Part 2: Checkouts, Branching, & Merging

This section deals with navigating repository history, creating & merging branches, and understanding conflicts.

The Hangar Workflow

The hangar workflow is intended to mimic common git workflows in which small incremental changes are made and committed on dedicated topic branches. After the topic has been adequatly set, topic branch is merged into a separate branch (commonly referred to as master, though it need not to be the actual branch named "master"), where well vetted and more permanent changes are kept.

Create Branch -> Checkout Branch -> Make Changes -> Commit

Making the Initial Commit

Let’s initialize a new repository and see how branching works in Hangar:

from hangar import Repository
import numpy as np

repo = Repository(path='/Users/rick/projects/tensorwerk/hangar/dev/mnist/')

repo_pth = repo.init(user_name='Test User', user_email='test@foo.com', remove_old=True)

Hangar Repo initialized at: /Users/rick/projects/tensorwerk/hangar/dev/mnist/.hangar

When a repository is first initialized, it has no history, no commits.

repo.log() # -> returns None

Though the repository is essentially empty at this point in time, there is one thing which is present: a branch with the name: "master".

repo.list_branches()

['master']

This "master" is the branch we make our first commit on; until we do, the repository is in a semi-unstable state; with no history or contents, most of the functionality of a repository (to store, retrieve, and work with versions of data across time) just isn’t possible. A significant portion of otherwise standard operations will generally flat out refuse to execute (ie. read-only checkouts, log, push, etc.) until the first commit is made.

One of the only options available at this point is to create a write-enabled checkout on the "master" branch and to begin to add data so we can make a commit. Let’s do that now:

co = repo.checkout(write=True)

As expected, there are no columns nor metadata samples recorded in the checkout.

print(f'number of metadata keys: {len(co.metadata)}')
print(f'number of columns: {len(co.columns)}')

number of metadata keys: 0
number of columns: 0

Let’s add a dummy array just to put something in the repository history to commit. We’ll then close the checkout so we can explore some useful tools which depend on having at least one historical record (commit) in the repo.

dummy = np.arange(10, dtype=np.uint16)
col = co.add_ndarray_column('dummy_column', prototype=dummy)
col['0'] = dummy
initialCommitHash = co.commit('first commit with a single sample added to a dummy column')
co.close()

If we check the history now, we can see our first commit hash, and that it is labeled with the branch name "master"

repo.log()

* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e (master) : first commit with a single sample added to a dummy column

So now our repository contains: - A commit: a fully independent description of the entire repository state as it existed at some point in time. A commit is identified by a commit_hash. - A branch: a label pointing to a particular commit / commit_hash.

Once committed, it is not possible to remove, modify, or otherwise tamper with the contents of a commit in any way. It is a permanent record, which Hangar has no method to change once written to disk.

In addition, as a commit_hash is not only calculated from the commit ’s contents, but from the commit_hash of its parents (more on this to follow), knowing a single top-level commit_hash allows us to verify the integrity of the entire repository history. This fundamental behavior holds even in cases of disk-corruption or malicious use.

Working with Checkouts & Branches

As mentioned in the first tutorial, we work with the data in a repository through a checkout. There are two types of checkouts (each of which have different uses and abilities):

`Checking out a branch / commit for reading: <api.rst#read-only-checkout>`__ is the process of retrieving records describing repository state at some point in time, and setting up access to the referenced data.

• Any number of read checkout processes can operate on a repository (on any number of commits) at the same time.

`Checking out a branch for writing: <api.rst#write-enabled-checkout>`__ is the process of setting up a (mutable) staging area to temporarily gather record references / data before all changes have been made and staging area contents are committed in a new permanent record of history (a commit).

• Only one write-enabled checkout can ever be operating in a repository at a time.

• When initially creating the checkout, the staging area is not actually “empty”. Instead, it has the full contents of the last commit referenced by a branch’s HEAD. These records can be removed / mutated / added to in any way to form the next commit. The new commit retains a permanent reference identifying the previous HEAD commit was used as its base staging area.

• On commit, the branch which was checked out has its HEAD pointer value updated to the new commit’s commit_hash. A write-enabled checkout starting from the same branch will now use that commit’s record content as the base for its staging area.

