GitTorrent – An Oxymoron?

The idea of combining peer-to-peer distribution (torrents) with source control (git) has recently been promoted (GitTorrent-2015-Announcement) but it actually has been around for awhile (GitTorrent-GoogleCode).

The appeal for this amalgamation is the combining distributing software products (packages, ISOs, etc.) via torrents while keeping all the benefits of source control systems such as Git.

The conflict between the ideas is fairly obvious. Torrents are not viewed as providing reliable service for the fundamental reason that they are not reliable. They optimize crowd distribution of popular material. Git needs to retain historic boring archives of old versions which will probably never be fetched.

There are several inherent contradictions between the “Git” and “Torrent” worlds:

  • Source Control takes responsibility for retention of every version of an object, forever.
  • Peer-to-peer networking is dependent on a social control which assumes that there is demand to access the content being shared. It is easy for a peer-to-peer network to identify and shame “leaches” who only download material but never donate bandwidth to the community. Measuring who is actually retaining seldom retrieved versions  would be far harder.
  • A primary feature of source control systems is to attribute every byte of the current version of a file/object to the specific update when it was introduced, and who to “blame” for that update. Source control systems frequently reqiure each edit to be password and/or certificate authenticated. Torrents which required tracable attribution for every contribution would see a major portion of their seeders unwilling to “contribute” material. The MPAA would probably claim that this is most of the content being torrented.

However, there are some major architectural similarities between Object Storage solutions designed for Enterprise storage (whether in-house and/or cloud-based) and peer-to-peer networking:

  • Both rely on a large number of storage targets rather than upon high availability of specific storage targets. Many low-cost modestly-available targets end up being cheaper and more reliable than a few highly available targets. But peer-to-peer targets don’t just “fail”, they are withdrawn from availability by their owners. That is a problem that Enterprise storage does not have to address.
  • Both benefit from the use of cryptographic hashes to identify chunks. Software distribution has been publishing checksums of packages and ISOs for some time, but they are typically maintained separately from the storage. This is unforntate because it requires the end user to remember to apply the fingerprint to the downloaded content, something that is probably actually done several times a day.

Actually peer-to-peer networks are more in need of cryptographic hashes than Enterprise storage clusters, but so far they tend to use cheaper hash algorithms. Every contributing sever cannot be vetted by the peer-to-peer network. Volunteer help seldom sticks around for along validation process. Therefore peer-to-peer networking must be designed to prevent malevolent and/or misbehaving servers from corrupting the distributed storage cluster.


One of the interesting statements that Chris Ball made in his blog announcement of the 2015 vintage GitTorrent was that FTP was unsuitable as a distribution method because there’s no master index of which hosts have which files in FTP, so we wouldn’t know where to look for anything.

But BitTorrent style directories are not the real solution either. They are the equivalent of HDFS or pNFS – centralized metadata referencing distributed chunks/blocks.

When you fully embrace identifying chunks with Cryptographic Hashes you just have to go all the way (which is what we did with NexentaEdge). Everything needs to be a chunk identified by a cryptographic hash of its payload. That includes the metadata. You just have to name some of the metadata chunks to anchor the references to the other chunks.

The root metadata chunks need to have an additional lookup driven by the cryptographic hash of their fully qualified name. To make a globally unique name just scope it by a domain name.

The Name Hash of a fully qualified, domain name scoped, object identifier needs to be able to find a list of the versions of the named object.

Each chunk is either considered to be “named” or “unnamed”. Named chunks are located by their Name Hash. Unnamed chunks by the Content Hash. Whichever hash is selected yields some form of partition of the global namespace. For example the configuration might specify that there are 1024 partitions. You then query for the desired hash on servers registered to support that partition.

To find the Metadata for an object:

  • Hash the fully qualified name of the object.
  • Map the hash value to a partition.
  • Ask the servers supporting that partition to return the most recent version of the Manifest for this object, or refer you to a later version that they know about but do not have a copy of.
  • Note that the returned Manifest still has a cryptographic hash of its content, allowing end-to-end validation of the manifest.
  • Pick the most recent valid version returned.

Then used the saved content hashes in the Manifest to find the payload chunks.

  • Map the content hash to its partition.
  • Ask the servers supporting that partition to return this chunk, or to refer you to another server that has it.
  • Only the first replica found actually found needs to be returned.

Note: this is slightly different from what NexentaEdge does, but NexentaEdge has the benefit of having all of the storage servers for a given cluster being at a single location. This enables use of multicast messaging to query all members of a “partition” in a single operation. There isno way to use multicast reliably over wide area networks.

In the following diagram various payload chunks are put to various partitions. Their content hashes are saved, and then referenced in version manifests which are stored (and indexed) on the partition derived from the name hash of the object name.


Object Storage Features Applied to Peer-to-Peer Networking

Despite the inherent conflicts there are several features of enterprise object storage that are applicable to the problem of peer-to-peer storage.

