Irmin is a library to persist and synchronize distributed data structures both on-disk and in-memory. It enables a style of programming very similar to the Git workflow, where distributed nodes fork, fetch, merge and push data between each other. The general idea is that you want every active node to get a local (partial) copy of a global database and always be very explicit about how and when data is shared and migrated.
The design consists of three main components:
- a low-level immutable and consistent key/value data-store
- a DAG persisted in that datastore
- a tag store which associate names to keys of the low-level data-store.
The immutability of the low-level block store makes it significantly easier to apply replication and network coding techniques to improve resilience via replication, and to optimise scheduling across many hosts.
- Blog: Introducing Irmin: Git-like distributed, branchable storage
- Blog: Using Irmin to add fault-tolerance to the Xenstore database
Projects using Irmin
0.11.0 - May 2016
- Use Logs (#342, @talex5)
- Improve non-unix portablity of Irmin_fs (#345, @samoht)
- Change the signature of Store.iter to defer opening the file only when needed. This was causing a file-descriptor early exhaustion on Windows (#345, @samoht)
- Fix paths for references on Windows (#345, @samoht)
- Port to ocaml-git 1.8.0
- Rather large API change in Irmin.Private.Contents.Store and Irmin.Private.Commit.Store to make it easier to build new and efficient Irmin backends. (#346, @samoht)
- Fix performance problem in the computation of LCAs (#351, @talex5)
- Fix sort order for Git trees (#352, @talex5)
0.10.1 - Nov 2015
- Support for launchd: the --address argument of the CLI now supports a URI launchd://<name> where <name> corresponds to the section in the property list file (#321, by @djs55)
- Expose /watch-rec in the REST API (#326, by @samoht)
- Expose Store.Key = Contents.Path in Irmin.Maker. Otherwise, the type of steps is abstract. (#327, by @talex5)
0.10.0 - Oct 2015
- Fix the Irmin_mem backend to work when equal keys might be not structurally equal (Pervasives.(=) is evil)
- Fix Hash.SHA1.equal to always return true when the underlying bigarrays are equals. Before that, this was only the case when the whole Cstruct.t where identical: ie. same bigarray but also same offset in the Cstruct.t value, which is obviously not always the case. Apply the same fix to Hash.SHA1.compare and Hash.SHA1.hash.
- Renamed "tag" to "branch" in the API, as "tag" is confusing for Git users. BC.tag is now BC.name and BC.branch is now BC.head_ref. The various "Tag" modules are now called "Ref" ("Branch" would be confusing here since they only store references to commits, not the branch contents). Note: The remote HTTP protocol still uses "tag".
- Remove Irmin_http_server.listen. Instead, return the Cohttp configuration for the server and let the user perform the listen. The resulting API is simpler (removes timeout and uri parameters), more flexible, and easier to use from Mirage.
- Remove Irmin.task from API of internal stores (commit, node, etc). Tasks are now passed explicitly to operations that need them, so it is now explicit which operations create commits. For example, the API now makes it clear that lcas doesn't change anything, while lca requires a task because it may create commits. Apart from simplifying the code, this change also makes it possible to create the internal stores once, not once per commit message. Note: this does not affect the main BC API, so most users will see no difference.
- Remove Irmin.Basic. This was a functor that took a functor for making stores and returned a functor for making stores with strings for branch names and SHA1 for the hash. It's easier to write the application out in full than to explain to people what it does. This change also makes it possible for back-ends to provide extra operations in a type-safe way. In particular, Irmin_git.Internals has moved inside the store type and the runtime check that it is only used with the correct store type is now enforced at compile time instead.
- Removed AO.config. It was only used by the removed Git.Internals hack.
- Moved AO.create to AO_MAKER.
- Remove dummy functions that are no longer needed with the new API:
- View.task is gone (it never did anything).
- View.create is gone (it ignored both its arguments and called View.empty).
- Ir_node.Graph.Store.create (unused, but previously required by AO).
- Ir_commit.History.Store.create (same).
- Removed the unused-and-not-exported Ir_bc.Make and Ir_bc.MAKER features.
- Combine Ir_bc.STORE_EXT and Ir_s.STORE. Ir_s was the only consumer of the Ir_bc.STORE_EXT interface, and all it did was repack the values to match its own interface. Now, Ir_bc exports the final public API directly, which simplifies the code.
