Object Storage Service (Pithos)

Pithos is an online storage service based on the OpenStack Object Storage API with several important extensions. It uses a block-based mechanism to allow users to upload, download, and share files, keep different versions of a file, and attach policies to them. It follows a layered, modular implementation. Pithos was designed to be used as a storage service by the total set of the Greek research and academic community (counting tens of thousands of users) but is free and open to use by anybody, under a BSD-2 clause license.

A presentation of Pithos features and architecture is here.

Introduction

In 2008 the Greek Research and Technology Network (GRNET) decided to offer an online storage service to the Greek research and academic community. The service, called Pithos, was implemented in 2008-2009, and was made available in spring 2009. It now has more than 12,000 users.

In 2011 GRNET decided to offer a new, evolved online storage service, to be called Pithos. Pithos is designed to address the main requirements expressed by the Pithos users in the first two years of operation:

• Provide both a web-based client and native desktop clients for the most common operating systems.
• Allow not only uploading, downloading, and sharing, but also synchronization capabilities so that uses are able to select folders and have then synchronized automatically with their online accounts.
• Allow uploading of large files, regardless of browser capabilities (depending on the version, browsers may place a 2 GBytes upload limit).
• Improve upload speed; not an issue as long as the user is on a computer connected to the GRNET backbone, but it becomes important over ADSL connections.
• Allow access by non-Shibboleth (http://shibboleth.internet2.edu/). accounts. Pithos delegates user authentication to the Greek Shibboleth federation, in which all research and academic institutions belong. However, it is desirable to have the option to open up Pithos to non-Shibboleth authenticated users as well.
• Use open standards as far as possible.

In what follows we describe the main features of Pithos, the elements of its design and the capabilities it affords. We touch on related work and we provide some discussion on our experiences and thoughts on the future.

Pithos Features

Pithos is based on the OpenStack Object Storage API (Pithos used a home-grown API). We decided to adopt an open standard API in order to leverage existing clients that implement the API. In this way, a user can access Pithos with a standard OpenStack client - although users will want to use a Pithos client to use features going beyond those offered by the OpenStack API. The strategy paid off during Pithos development itself, as we were able to access and test the service with existing clients, while also developing new clients based on open source OpenStack clients.

The major extensions on the OpenStack API are:

• The use of block-based storage in lieu of an object-based one. OpenStack stores objects, which may be files, but this is not necessary - large files (longer than 5GBytes), for instance, must be stored as a series of distinct objects accompanied by a manifest. Pithos stores blocks, so objects can be of unlimited size.
• Permissions on individual files and folders. Note that folders do not exist in the OpenStack API, but are simulated by appropriate conventions, an approach we have kept in Pithos to avoid incompatibility.
• Fully-versioned objects.
• Metadata-based queries. Users are free to set metadata on their objects, and they can list objects meeting metadata criteria.
• Policies, such as whether to enable object versioning and to enforce quotas. This is particularly important for sharing object containers, since the user may want to avoid running out of space because of collaborators writing in the shared storage.
• Partial upload and download based on HTTP request headers and parameters.
• Object updates, where data may even come from other objects already stored in Pithos. This allows users to compose objects from other objects without uploading data.
• All objects are assigned UUIDs on creation, which can be used to reference them regardless of their path location.

Pithos Design

Pithos is built on a layered architecture (see Figure). The Pithos server speaks HTTP with the outside world. The HTTP operations implement an extended OpenStack Object Storage API. The back end is a library meant to be used by internal code and other front ends. For instance, the back end library, apart from being used in Pithos for implementing the OpenStack Object Storage API, is also used in our implementation of the OpenStack Image Service API. Moreover, the back end library allows specification of different namespaces for metadata, so that the same object can be viewed by different front end APIs with different sets of metadata. Hence the same object can be viewed as a file in Pithos, with one set of metadata, or as an image with a different set of metadata, in our implementation of the OpenStack Image Service.

The data component provides storage of block and the information needed to retrieve them, while the metadata component is a database of nodes and permissions. At the current implementation, data is saved to the filesystem and metadata in an SQL database. In the future, data will be saved to some distributed block storage (we are currently evaluating RADOS - http://ceph.newdream.net/category/rados), and metadata to a NoSQL database.