Creating a branch

A branch is an individual series of changes / commits which diverge from the main history of the repository at some point in time. All changes made along a branch are completely isolated from those on other branches. After some point in time, changes made in a disparate branches can be unified through an automatic merge process (described in detail later in this tutorial). In general, the Hangar branching model is semantically identical to the Git one; The one exception is that in Hangar, a branch must always have a name and a base_commit. (No “Detached HEAD state” is possible for a write-enabled checkout). If No base_commit is specified, the current writer branch HEAD commit is used as the base_commit hash for the branch automatically.

Hangar branches have the same lightweight and performant properties which make working with Git branches so appealing - they are cheap and easy to use, create, and discard (if necessary).

To create a branch, use the create_branch() method.

branch_1 = repo.create_branch(name='testbranch')

branch_1

BranchHead(name='testbranch', digest='a=eaee002ed9c6e949c3657bd50e3949d6a459d50e')

We use the list_branches() and log() methods to see that a new branch named testbranch has been created and is indeed pointing to our initial commit.

print(f'branch names: {repo.list_branches()} \n')
repo.log()

branch names: ['master', 'testbranch']

* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e (master) (testbranch) : first commit with a single sample added to a dummy column

If instead, we actually specify the base commit (with a different branch name) we see we do actually get a third branch. pointing to the same commit as master and testbranch

branch_2 = repo.create_branch(name='new', base_commit=initialCommitHash)

branch_2

BranchHead(name='new', digest='a=eaee002ed9c6e949c3657bd50e3949d6a459d50e')

repo.log()

* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e (master) (new) (testbranch) : first commit with a single sample added to a dummy column

Making changes on a branch

Let’s make some changes on the new branch to see how things work.

We can see that the data we added previously is still here (dummy arrayset containing one sample labeled 0).

co = repo.checkout(write=True, branch='new')

co.columns

Hangar Columns
    Writeable         : True
    Number of Columns : 1
    Column Names / Partial Remote References:
      - dummy_column / False

co.columns['dummy_column']

Hangar FlatSampleWriter
    Column Name              : dummy_column
    Writeable                : True
    Column Type              : ndarray
    Column Layout            : flat
    Schema Type              : fixed_shape
    DType                    : uint16
    Shape                    : (10,)
    Number of Samples        : 1
    Partial Remote Data Refs : False

co.columns['dummy_column']['0']

array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=uint16)

Let’s add another sample to the dummy_arrayset called 1

arr = np.arange(10, dtype=np.uint16)
# let's increment values so that `0` and `1` aren't set to the same thing
arr += 1

co['dummy_column', '1'] = arr

We can see that in this checkout, there are indeed two samples in the dummy_arrayset:

len(co.columns['dummy_column'])

2

That’s all, let’s commit this and be done with this branch.

co.commit('commit on `new` branch adding a sample to dummy_arrayset')
co.close()

How do changes appear when made on a branch?

If we look at the log, we see that the branch we were on (new) is a commit ahead of master and testbranch

repo.log()

* a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94 (new) : commit on `new` branch adding a sample to dummy_arrayset
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e (master) (testbranch) : first commit with a single sample added to a dummy column

The meaning is exactly what one would intuit. We made some changes, they were reflected on the new branch, but the master and testbranch branches were not impacted at all, nor were any of the commits!

Merging (Part 1) Fast-Forward Merges

Say we like the changes we made on the new branch so much that we want them to be included into our master branch! How do we make this happen for this scenario??

Well, the history between the HEAD of the new and the HEAD of the master branch is perfectly linear. In fact, when we began making changes on new, our staging area was identical to what the master HEAD commit references are right now!

If you’ll remember that a branch is just a pointer which assigns some name to a commit_hash, it becomes apparent that a merge in this case really doesn’t involve any work at all. With a linear history between master and new, any commits exsting along the path between the HEAD of new and master are the only changes which are introduced, and we can be sure that this is the only view of the data records which can exist!

What this means in practice is that for this type of merge, we can just update the HEAD of master to point to the HEAD of "new", and the merge is complete.

This situation is referred to as a Fast Forward (FF) Merge. A FF merge is safe to perform any time a linear history lies between the HEAD of some topic and base branch, regardless of how many commits or changes which were introduced.

