Deterministic
The same input tree always produces the same archive bytes. Timestamps and ordering are normalised, so archive hashes are comparable across machines.
v0.9.4 · MIT licensed
A boring archive format for build artifacts
Haven packs directories into a single file with content-addressed chunks, so repeated builds of the same tree cost almost nothing to store. One static binary, no runtime dependencies, deterministic output.
Why another archive format
Most CI pipelines end up tarring a build directory and pushing it somewhere. That works until the directory is 400 MB and 98% of it is identical to yesterday's build. Haven splits files into content-addressed chunks and stores each unique chunk once, so the incremental cost of an artifact is roughly the size of what actually changed.
It is not a backup tool and not a package manager. It does one job: take a directory, produce a single verifiable file, and unpack it byte-for-byte somewhere else.
Properties
Deduplicated
Rolling-hash chunking with a 64 KiB target. Shared chunks between archives in the same store are written once.
Verifiable
Every archive carries a manifest of chunk digests. haven verify
re-hashes the contents and fails loudly on mismatch.
Streamable
Archives can be written to and read from a pipe. No seeking required for the common create and extract paths.
Single binary
Statically linked, around 4 MB. No shared libraries, no interpreter, nothing to install on the target host.
Small surface
Six subcommands. The format specification is nine pages. Both are meant to stay that way.
Thirty seconds
$ haven create --store ./cache -o dist-2f9c.hv ./dist
scanned 1841 files 412.6 MiB
chunked 6702 chunks 409.1 MiB
written 213 chunks 18.4 MiB (95.5% deduplicated)
archive dist-2f9c.hv 18.7 MiB
$ haven verify dist-2f9c.hv
ok 1841 files 6702 chunks manifest digest b3:41d9e0c7…
$ haven extract dist-2f9c.hv --into /srv/release/2f9c
restored 1841 files 412.6 MiB in 2.31sThe 95.5% figure is what a second build of an unchanged frontend bundle typically looks like. Cold archives dedupe far less; see the measurements.
Status
The format is frozen for the 0.9 line. Anything that would change archive bytes is deferred to 1.0.