README.md

Pijul for Git users

Introduction

Pijul is an in-development distributed version control system that implements repositories using a different model than Git's. This model enables cool features and avoids common problems which we are going to explore in this tutorial. (The good stuff appears from the "Maybe we don't need branches" section.) It is assumed that the reader uses Git in a daily base, i.e. knows not only how to commit, push and pull but also has had the need to cherry-pick commits and rebase branches at least once.

Creating a repo

$ mkdir pijul-tutorial
$ cd pijul-tutorial
$ pijul init

Nothing new here.

Adding files to the repository

Just like in Git after we create a file it must be explicitly added to the repository:

$ pijul add <files>

There's a difference though: in Pijul a file is just a UNIX file, i.e. directories are also files so we don't need to create .keep files to add empty directories to our repos. Try this:

$ mkdir a-dir
$ touch a-dir/a-file
$ pijul add a-dir/
$ pijul status

The output will be:

On branch master

Changes not yet recorded:
  (use "pijul record ..." to record a new patch)

        new file:  a-dir

Untracked files:
  (use "pijul add <file>..." to track them)

        a-dir/a-file

To add files recursively we must use the --recursive flag.

Signing keys

Pijul can sign patches automatically, so let's create a signing key before we record our first patch:

$ pijul key gen --signing

The key pair will be located in ~/.local/share/pijul/config. At the moment the private key is created without a password so treat it with care.

Recording patches

From the user perspective this is the equivalent to Git's commit operation but it is interactive by default:

$ pijul record

added file a-dir

Shall I record this change? (1/2) [ynkadi?] y

added file a-dir/a-file

Shall I record this change? (2/2) [ynkadi?] y

What is your name <and email address>? Someone's name
What is the name of this patch? Add a dir and a file
Recorded patch 6fHCAzzT5UYCsSJi7cpNEqvZypMw1maoLgscWgi7m5JFsDjKcDNk7A84Cj93ZrKcmqHyPxXZebmvFarDA5tuX1jL

Here y means yes, n means no, k means undo and remake last decision, a means include this and all remaining patches, d means include neither this patch nor the remaining patches and i means ignore this file locally (i.e. it is added to .pijul/local/ignore).

Let's change a-file:

$ echo Hello > a-dir/a-file

$ pijul record
In file "a-dir/a-file"

+ Hello

Shall I record this change? (1/1) [ynkad?] y

What is the name of this patch? Add a greeting
Recorded patch 9NrFXxyNATX5qgdq4tywLU1ZqTLMbjMCjrzS3obcV2kSdGKEHzC8j4i8VPBpCq8Qjs7WmCYt8eCTN6s1VSqjrBB4

Ignoring files

We saw that when recording a patch we can chose to locally ignore a file, but we can also create a .pijulignore or .ignore file in the root of our repository and record it. All those files accept the same patterns as a .gitignore file.

Just like in Git if we want to ignore a file that was recorded in a previous patch we must remove that file from the repository.

Removing files from the repository

$ pijul remove <files>

The files will be shown as untracked again whether they were recorded with a previous patch or not, so this has the effect of git reset <files> or git rm --cached depending on the previous state of these files.

Removing a patch

$ pijul unrecord

This command is interactive. Alternatively, one can use pijul unrecord <patch> to remove one or more patches, knowing their hash. Patch hashes can be obtained with pijul log.

Unrecording and recording the same patch again will leave the repository in the same state.

There are cases where a patch depends on a previous one. For example if a patch edits (and only edits) file A it will depend on the patch that created that file. We can see these dependencies with pijul dependencies and they are managed automatically. This is why pijul unrecord <patch> might sometimes refuse to work.

Discarding changes

$ pijul revert

This is like git checkout applied to files (instead of branches).

Branches

To create a new branch we use the pijul fork <branch-name> command and to switch to another branch we use pijul checkout <branch-name>.

To apply a patch from another branch we use the pijul apply <patch-hash> command. Notice that this doesn't produce a different patch with a different hash like git cherry-pick does.

Finally to delete a branch we have the delete-branch subcommand, but:

Maybe we don't need branches

Because in Git each commit is related to a parent (except for the first one), branches are useful to avoid mixing up unrelated work. We don't want our history to look like this:

* More work for feature 3
|
* More work for feature 1
|
* Work for feature 3
|
* Work for feature 2
|
* Work for feature 1

And if we need to push a fix for a bug ASAP we don't want to also push commits that still are a work in progress so we create branches for every new feature and work in them in isolation.