- Moved module types into ir_s.mli and removed ir_s.ml. Before, all module types were duplicated in the .ml and .mli files.
- BC stores now contain a Repo module. A Repo.t represents a repository as a whole, rather than any particular branch. Operations which do not look at the current branch have been moved to this module. They are: branches, remove_branch, heads, watch_branches, import, export, and task_of_head. When updating old code, you can use BC.repo t to get a Repo.t from a branch. Note that heads previously ensured that the current branch's head was included in the returned set (which made a difference for anonymous branches). This feature has been removed. In the future, the plan is to use OCaml's GC to track which anonymous branches are still being used and return all of them.
- The internal stores (commit, node, etc) used to implement a full BC store are now created by the back-ends, not by Ir_bc. This allows back-ends to use their own APIs for this. In particular, back-ends can now share resources (such as a database connection) between stores. Internal stores no longer need to deal with config values at all.
- Sync.create now takes a Repo.t, not a config, allowing Repo.config to be removed and allowing sharing of the back-end's internal state with the sync code. For example, the Git back-end no longer needs to create a new Git store object for sync.
- Change type head to type commit_id. head was confusing because it applied to all commits, not just branch heads. Putting id in the name makes it clear that this is just data and (for example) holding an ID will not prevent the corresponding commit from being GC'd (once we have GC). of_head is now of_commit_id, task_of_head is now task_of_commit_id, Internals.commit_of_head is now Internals.commit_of_id and BC.Head is now BC.Hash.
0.9.10 - Oct 2015
- Expose the Git compression level (#104, #298 by @samoht)
- Add an optional config argument to all the backend's config functions. This allow the backends to composed more easily. (initial patch by @nasrallahmounir, integration by @samoht)
- Add signatures for immutable link store, to store links between keys: Irmin.LINK and Irmin.LINK_MAKER. Add Irmin_mem.Link and Irmin_fs.Link which implement Irmin.LINK_MAKER in these backends (initial patch by @nasrallahmounir, integration by @samoht)
- Add signatures for raw values (ie. whose values are of type Cstruct.t): Irmin.RAW and raw store maker: Irmin.AO_MAKER_RAW (initial patch by @nasrallahmounir, integration by @samoht)
- Expose Irmin.Hash.digest_size (initial patch by @nasrallahmounir, integration by @samoht)
- Expose /view to the REST API (#292, by @samoht)
- Expose Irmin.Private.merge_node (#292 by @samoht)
- Change the JSON stream API, which requres ezjsonm.0.4.2. (#266, #269, #273 by @samoht)
- Fix a race when a lot of processes are trying to add a watch at the same time. (#270, #271, by @samoht)
- Expose Irmin_git.Irmin_value_store functor. This provides the Irmin Contents/Node/Commit APIs on top of a Git-type store. This is useful for backends that want to store data using the Git object format, to be able to sync with Git, but without using Git's filesystem layout and locking. (#268 by @talex5)
- Remove the first-class module API. It's confusing to duplicate the API (#293, by @talex5)
0.9.9 - Aug 2015
- Allow raw bodies in queries and responses for the REST API. This is controlled by the Content-type field set by the client: by default, we still use JSON (or use application/json) but using application/octet-stream will avoid having to hex-encode large binary blobs to make them JSON-compatible. This feature is still experimental (especially when using Git on the server) (#255)
- Adapt to ocaml-git.1.7.1 (which works with lwt.2.5.0)
- Expose Store.config for all the stores (AO, RW, etc.)
- Expose Irmin_git.Internals to be able to get back the Git commit objects from an head value (#245, #241)
- Expose Irmin.Private.remove_node
- Remove the special __root__ filename in Irmin stores and in views (#233)
- This fixes View.update_path when the view contains a value at its root. Now the updated path contains a the value stored at the root of the view.
- Writing a value to the root of a store is now an error
- Reading a value at the root of a store always return None
- Make the HTTP backend re-raise the Invalid_argument and Failure exceptions that were raised by the server.
0.9.8 - Jul 2015
- Fix wrong interaction of in-memory views and temporary branches in the store (#237)
- Fix Irmin.update_tag for HTTP clients
- Initial MirageOS support. Expose Mirage_irmin.KV_RO to surface an Irmin store as a read-only key/value store implementing `V1_LWT.KV_RO (#107)
- Expose `Irmin_git.Memory_ext. This allows the Git memory backend to be configured with a non-empty conduit context.
