Allow disabling SWDB writing #662
Conversation
This is still in use by some C clients, notably rpm-ostree.
It confused me when reading this code that the `Swdb` constructors weren't all together.
This is actually a false positive since we only set it and use it if `setGPGHomeEnv` is set. But of course, GCC doesn't know that. But still, let's silence it so that we don't learn to ignore compiler warnings.
We already declared an `swdb` variable higher up defined to `priv->swdb`. This function had a mix of `swdb` and `priv->swdb`. Only use the former for consistency.
Putting the else on a newline here is weird. Let's just be consistent.
Add a new `dnf_transaction_new_with_flags()` which allows for setting flags right from the start. This is most a convenience function but will be necessary in the next patch, which requires being able to set flags that affect initialization itself and so can't be a separate call.
Not all users of libdnf really need the functionality added by the SWDB database. This is especially true for rpm-ostree-based variants of Fedora and RHEL, which only use libdnf for composing on the server and package layering on the client. Some major differences that render the SWDB not useful for rpm-ostree: - rpm-ostree already keeps track of requested packages in a separate database (really, text file). - rpm-ostree always applies package installs from scratch on top of the base OSTree, so the concept of history is simply different (see e.g. coreos/rpm-ostree#1489). - rpm-ostree updates are performed offline, i.e. not on the live system. As such, we don't need to track the state of the RPM transaction as it happens in a separate database. If librpm throws an error, we just throw out the whole offline rootfs. - OSTree commits themselves already contain the list of packages that were installed in them, so that we can consult the commit metadata (or even the rpmdb directly) if we need to render the RPM diff on a system or OSTree branch.
|
This will allow rpm-ostree to stop using a fork of libdnf and go back to using master (see e.g. coreos/rpm-ostree#1381). We can then work towards stopping to bundle rpm-ostree at least on f29+ and el8. |
|
thanks @jlebon! |
This actually doesn't make sense. That would be a regression from pre-SWDB in which the bootstrap transaction that creates the environment, as well as the further interactions with that environment get recorded in the transaction database. This is incredibly valuable for when you're working from a debug environment or similar, so I don't see how this makes any sense to do. One of the consequences of this, for example, would be that Anaconda netinstall transactions wouldn't be recorded anymore. That's a huge problem. |
|
Gotcha, makes sense. I initially was going to implement this by simply passing a flag from rpm-ostree, but then thought it might make more sense anyway to key off of the install root. Anyway, will switch back to using an explicit flag. |
(Don't want to distract too much on this PR, nice work! I just want to say though that I am very hesitant about going back to dynamically linking libdnf. There were two silent ABI breaks in the past that were an enormous pain to debug. If we had gating tests in bodhi that run rpm-ostree's test suite though that would ameliorate my concern there though) |
|
OK, switched to just exposing the new functionality for rpm-ostree! ⬆️
Yeah, understood. I do hope we can eventually go back to unbundling though. Definitely agreed we should make that conditional on improving the CI story first. |
Do we have that ability today? If gating isn't an option, is non-gating? |
|
@jlebon There's no getter function for the new swdb flag. Can you please add one? |
Add new getter & setter functions to allow the context's transaction to be initialized with the new `DISABLE_SWDB` flag. This will be used by rpm-ostree.
|
Done! ⬆️ |
|
@jlebon I wrote an alternative patch: #664 Could you check if it makes sense to you and if it would work with rpm-ostree code? |
|
☔ The latest upstream changes (presumably 2b32f55) made this pull request unmergeable. Please resolve the merge conflicts. |
|
Prep patches merged. |
Only the last three commits are new here. The rest are from #660. I'll just copy the commit messages from the last two:
[transaction] Add DNF_TRANSACTION_FLAG_DISABLE_SWDB
Not all users of libdnf really need the functionality added by the SWDB
database. This is especially true for rpm-ostree-based variants of
Fedora and RHEL, which only use libdnf for composing on the server and
package layering on the client.
Some major differences that render the SWDB not useful for rpm-ostree:
database (really, text file).
base OSTree, so the concept of history is simply different (see e.g.
Consider an
rpm-ostree historycommand coreos/rpm-ostree#1489).As such, we don't need to track the state of the RPM transaction as it
happens in a separate database. If librpm throws an error, we just
throw out the whole offline rootfs.
were installed in them, so that we can consult the commit metadata (or
even the rpmdb directly) if we need to render the RPM diff on a system
or OSTree branch.
[context] Disable SWDB in alternate install roots
This gets to the heart of the distinction. If one is using libdnf to
operate on the live system, then an SWDB tracking those mutations makes
sense. If not, then it's likely libdnf is only being used as a low-level
tool to generate some other artifacts. If we detect such cases,
automatically, disable writing an SWDB.
Closes: #645
Requires: #660