In both Enterprise and peer-to-peer storage, objects can be broken down into chunks Identified by a Cryptographic Hash. Each chunk represents a portion of an Object Version’s payload or metadata. Payload chunks can be  identified by a cryptographic hash of their payload. Metadata chunks can be also identified by a cryptographic hash of their payload, but also need to be found using only the object name. This can be done using the cryptographic hash of the object’s fully qualified name.

Using cryptographic hashes to identify chunks has several benefits.

  • Cryptographic Hashes Self-Validate the Chunk. As each chunk is retrieved the client can validate that it’s payload is compatible with its cryptographic hash.
  • Cryptographic Hashes enable distributed deduplication.
  • Cryptographic Hashes can drive the distribution of Chunks. The cryptographic hash of the content and/or name can determine which storage targets hold each chunk.

Breaking up Storage

Distributing chunks by cryptographic hashes allows each peer-to-peer server to only deal with a subset of the objects in the system.

A client looking for a peer-to-peer replica of a chunk only need to contact the set of server dealing with the desired slice of the “hash space”. This applies whether seeking metadata or payload.

Authenticating Users

In enterprise object storage systems, users typically authenticate using a tenant-specific authorization server, such as Active Directory or LDAP.

For peer-to-peer storage a given domain name would publish a public key. The holder of the matching private key could either retain the exclusive right to make new versions or just the right to delete posts made by others.

The system should probably also require each version be signed with a key associated with the email address.

Remaining Issues

One of the toughest challenges for anyone building a peer-to-peer storage network is determining how many replicas (or erasure encoded slices) are needed to make the chance of losing content to be effectively zero (or some other TBD threshold).

This is related to the other challenge of how to encourage users to provide storage.

The costs of long-haul communications are such that a peer-to-peer storage network is unlikely to be cost effective on a $$$ per IOP or TB basis. But there are other benefits of a peer-to-peer storage community that may be appealing:

  • There is no need to tell anyone other than your end users what your are storing. Metadata can be encrypted. If you know the key you can decrypt the metadata and confirm the name and version of the version metadata retrieved. If you do not know the key, you have a blob with a name that hashes to X.
  • Content Distribution (albeit a distributed mirroring system may be more valuable here). Those fetching from a mirror do not have to independtly validate the content, corrupted or altered data cannot be retrieved from the system.
  • There is another, related to dynamic ingest of data, which we’ll cover in my next blog.


This post touches upon features of Nexenta’s object storage product NexentaEdge. Portions of NexentaEdge are covered by patent applications and/or patents. Nexenta reserves all rights under those patents.


It is a conventional wisdom: peer-to-peer storage works for distributing of popular items. A torrent swarm is never more efficient than when delivering the latest episode of Game of Thrones.

But there is another usage of peer-to-peer networking that is almost the opposite – inbound swarming. This idea was originally suggested by Robert Novak while he was at Nexenta, but many details have been flushed out since then. An inbound swarm is quite unlike a Game of Thrones swarm: it carries a horde of distinct files that are bound to only one destination, and most will never be accessed.

Turning Peer-to-Peer Inside-Out

The use-case is simple: imagine that you have a very large number of people generating content at a location with limited upstream capacity. The “limited” capacity can be quite high, it just has to be limited relative to the demand.

If the use case is not clear, consider a technical conference at the Moscone Center (or any venue). Or consider a sporting event that wasn’t built with a WIFI router for each row of seat. In either case you have a lot of content being generated by smartphones and tablets, a lot more content than can be transmitted live from the venue.

This content has some special characteristics:

  • It is undoubtedly targeted to some form of object storage: a photo archive site, a cloud storage folder or perhaps even a home server.
  • 99.9% of it will reach its destination without any need for peer-to-peer networking, eventually. Specifically when the smartphone or tablet returns to the office or home it will upload everything via WiFi without any real problems.
  • It needs to be backed up. This content is at risk of failure, either of the smartphone/tablet or of the SD card that the content is stored on.

Peer-to-peer networking can provide a backup service to these mobile devices tailored exactly to these scenarios.

Basically, the mobile devices back each other up. But this needs to be automatic, between collaborating users without relying on negotiating trust on an individual basis.

To establish this trust the application would have to:

  • Throttle the amount of storage and bandwidth each volunteer provided. This is a feature of torrenting software already.
  • Promise the source of the material that nobody would be able to view or alter the content before it reached its final destination.
  • Promise the intermediate holders of the data that they could in no way be held responsible for any of the content because there was absolutely no way for them to review it.

Once the trust issues are solved there is more than enough bandwidth within a Conference center, and more than enough storage for all the content being generated. Content just needs to be held on multiple devices until the bandwidth is available to deliver the blob to its ultimate destination.

Each file/object is:

  • Encrypted. The payload is completely undecipherable until it reaches its destination.
  • Fingerprinted. There is a unique id for each file/object calculated by a cryptographic hash of the file/object content.