- Expose Irmin.SYNC
- Transmit client tasks to the HTTP server on DELETE too (#227, @dsheets)
- Do note expose private types in the public interface (#234, @koleini)
- Fix missing zero padding for date pretty-printing (#228, @dsheets)
- Update the tests to use ocaml-git.1.6.0
- Improve the style of the HTTP commit graph.
- Constraint the string tags to contain only alpha-numeric characters and few mores (-, _, '.' and /) (#186)
- Fix a race condition in Irmin.clone. (#221)
- Escpate double quotes in the output of commit messages to workaround HTML display issues. (#222)
0.9.7 - Jul 2015
- Add a version check for HTTP client and server. The client might add the version in the HTTP headers using the X-IrminVersion header - the server might decide to enfore the version check or not. The server always reply with its version in the JSON reply, using a version field. The client might use that information to bail out nicely instead of failing because of some random unmarshalling errors due to API changes (#167)
- Fix a regression in 0.9.5 and 0.9.6 when inserting new child in Git trees. This could cause a tree to have duplicate childs having the same names, which would confuse the merge functions, make git fsck and git gc complain a lot (with good reasons) and do some fency things with git index. The regression has been introduced while trying to fix #190 (the fix is in #229)
0.9.6 - Jul 2015
- Fix the datamodel: it is not possible to store data in intermediate nodes anymore (#209)
- Fix serialization of slices (#204)
- Do not fail silently when the synchronisation fails (#202)
- Fix a race in the HTTP backend between adding a watch and updating the store. In some cases, the watch callback wasn't able to see the first few updates (#198)
- Fix a race for all the on-disk backends between adding a watch and updating the store. This is fixed by making Irmin.Private.Watch.listen_dir and Irmin.Private.Watch.set_listen_dir_hook synchronous.
- Update the tests to use alcotest >= 0.4. This removes the dependency towards OUnit and nocrypto for the tests.
- Make the file-locking code a bit more robust
0.9.5 - Jun 2015
- Fix Irmin.export for the HTTP backend (#196, patch from Alex Zatelepin)
- Fix a race in Irmin.export (#196, patch from Alex Zatelepin)
- Add Task.empty (the empty task) and Task.none (the empty task constructor)
- Completely rewrite the notification mechanism. All the watch functions now take a callback as argument and return a de-allocation function. The callbacks receive a heads values (the last and current ones) and diff values. (#187)
- Add Irmin.watch_head to watch for the changes of the current branch's head
- Add Irmin.watch_tags to watch for the changes of all the tags in the store
- Add Irmin.watch_key to watch for the changes of the values associated to a given key (this is not recursive anymore).
- Add View.watch_path to watch for the changes in a subtree. The function return views and the user can use View.diff to compute differences between views if needed.
- Transfer the HTTP client task to the server to make the commit messages relative to the client state (and not the server's) (#136)
- Fix View.remove to clean-up empty directories (#190)
- Fix the ordering of tree entries in the Git backend (#190)
- Allow to create a new head from a view and a list of parents with View.make_head (#188)
- Allow to create an empty temporary branch with Irmin.empty (#161)
- Use a pure OCaml implementation of SHA1, do not depend on nocrypto anymore (#183, by @talex5)
- Remove Irmin.Snapshot. Nobody was using it and it can be easily replaced by Irmin.head, Irmin.watch_head and Irmin.update_head.
- Change signature of Irmin.iter to include the values and move it into the Irmin.RO signature.
- Add Irmin.fast_forward_head (#172)
- Add Irmin.compare_and_set_head (#171)
- Simplify the RW_MAKER signature (#158)
- Fix Irmin_git.RW_MAKER (#159)
- Improve the efficiency of the LCA computation (#174, with @talex5 help)
- By default, explore the full graph when computing the LCAs. The previous behavior was to limit the depth of the exploration to be 256 by default.
0.9.4 - Mar 2015
- Ensure that Irmin.update and Irmin.merge are atomic.
- Fix Irmin.clone of an empty branch
- Add Irmin.RW.compare_and_test that the backends now have to implement to guarantee atomicity of Irmin's high-level operations.
- Add Irmin.Private.Lock to provide per-handler, per-key locking. This can be used by backend to implement simple locking policies.
