinfo
files) that were already uploaded to nix-casync
.[^1]: Nix won't upload the same store path multiple times, as it checks
$outhash.narinfo
for existence first - so this only applies to multiple
.narinfo
files referring to the same .nar
file.