Block-based Storage for the Client

Since an object is saved as a set of blocks in Pithos, object operations are no longer required to refer to the whole object. We can handle parts of objects as needed when uploading, downloading, or copying and moving data.

In particular, a client, provided it has access permissions, can download data from Pithos by issuing a GET request on an object. If the request includes the hashmap parameter, then the request refers to a hashmap, that is, a set containing the object’s block hashes. The reply is of the form:

{"block_hash": "sha1",
 "hashes": ["7295c41da03d7f916440b98e32c4a2a39351546c", ...],
 "block_size":131072,
 "bytes": 242}

The client can then compare the hashmap with the hashmap computed from the local file. Any missing parts can be downloaded with GET requests with an additional Range header containing the hashes of the blocks to be retrieved. The integrity of the file can be checked against the X-Object-Hash header, returned by the server and containing the root Merkle hash of the object’s hashmap.

When uploading a file to Pithos, only the missing blocks will be submitted to the server, with the following algorithm:

• Calculate the hash value for each block of the object to be uploaded.
• Send a hashmap PUT request for the object. This is a PUT request with a hashmap request parameter appended to it. If the parameter is not present, the object’s data (or part of it) is provided with the request. If the parameter is present, the object hashmap is provided with the request.
• If the server responds with status 201 (Created), the blocks are already on the server and we do not need to do anything more.
• If the server responds with status 409 (Conflict), the server’s response body contains the hashes of the blocks that do not exist on the server. Then, for each hash value in the server’s response (or all hashes together) send a POST request to the server with the block’s data.

In effect, we are deduplicating data based on their block hashes, transparently to the users. This results to perceived instantaneous uploads when material is already present in Pithos storage.

Block-based Storage Processing

Hashmaps themselves are saved in blocks. All blocks are persisted to storage using content-based addressing. It follows that to read a file, Pithos performs the following operations:

• The client issues a request to get a file, via HTTP GET.
• The API front end asks from the back end the metadata of the object.
• The back end checks the permissions of the object and, if they allow access to it, returns the object’s metadata.
• The front end evaluates any HTTP headers (such as If-Modified-Since, If-Match, etc.).
• If the preconditions are met, the API front end requests from the back end the object’s hashmap (hashmaps are indexed by the full path).
• The back end will read and return to the API front end the object’s hashmap from the underlying storage.
• Depending on the HTTP Range header, the API front end asks from the back end the required blocks, giving their corresponding hashes.
• The back end fetches the blocks from the underlying storage, passes them to the API front end, which returns them to the client.

Saving data from the client to the server is done in several different ways.

First, a regular HTTP PUT is the reverse of the HTTP GET. The client sends the full object to the API front end. The API front end splits the object to blocks. It sends each block to the back end, which calculates its hash and saves it to storage. When the hashmap is complete, the API front end commands the back end to create a new object with the created hashmap and any associated metadata.

Secondly, the client may send to the API front end a hashmap and any associated metadata, with a special formatted HTTP PUT, using an appropriate URL parameter. In this case, if the back end can find the requested blocks, the object will be created as previously, otherwise it will report back the list of missing blocks, which will be passed back to the client. The client then may send the missing blocks by issuing an HTTP POST and then retry the HTTP PUT for the hashmap. This allows for very fast uploads, since it may happen that no real data uploading takes place, if the blocks are already in data storage.

Copying objects does not involve data copying, but is performed by associating the object’s hashmap with the new path. Moving objects, as in OpenStack, is a copy followed by a delete, again with no real data being moved.

Updates to an existing object, which are not offered by OpenStack, are implemented by issuing an HTTP POST request including the offset and the length of the data. The API front end requests from the back end the hashmap of the existing object. Depending on the offset of the update (whether it falls within block boundaries or not) the front end will ask the back end to update or create new blocks. At the end, the front end will save the updated hashmap. It is also possible to pass a parameter to HTTP POST to specify that the data will come from another object, instead of being uploaded by the client.

Pithos Back End Nodes

Pithos organizes entities in a tree hierarchy, with one tree node per path entry (see Figure). Nodes can be accounts, containers, and objects. A user may have multiple accounts, each account may have multiple containers, and each container may have multiple objects. An object may have multiple versions, and each version of an object has properties (a set of fixed metadata, like size and mtime) and arbitrary metadata.