- Add Lwt.t to the return type of Irmin.tag and Irmin.tag_exn
- Do not throw [Not_found]. Now all the _exn function raise Invalid_argument (#144)
- Remove Irmin.switch and Irmin.detach
- Add Irmin.history to get the branch history as a DAG of heads (#140).
- Fix performance of lcas computation (#160)
- Add Irmin.Merge.promise combinators
0.9.3 - Jan 2015
- Fix the invalidation of the view caches (report by @gregtatcam). This was causing some confusing issues where views' sub-keys where not properly updated to to their new values when the view is merged back to the store. The issues is a regression introduced in 0.9.0.
- Add post-commit hooks for the HTTP server.
- Add Irmin.watch_tags to monitor tag creation and desctructions.
- Fix Irmin.push
- Add Irmin.with_hrw_view to easily use transactions.
- Add a phantom type to Irmin.t to denote the store capabilities read-only, read-write or branch-consistent.
- The ~old argument of a merge function can now be optional to signify that there is no common ancestor.
- Expose Irmin.with_rw_view to create a temporary, in-memory and mutable view of the store. This can be used to perform atomic operations in the store (ie. non-persistent transactions).
- Simplify the view API again
- Expose the task of previous commits. This let the user access the Git timestamp and other info such as the committer name (#90)
- The user-defined merge functions now takes an unit -> 'a result Lwt.t argument for ~old (instead of 'a). Evalutating the function will compute the least-common ancestors. Merge functions which ignore the old argument don't have to pay the cost of computing the lcas anymore.
- Expose S.lca to get the least common ancestors
- Update to ocaml-git 1.4.6
0.9.2 0 Jan 2015
- Fix S.of_head for the HTTP client (regression introduced in 0.9.0)
- Fix regression in displaying the store's graph over HTTP introduced by 0.9.0.
- Fix regression in watch handling introduced in 0.9.0.
- Fix regressions in Views introduced in 0.9.0. (thx @buzzheavyyear for the report)
- Always add a commit when calling a update function (Irmin.update Irmin.remove, Irmin.remove_rec) even if the contents' store have not changed.
- The [head] argument of [Git_unix.config] now has a proper type.
- Expose synchronisation functions for basic Irmin stores.
- The user-provided merge function now takes optional values. The function is now called much more often during recursive merges (even if one of the 3 buckets of the 3-way merge function is not filled -- in that case, it uses None).
- Also expose the type of the keys in the type basic Irmin stores. Use ('key, 'value) Irmint.t instead of 'value Irmin.t.
- The user-defined merge functions now take the current filename being merged as an additional argument.
- The user-defined Contents should expose a Path sub-module. Keys of the resulting Irmin store will be of type Path.t.
- Fix irmin init --help. (#103)
0.9.1 - Dec 2014
- Port to Cohttp 0.14.0+ HTTP interface (#102)
0.9.0 - Dec 2014
- Improve the efficiency of the Git backend
- Expose a cleaner API for the Unix backends
- Expose a cleaner public API
- Rename Origin into Task and use it pervasively through the API
- Expose a high-level REST API over HTTP (#80)
- Fix the Git backend to stop constantly overwrite .git/HEAD (#76)
- Add a limit on concurrently open files (#93, #75)
- Add remove_rec to remove directories (#74, #85)
- Remove dependency to core_kernel (#22, #81)
- Remove dependency to cryptokit andsha1and usenocrypto` instead
- Remove dependency to caml4
- Fix writing contents at the root of the store (#73)
- More efficient synchronization protocol between Irmin stores (#11)
0.8.3 - Jun 2014
- Views now keep track of their parent commits - this makes View.merge_path looks like a merge between branches. All the view operations are squashed in a unique commit.
- Better graphs, where we only show the commit history (the full graph is still available using --full on the command-lineor or ?full=1 on the web interface)
- By default, do not call dot when dumping a graph on the command-line. dot does not like big graphs, but that's still useful to have the .dot file to analyze it.
0.8.2 - Jun 2014
- Support backend specific protocols for push/pull
- The Irmin Git backend can now sync with remote Git repositories
- Simplify the organisation of the libraries: irmin, irmin.backend, irmin.server and irmin.unix (check how the example are compiled)
- Small refactoring to ease the use of the API. Now use open Irmin_unix at the top of your file and use less functor in your code (again, check the examples)
0.8.1 - Jun 2014
- Fix the behavior of IrminMemory.Make to return an hanlder to a shared datastore instead of creating a fresh one. Add IrminMemory.Fresh to return a fresh in-memory datastore maker.
