You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Discuss: not update Cargo.toml minor/patch version? #14962
Hi, I see we've checked in Cargo.lock recently #14135, and I think it's good!
But I'm not sure whether this was discussed: what about not updating Cargo.toml (for minor/patch versions), but only Cargo.lock?
From the discussion, I can see the main motivation is to have reproducible build (agains near latest dependencies) in CI. To achieve this, Cargo.lock (updated by bot) is enough.
Whether or not updating Cargo.toml means whether or not force downstream users to use only the latest dependency versions. Personally I prefer a more tolerable version range, so that downstream can update deps 1 by 1 and audit each dep's changes.
FYI in iceberg-rust, we have similar discussions on this topic, and we prefer to have a wider range of versions support, to allow users to choose their dep version (by not updating Cargo.toml too often) https://lists.apache.org/thread/pv3onm41229lovs1odqg94fdc60wcp73
From the discussion, I can see the main motivation is to have reproducible build (agains near latest dependencies) in CI. To achieve this, Cargo.lock (updated by bot) is enough.
Another reason we always increase Cargo.toml versions of submodules is that we don't test DataFusion building with other versions.
For example, if one of the dependencies like arrow releases a new minor version that has a new function, and then we update DataFusion to use that new function, then DataFusion actually really does require the newer version
We don't spend much time
Whether or not updating Cargo.toml means whether or not force downstream users to use only the latest dependency versions. Personally I prefer a more tolerable version range, so that downstream can update deps 1 by 1 and audit each dep's changes.
I think having a wider version range would be valuable and make DataFusion easier to use by downstream projects as well. The key thing is to figure out how to test that the ranges encoded in Cargo.toml are actually accurate
In iceberg-rust, we currently use it to do MSRV test together with min version deps test (some deps cannot be built on MSRV in newer toolchain versions )
I think having a wider version range would be valuable and make DataFusion easier to use by downstream projects as well. The key thing is to figure out how to test that the ranges encoded in Cargo.toml are actually accurate
If we allow/support a version range, we should test with lowest versions and highest versions. If both version-sets pass the tests, we can 🤞-project that any allowed versions pass the test.
Activity
alamb commentedon Mar 3, 2025
Another reason we always increase Cargo.toml versions of submodules is that we don't test DataFusion building with other versions.
For example, if one of the dependencies like
arrowreleases a new minor version that has a new function, and then we update DataFusion to use that new function, then DataFusion actually really does require the newer versionWe don't spend much time
I think having a wider version range would be valuable and make DataFusion easier to use by downstream projects as well. The key thing is to figure out how to test that the ranges encoded in Cargo.toml are actually accurate
xxchan commentedon Mar 3, 2025
There's a flag
direct-minimal-versionshttps://doc.rust-lang.org/cargo/reference/unstable.html#direct-minimal-versions which fits perfectly for minimal version testing,but maybe has some minor problems though rust-lang/cargo#5657 (comment)In iceberg-rust, we currently use it to do MSRV test together with min version deps test (some deps cannot be built on MSRV in newer toolchain versions )
alamb commentedon Mar 25, 2025
@logan-keede filed a related ticket / discussion
cargo-semver-checks#15408findepi commentedon Aug 4, 2025
If we allow/support a version range, we should test with lowest versions and highest versions. If both version-sets pass the tests, we can 🤞-project that any allowed versions pass the test.
I agree the project time budget is a constrain.