The filesystem on ACID
AcidFS allows interaction with the filesystem using transactions with ACID semantics. Git is used as a back end, and AcidFS integrates with the transaction package allowing use of multiple databases in a single transaction.
- Changes to the filesystem will only be persisted when a transaction is committed and if the transaction succeeds.
- Within the scope of a transaction, your application will only see a view of the filesystem consistent with that filesystem's state at the beginning of the transaction. Concurrent writes do not affect the current context.
- A full history of all changes is available, since files are stored in a backing Git repository. The standard Git toolchain can be used to recall past states, roll back particular changes, replicate the repository remotely, etc.
- Changes to a AcidFS filesystem are synced automatically with any other database making use of the transaction package and its two phase commit protocol, eg. ZODB or SQLAlchemy.
- Most common concurrent changes can be merged. There's even a decent chance concurrent modifications to the same text file can be merged.
- Transactions can be started from an arbitrary commit point, allowing, for example, a web application to apply the results of a form submission to the state of your data at the time the form was rendered, making concurrent edits to the same resource less risky and effectively giving you transactions that can span request boundaries.
The motivation for this package is the fact that it often is convenient for certain very simple problems to simply write and read data from a fileystem, but often a database of some sort winds up being used simply because of the power and safety available with a system which uses transactions and ACID semantics. For example, you wouldn't want a web application with any amount of concurrency at all to be writing directly to the filesystem, since it would be easy for two threads or processes to both attempt to write to the same file at the same time, with the result that one change is clobbered by another, or even worse, the application is left in an inconsistent, corrupted state. After thinking about various ways to attack this problem and looking at Git's datastore and plumbing commands, it was determined that Git was a very good fit, allowing a graceful solution to this problem.
In a nutshell:
- Only platforms where fcntl is available are supported. This excludes Microsoft Windows and probably the JVM as well.
- Kernel level locking is used to manage concurrency. This means AcidFS cannot handle multiple application servers writing to a shared network drive.
- The type of locking used only synchronizes other instances of AcidFS. Other processes manipulating the Git repository without using AcidFS could cause a race condition. A repository used by AcidFS should only be written to by AcidFS in order to avoid unpleasant race conditions.
All of the above limitations are a result of the locking used to synchronize commits. For the most part, during a transaction, nothing special needs to be done to manage concurrency since Git's storage model makes management of multiple, parallel trees trivially easy. At commit time, however, any new data has to be merged with the current head which may have changed since the transaction began. This last step should be synchronized such that only one instance of AcidFS is attempting this at a time. The mechanism, currently, for doing this is use of the fcntl module which takes advantage of an advisory locking mechanism available in Unix kernels.
AcidFS is easy to use. Just create an instance of acidfs.AcidFS and start using the filesystem:
import acidfs fs = acidfs.AcidFS('path/to/my/repo') fs.mkdir('foo') with fs.open('/foo/bar', 'w') as f: print >> f, 'Hello!'
If there is not already a Git repository at the path specified, one is created. An instance of AcidFS is not thread safe. The same AcidFS instance should not be shared across threads or greenlets, etc.
The transaction package is used to commit and abort transactions:
import transaction transaction.commit() # If no exception has been thrown, then changes are saved! Yeah!
The transaction package has built in support for providing metadata about a particular transaction. This metadata is used to set the commit data for the underlying git commit for a transaction. Use of these hooks is optional but recommended to provide meaningful audit information in the history of your repository. An example is the best illustration:
import transaction current = transaction.get() current.note('Added blog entry: "Bedrock Bro Culture: Yabba Dabba Dude!"') current.setUser('Fred Flintstone') current.setExtendedInfo('email', 'firstname.lastname@example.org')
A users's name may also be set by using the
current.setExtendedInfo('user', 'Fred Flintstone')
acidfs_email are available in extended info in
case you are sharing a transaction with a system that has a different notion of
what user and email should be set to. Substance D, for examples, sets the user
to an integer OID that represents the user in its system, but that might not be
what you want to see in the Git log for your repository:
current.setExtendedInfo('acidfs_user', 'Fred Flintstone') current.setExtendedInfo('acidfs_email', 'email@example.com')
The transaction might look something like this in the git log:
commit 3aa61073ea755f2c642ef7e258abe77215fe54a2 Author: Fred Flintstone <firstname.lastname@example.org> Date: Sun Sep 16 22:08:08 2012 -0400 Added blog entry: "Bedrock Bro Culture: Yabba Dabba Dude!"