- The HTTP server now outputs some nice graph (using dagre-d3). Don't expect to display very large graphs
- More friendly tag names in the Git backend (no need to prefix everything by refs/heads/ anymore)
- Partial support for recursive stores (WIP)
0.8.0 - May 2014
- Spring clean-ups in the API. Separation in IrminBranch for fork/join operations, IrminSnapshot for snapshot/revert operations and IrminDump for import/export operations. The later two implementation can be derived automaticaly from a base IrminBranch implementation. The update and merge operations are supported on each backend
- No need to have bin_io converter for contents anymore
- No need to have JSON converter for contents anymore
- No more IrminDispatch
- Add an optional branch argument to Irmin.create to use an already existing branch
- Fix order of arguments in Irmin.merge
0.7.0 - May 2015
- Feature: support for in-memory transactions. They are built on top of views.
- Feature: add support for views: these are temporary stores with lazy reads + in-memory writes; they can be used to convert back and forth an OCaml value into a store, or to have a fast stagging area without the need to commit every operation to the store.
- Support custom messages in commit messages
- Improve the IrminMerge API
- Backend: add a 'dispatch' backend for combining multiple backends into one. This can be used to have a P2P store where there is well-defined mapping between keys and host (as a DHT).
- Fix: limit the number of simulteanous open files in the Git and the file-system backend
- Speed-up the in-memory store
- Speed-up the import/export codepath
- Speed-up the reads
- Speed-up IrminValue.Mux
- Deps: use ocaml-sha instead of cryptokit
0.6.0 - Apr 2014
- Support for user-defined contents (with custom merge operators)
- Support for merge operations
- Rename IrminTree to IrminNode to reflect the fact that we can support arbitrary immutable graphs (it's better if they are DAGs but that's not mandatory)
- Rename IrminBlob to IrminContents to reflect the fact that we also support structured contents (as JSON objects)
- Support for linking the library without linking to camlp4 as well (#23)
0.5.1 - Mar 2014
- Port to use Cohttp 0.10.0 interface.
0.5.0 - Feb 2014
- More consistent support for notifications. irmin watch works now for all backends.
- Support for different blob formats on the command-line
- Support for JSON blobs
- More flexible irmin fetch command: we can now choose the backend to import the data in
- Fix import of Git objects when the blobs were not imported first
- Support non-UTF8 strings as path name and blob contents (for all backends, including the JSON one)
- Speed-up the slow tests execution time
- Improve the output graph when objects of different kinds might have the same SHA1
0.4 - Jan 2014
- The command-line tool now looks in the environment for the variable IRMIN to configure its default backend
- Add a Git backend
- Add Travis CI scripts to the repo
- Use Lwt_bytes and Lwt_unix instead of the custom-made IrminChannel
- Use bin_prot instead of a custom binary protocol
- Major refactoring: Value is now Blob, Revision is now Commit and Tag becomes Reference (rational: consistency with Git names)
- Use core_kernel instead of building a custom Identiable.S
- Use dolog instead of a custom log library
- Use mstruct (mutable buffers on top of cstruct) which is now released independently
0.3 - Dec 2013
- Fix a fd leak in the filesystem bakend
- Functorize the CRUD interface over the HTTP client implementation
- Use oasis to build the project
- Use the now released separately ezjsonm and alcotest libraries
0.2 - Nov 2013
- Fix the HTTP server responses
- More high-level tests
- Add unit-tests for the client CRUD interfaces (over memory and/or filesystem)
- Fix issues with the Tree API
- Implement a relatively efficent Import/Export scheme (#3)
- For more safety, the marshalled values are now typed in the binary protocol
- Add functions to dump the contents of the store as a Graphviz graph
- Polish the CLI which now looks usable enough
- Optimize the CRUD backend by executing high-level API functions on the server
- Improve and make the CLI easier to use
- Implement clone/pull/push/snapshot/revert in the CLI
0.1 - Oct 2013
- Use an HTTP server as a front-end
- Initial support for in-memory and filesystem backends
- Simple signature for backends
- Binary protocol for storing values and metadata and for future network exchange