The tree hierarchy has up to three levels, since, following the OpenStack API, everything is stored as an object in a container. The notion of folders or directories is through conventions that simulate pseudo-hierarchical folders. In particular, object names that contain the forward slash character and have an accompanying marker object with a Content-Type: application/directory as part of their metadata can be treated as directories by Pithos clients. Each node corresponds to a unique path, and we keep its parent in the account/container/object hierarchy (that is, all objects have a container as their parent).

Pithos Back End Versions

For each object version we keep the root Merkle hash of the object it refers to, the size of the object, the last modification time and the user that modified the file, and its cluster. A version belongs to one of the following three clusters (see Figure):

• normal, which are the current versions
• history, which contain the previous versions of an object
• deleted, which contain objects that have been deleted

This versioning allows Pithos to offer to its user time-based contents listing of their accounts. In effect, this also allows them to take their containers back in time. This is implemented conceptually by taking a vertical line in the Figure and presenting to the user the state on the left side of the line.

Pithos Back End Permissions

Pithos recognizes read and write permissions, which can be granted to individual users or groups of users. Groups as collections of users created at the account level by users themselves, and are flat - a group cannot contain or reference another group. Ownership of a file cannot be delegated.

Pithos also recognizes a “public” permission, which means that the object is readable by all. When an object is made public, it is assigned a URL that can be used to access the object from outside Pithos even by non-Pithos users.

Permissions can be assigned to objects, which may be actual files, or directories. When listing objects, the back end uses the permissions as filters for what to display, so that users will see only objects to which they have access. Depending on the type of the object, the filter may be exact (plain object), or a prefix (like path/* for a directory). When accessing objects, the same rules are used to decide whether to allow the user to read or modify the object or directory. If no permissions apply to a specific object, the back end searches for permissions on the closest directory sharing a common prefix with the object.

Related Work

Commercial cloud providers have been offering online storage for quite some time, but the code is not published and we do not know the details of their implementation. Rackspace has used the OpenStack Object Storage in its Cloud Files product. Swift is an open source implementation of the OpenStack Object Storage API. As we have pointed out, our implementation maintains compatibility with OpenStack, while offering additional capabilities.

Discussion

Pithos is implemented in Python as a Django application. We use SQLAlchemy as a database abstraction layer. It is currently about 17,000 lines of code, and it has taken about 50 person months of development effort. This development was done from scratch, with no reuse of the existing Pithos code. That service was written in the J2EE framework. We decided to move from J2EE to Python for two reasons: first, J2EE proved an overkill for the original Pithos service in its years of operation. Secondly, Python was strongly favored by the GRNET operations team, who are the people taking responsibility for running the service - so their voice is heard.

Apart from the service implementation, which we have been describing here, we have parallel development lines for native client tools on different operating systems (MS-Windows, Mac OS X, Android, and iOS). The desktop clients allow synchronization with local directories, a feature that existing users of Pithos have been asking for, probably influenced by services like DropBox. These clients are offered in parallel to the standard Pithos interface, which is a web application build on top of the API front end - we treat our own web application as just another client that has to go through the API front end, without granting it access to the back end directly.

We are carrying the idea of our own services being clients to Pithos a step further, with new projects we have in our pipeline, in which a digital repository service will be built on top of Pithos. It will use again the API front end, so that repository users will have all Pithos capabilities, and on top of them we will build additional functionality such as full text search, Dublin Core metadata storage and querying, streaming, and so on.

At the time of this writing (March 2012) Pithos is in alpha, available to users by invitation. We will extend our user base as we move to beta in the coming months, and to our full set of users in the second half of 2012. We are eager to see how our ideas fare as we will scaling up, and we welcome any comments and suggestions.

Acknowledgments

Pithos is financially supported by Grant 296114, “Advanced Computing Services for the Research and Academic Community”, of the Greek National Strategic Reference Framework.

Availability

The Pithos code is available under a BSD 2-clause license from: https://code.grnet.gr/projects/pithos/repository

The code can also be accessed from its source repository: https://code.grnet.gr/git/pithos/

More information and documentation is available at: http://www.synnefo.org/docs/pithos/latest/index.html