For other situations, a more complicated Three Way Merge is required. This merge method will be explained a bit more later in this tutorial.

co = repo.checkout(write=True, branch='master')

Performing the Merge

In practice, you’ll never need to know the details of the merge theory explained above (or even remember it exists). Hangar automatically figures out which merge algorithms should be used and then performed whatever calculations are needed to compute the results.

As a user, merging in Hangar is a one-liner! just use the merge() method from a write-enabled checkout (shown below), or the analogous methods method from the Repository Object repo.merge() (if not already working with a write-enabled checkout object).

co.merge(message='message for commit (not used for FF merge)', dev_branch='new')

Selected Fast-Forward Merge Strategy

'a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94'

Let’s check the log!

repo.log()

* a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94 (master) (new) : commit on `new` branch adding a sample to dummy_arrayset
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e (testbranch) : first commit with a single sample added to a dummy column

co.branch_name

'master'

co.commit_hash

'a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94'

co.columns['dummy_column']

Hangar FlatSampleWriter
    Column Name              : dummy_column
    Writeable                : True
    Column Type              : ndarray
    Column Layout            : flat
    Schema Type              : fixed_shape
    DType                    : uint16
    Shape                    : (10,)
    Number of Samples        : 2
    Partial Remote Data Refs : False

As you can see, everything is as it should be!

co.close()

Making changes to introduce diverged histories

Let’s now go back to our testbranch branch and make some changes there so we can see what happens when changes don’t follow a linear history.

co = repo.checkout(write=True, branch='testbranch')

co.columns

Hangar Columns
    Writeable         : True
    Number of Columns : 1
    Column Names / Partial Remote References:
      - dummy_column / False

co.columns['dummy_column']

Hangar FlatSampleWriter
    Column Name              : dummy_column
    Writeable                : True
    Column Type              : ndarray
    Column Layout            : flat
    Schema Type              : fixed_shape
    DType                    : uint16
    Shape                    : (10,)
    Number of Samples        : 1
    Partial Remote Data Refs : False

We will start by mutating sample 0 in dummy_arrayset to a different value

old_arr = co['dummy_column', '0']
new_arr = old_arr + 50
new_arr

array([50, 51, 52, 53, 54, 55, 56, 57, 58, 59], dtype=uint16)

co['dummy_column', '0'] = new_arr

Let’s make a commit here, then add some metadata and make a new commit (all on the testbranch branch).

co.commit('mutated sample `0` of `dummy_column` to new value')

'a=fcd82f86e39b19c3e5351dda063884b5d2fda67b'

repo.log()

* a=fcd82f86e39b19c3e5351dda063884b5d2fda67b (testbranch) : mutated sample `0` of `dummy_column` to new value
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e : first commit with a single sample added to a dummy column

co.metadata['hello'] = 'world'

co.commit('added hellow world metadata')

'a=69a08ca41ca1f5577fb0ffcf59d4d1585f614c4d'

co.close()

Looking at our history how, we see that none of the original branches reference our first commit anymore.

repo.log()

* a=69a08ca41ca1f5577fb0ffcf59d4d1585f614c4d (testbranch) : added hellow world metadata
* a=fcd82f86e39b19c3e5351dda063884b5d2fda67b : mutated sample `0` of `dummy_column` to new value
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e : first commit with a single sample added to a dummy column

We can check the history of the master branch by specifying it as an argument to the log() method.

repo.log('master')

* a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94 (master) (new) : commit on `new` branch adding a sample to dummy_arrayset
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e : first commit with a single sample added to a dummy column

Merging (Part 2) Three Way Merge

If we now want to merge the changes on testbranch into master, we can’t just follow a simple linear history; the branches have diverged.

For this case, Hangar implements a Three Way Merge algorithm which does the following: - Find the most recent common ancestor commit present in both the testbranch and master branches - Compute what changed between the common ancestor and each branch’s HEAD commit - Check if any of the changes conflict with each other (more on this in a later tutorial) - If no conflicts are present, compute the results of the merge between the two sets of changes - Create a new commit containing the merge results reference both branch HEADs as parents of the new commit, and update the base branch HEAD to that new commit’s commit_hash

co = repo.checkout(write=True, branch='master')

Once again, as a user, the details are completely irrelevant, and the operation occurs from the same one-liner call we used before for the FF Merge.

co.merge(message='merge of testbranch into master', dev_branch='testbranch')

Selected 3-Way Merge Strategy

'a=002041fe8d8846b06f33842964904b627de55214'