But in Pijul patches usually commute: in the same way that 3 + 4 + 8 produces exactly the same result than 4 + 3 + 8, if we apply patch B to our repo before we apply patch A and then C the result will be exactly the same that our coworkers will get if they apply patch A before patch C and then patch B. Now if patch C has a dependency called D (as we saw in Removing a patch) they cannot commute, but the entire graph commutes with other patches, i.e if I apply patch A before patch B and then patches CD I would get the same repository state than if I applied patch B before patches CD and then patch A. So Alice could have the same history as in the previous example while Bob could have

* More work for feature 1
|
* Work for feature 2
|
* More work for feature 3
|
* Work for feature 3
|
* Work for feature 1

And the repos would be equivalents; that is, the files would be the same. Why is that useful?

We can start working on a new feature without realizing that it is actually a new feature and that we need a new branch. We can create all the patches we need for that feature (e.g. the patches that implement it, the patches that fix the bugs introduced by it, and the patches that fix typos) in whatever order we want. Then we can unrecord these patches and record them again as just one patch without a rebase. (There's actually no rebase operation in Pijul.)

But this model really shines when we start to work with:

Remotes

At the moment, pushing works over SSH, the server only needs to have Pijul installed. The Nest is a free service that hosts public repositories. We can reuse our current SSH key pair or create a new pair with

$ pijul key gen --ssh

This new key pair will be stored in the same directory used for the signing keys and we can add it to The Nest like we do with SSH keys in GitHub.

Now that we have an account on The Nest we can upload our signing key with pijul key upload.

Now let's push something:

$ pijul push <our-nest-user-name>@nest.pijul.com:<our-repo>

Unless we pass the --all flag Pijul will ask us which patches we want to push. So we can keep a patch locally, unrecord it, record it again, decide that actually we don't need it and kill it forever or push it a year later when we finally decide that the world needs it. All without branches.

If we don't want to specify the remote every time we push we can set it as default with the --set-default flag.

Of course to pull changes we have the pijul pull command.

Both commands have --from-branch (source branch), --to-branch (destination branch) and --set-remote (create a local name for the remote) options.

BTW if we can keep patches for ourselves can we pull only the patches we want? Yes, that's called "partial clone". It was introduced in version 0.11 and works like this:

$ pijul pull --path <patch-hash> <remote>

Of course it will bring the patch and all its dependencies.

As we have seen we neither need branches, cherry-picking nor rebasing because of the patch theory behind Pijul.

Contributing with a remote

With Pijul we don't need forking either. The steps to contribute to a repo are:

  1. Clone it with pijul clone <repo-url>
  2. Make some patches!
  3. Go to the page of the repo in The Nest and open a new discussion
  4. The Nest will create a branch with the number of the discussion as a name
  5. Push the patches with pijul push <our-user-name>@nest.pijul.com:<repo-owner-user-name>/<repo-name> --to-branch :<discussion-number>

Then the repo owner could apply our patches to the master branch. You can also attach patches from your repos to a discussion when you create or participate in one.

Tags

A tag in Pijul is a patch that specifies that all the previous patches depend on each other to recreate the current state of the repo.

To create a tag we have the pijul tag command which will ask for a tag name.

After new patches are added to the repo we can recreate the state of any tag by creating a new branch:

pijul fork --patch <hash-of-the-tag> <name-of-the-new-branch>

Because tags are just patches we can look for their hashes with pijul log.

In summary

Forget about bad merges, feature branches, rebasing and conflicts produced by merges after cherry-picking.

Learning more

Pijul has an on-line manual but currently it is a little bit outdated. The best way to learn more is by executing pijul help. This will list all the subcommands and we can read more about any of them by running pijul help <subcommand>.

The subcommands are interactive by default but we can pass data to them directly from the command line to avoid being asked questions. All these options are explained in each subcommand's help.

For more information on the theory behind Pijul refer to Joe Neeman's blog post on the matter. He also wrote a post that explains how Pijul implements it.

A work in progress

As we said Pijul is an in-development tool: the UI could change in the future and there are some missing features. (Something like bisect would be super helpful.) But that's also an opportunity: the developers seem quite open to receive feedback.