If we now look at the log, we see that this has a much different look than before. The three way merge results in a history which references changes made in both diverged branches, and unifies them in a single commit

repo.log()

*   a=002041fe8d8846b06f33842964904b627de55214 (master) : merge of testbranch into master
|\
| * a=69a08ca41ca1f5577fb0ffcf59d4d1585f614c4d (testbranch) : added hellow world metadata
| * a=fcd82f86e39b19c3e5351dda063884b5d2fda67b : mutated sample `0` of `dummy_column` to new value
* | a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94 (new) : commit on `new` branch adding a sample to dummy_arrayset
|/
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e : first commit with a single sample added to a dummy column

Manually inspecting the merge result to verify it matches our expectations

dummy_arrayset should contain two arrays, key 1 was set in the previous commit originally made in new and merged into master. Key 0 was mutated in testbranch and unchanged in master, so the update from testbranch is kept.

There should be one metadata sample with they key hello and the value "world".

co.columns

Hangar Columns
    Writeable         : True
    Number of Columns : 1
    Column Names / Partial Remote References:
      - dummy_column / False

co.columns['dummy_column']

Hangar FlatSampleWriter
    Column Name              : dummy_column
    Writeable                : True
    Column Type              : ndarray
    Column Layout            : flat
    Schema Type              : fixed_shape
    DType                    : uint16
    Shape                    : (10,)
    Number of Samples        : 2
    Partial Remote Data Refs : False

co['dummy_column', ['0', '1']]

[array([50, 51, 52, 53, 54, 55, 56, 57, 58, 59], dtype=uint16),
 array([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10], dtype=uint16)]

co.metadata

Hangar Metadata
    Writeable: True
    Number of Keys: 1

co.metadata['hello']

'world'

The Merge was a success!

co.close()

Conflicts

Now that we’ve seen merging in action, the next step is to talk about conflicts.

How Are Conflicts Detected?

Any merge conflicts can be identified and addressed ahead of running a merge command by using the built in diff tools. When diffing commits, Hangar will provide a list of conflicts which it identifies. In general these fall into 4 categories:

1. Additions in both branches which created new keys (samples / columns / metadata) with non-compatible values. For samples & metadata, the hash of the data is compared, for columns, the schema specification is checked for compatibility in a method custom to the internal workings of Hangar.

2. Removal in Master Commit/Branch & Mutation in Dev Commit / Branch. Applies for samples, columns, and metadata identically.

3. Mutation in Dev Commit/Branch & Removal in Master Commit / Branch. Applies for samples, columns, and metadata identically.

4. Mutations on keys of both branches to non-compatible values. For samples & metadata, the hash of the data is compared; for columns, the schema specification is checked for compatibility in a method custom to the internal workings of Hangar.

Let’s make a merge conflict

To force a conflict, we are going to checkout the new branch and set the metadata key hello to the value foo conflict... BOO!. Then if we try to merge this into the testbranch branch (which set hello to a value of world) we see how hangar will identify the conflict and halt without making any changes.

Automated conflict resolution will be introduced in a future version of Hangar, for now it is up to the user to manually resolve conflicts by making any necessary changes in each branch before reattempting a merge operation.

co = repo.checkout(write=True, branch='new')

co.metadata['hello'] = 'foo conflict... BOO!'

co.commit ('commit on new branch to hello metadata key so we can demonstrate a conflict')

'a=95896880b33fc06a3c2359a03408f07c87bcc8c0'

repo.log()

* a=95896880b33fc06a3c2359a03408f07c87bcc8c0 (new) : commit on new branch to hello metadata key so we can demonstrate a conflict
* a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94 : commit on `new` branch adding a sample to dummy_arrayset
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e : first commit with a single sample added to a dummy column

When we attempt the merge, an exception is thrown telling us there is a conflict!

co.merge(message='this merge should not happen', dev_branch='testbranch')

Selected 3-Way Merge Strategy

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-57-1a98dce1852b> in <module>
----> 1 co.merge(message='this merge should not happen', dev_branch='testbranch')

~/projects/tensorwerk/hangar/hangar-py/src/hangar/checkout.py in merge(self, message, dev_branch)
   1027             dev_branch=dev_branch,
   1028             repo_path=self._repo_path,
-> 1029             writer_uuid=self._writer_lock)
   1030
   1031         for asetHandle in self._columns.values():

~/projects/tensorwerk/hangar/hangar-py/src/hangar/merger.py in select_merge_algorithm(message, branchenv, stageenv, refenv, stagehashenv, master_branch, dev_branch, repo_path, writer_uuid)
    136
    137     except ValueError as e:
--> 138         raise e from None
    139
    140     finally:

~/projects/tensorwerk/hangar/hangar-py/src/hangar/merger.py in select_merge_algorithm(message, branchenv, stageenv, refenv, stagehashenv, master_branch, dev_branch, repo_path, writer_uuid)
    133                 refenv=refenv,
    134                 stagehashenv=stagehashenv,
--> 135                 repo_path=repo_path)
    136
    137     except ValueError as e:

~/projects/tensorwerk/hangar/hangar-py/src/hangar/merger.py in _three_way_merge(message, master_branch, masterHEAD, dev_branch, devHEAD, ancestorHEAD, branchenv, stageenv, refenv, stagehashenv, repo_path)
    260         if conflict.conflict is True:
    261             msg = f'HANGAR VALUE ERROR:: Merge ABORTED with conflict: {conflict}'
--> 262             raise ValueError(msg) from None
    263
    264         with mEnv.begin(write=True) as txn:

ValueError: HANGAR VALUE ERROR:: Merge ABORTED with conflict: Conflicts(t1=[(b'l:hello', b'2=d8fa6800caf496e637d965faac1a033e4636c2e6')], t21=[], t22=[], t3=[], conflict=True)

Checking for Conflicts

Alternatively, use the diff methods on a checkout to test for conflicts before attempting a merge.

It is possible to diff between a checkout object and:

1. Another branch (diff.branch())

2. A specified commit (diff.commit())

3. Changes made in the staging area before a commit is made (diff.staged()) (for write-enabled checkouts only.)

Or via the CLI status tool between the staging area and any branch/commit (only a human readable summary is produced).

merge_results, conflicts_found = co.diff.branch('testbranch')

conflicts_found

Conflicts(t1=Changes(schema={}, samples=(), metadata=(MetadataRecordKey(key='hello'),)), t21=Changes(schema={}, samples=(), metadata=()), t22=Changes(schema={}, samples=(), metadata=()), t3=Changes(schema={}, samples=(), metadata=()), conflict=True)

conflicts_found.t1.metadata

(MetadataRecordKey(key='hello'),)

The type codes for a Conflicts namedtuple such as the one we saw:

Conflicts(t1=('hello',), t21=(), t22=(), t3=(), conflict=True)

are as follow:

• t1: Addition of key in master AND dev with different values.

• t21: Removed key in master, mutated value in dev.

• t22: Removed key in dev, mutated value in master.

• t3: Mutated key in both master AND dev to different values.

• conflict: Bool indicating if any type of conflict is present.

To resolve, remove the conflict

del co.metadata['hello']
# resolved conflict by removing hello key
co.commit('commit which removes conflicting metadata key')

'a=e69ba8aeffc130c57d2ae0a8131c8ea59083cb62'

co.merge(message='this merge succeeds as it no longer has a conflict', dev_branch='testbranch')

Selected 3-Way Merge Strategy

'a=ef7ddf4a4a216315d929bd905e78866e3ad6e4fd'

We can verify that history looks as we would expect via the log!

repo.log()

*   a=ef7ddf4a4a216315d929bd905e78866e3ad6e4fd (new) : this merge succeeds as it no longer has a conflict
|\
* | a=e69ba8aeffc130c57d2ae0a8131c8ea59083cb62 : commit which removes conflicting metadata key
* | a=95896880b33fc06a3c2359a03408f07c87bcc8c0 : commit on new branch to hello metadata key so we can demonstrate a conflict
| * a=69a08ca41ca1f5577fb0ffcf59d4d1585f614c4d (testbranch) : added hellow world metadata
| * a=fcd82f86e39b19c3e5351dda063884b5d2fda67b : mutated sample `0` of `dummy_column` to new value
* | a=c1cf1bd6863ed0b95239d2c9e1a6c6cc65569e94 : commit on `new` branch adding a sample to dummy_arrayset
|/
* a=eaee002ed9c6e949c3657bd50e3949d6a459d50e : first commit with a single sample added to a dummy column