This serves several purposes:

  • It prevents unauthorized alteration of the content by any of the volunteer devices holding a backup of it.
  • It allows easy detection of when the original device or a different backup device has already delivered the file/object.  It is likely that if such a service were deployed that the typical file/object would never be accessed. Most of the time the original device will make it safely back to the office or home and deliver the file itself. Like any backup image, the preferred scenario is that it is never accessed.
  • The designated destination could even be an intermediate destination to provide greater anonymity. The first destination would decrypt the object enough to know where it would should route it to. This doesn’t have to be NSA-proof obfuscation; but most people would rather not be broadcasting their Google Photo account identification to everyone at the venue with them

Pieces of this solution have been around for a long time. Routing schemes exist for ad hoc networks that are dynamically discovered. Store and forward relay is as old as email. But this solution is actually far simpler than either of those. Each replica is stored on at most one intermediate device, and is forwarded at most once. The only ad hoc network used is at the source venue. Forwarding would be limited to known “home” Wi-Fi networks that are connected to the Internet.


Crowd-Sourced Backup

So this is essentially a mobile “crowd-sourced” backup  system that is specifically tailored to the needs of files generated by mobile devices, such as photos of presentations at a conference, or photos of friends together at a convert, etc.

The goal is to provide backup replicas that are retained for perhaps half a day. There is no need for long term storage. Rather than seeking to distribute replicas to many destinations, it instead seeks many paths to a single destination per object.

The biggest gotcha with this idea is that it really takes a Mobile OS vendor to properly enable the on-site file transfers. There isn’t any incentive for users to download a “CrowdSharing” app unless they are likely to encounter other devices with the same App installed.

Without critical mass this doesn’t work. Do you remember the Zune’s Wireless Sharing feature? It would have been a big plus, if you ever ran across other people with Zune’s to share with.

So this is a great potential application, but realistically it probably will not go anywhere unless embraced by Google and/or Apple.

Hybrid storage arrays versus all-flash

Instead of introduction: a fair summary at TheRegister.

This, and many other writings on the topic reiterate the pros and cons, restate the importance of considering your use case and your workload, mention the cost per IOPS and cost per TB. The latter, by the way, ranges between $70/TB (low-end SATA) to thousands and tens of thousands (high-end all-flash array).

To be fair, per IOPS price for those high-end arrays still beats 7.2K RPM SATA by about 20 to 100 times.

In the meantime, tiering in the last couple years  went out of vogue, looks like. In part and specifically this (coming out of vogue) relates to automated within-LUN (sub-LUN or intra-LUN) and within-target hot <=> cold tiering by storage array transparently for the users and applications. Case in point: Compellent’s Data Progression, 3PAR’s Dynamic Optimization of the previous decade.

And many others. We don’t hear about this kind of tiering, the “automated data migration: much anymore.

The reason, one might say, is that SSDs become better and cheaper, denser and more reliable at the rate set by Moore’s law back in 1965. The real reason, I say, is different: hybrid is tough. Designing, building and delivering storage software to optimize IO datapath across hybrid array – is a rocket science.

Let’s make a mental experiment though (a leap of faith, really) and assume there’s a piece of such storage software, something like this when looking from 30,000 ft:


Wouldn’t we want then to combine excellent read performance (whereby active data timely and on-demand migrates to SSDs) and total capacity, while at the same time using the same flash tier for user writes, to complete them at flash latencies?

Of course we would.

The fact that an optimal and stable hybrid storage array(*) is not done yet does not mean that it cannot be done. Yes – the corresponding software is racy and  hard to debug (**). More importantly, it is tough to test across exploding number of use cases. But – it is doable. And once it is done and demonstrated, it might as well change the perspective and the common story line..

Hybrid storage arrays or all-flash?

(*)  Term “hybrid” is used in many different contexts. On the drive level, there are hybrid SSDs, HDDs and even SSHDs.  At the opposite extreme storage administrators build hybrid solutions combining all-flash arrays (the proverbial “primary tier”) and archival Cloud services. Hybrid storage array, hybrid volume (LUN) – is somewhere in the middle..

(**) Disclaimer: some of the related work is being done at Nexenta, and will be part of the next-gen release of the storage appliance..

Global Namespace for Docker

This text has been posted to Docker development community, at  – might be a bit too technical at times.

1. Terms

Global Namespace: often refers to the capability to aggregate remote filesystems via unified (file/directory) naming while at the same time supporting unmodified clients. Not to be confused with LXC pid etc. namespaces

2. sha256

Docker Registry V2 introduces content-addressable globally unique (*) digests for both image manifests and image layers. The default checksum is sha256.

Side note: sha256 covers a space of more than 10 ** 77 unique random digests, which is about as much as the number of atoms in the observable universe. Apart from this unimaginable number sha256 has all the good crypto-qualities including collision resistance, avalanche effect for small changes, pre-image resistance and second pre-image resistance.

The same applies to sha512 and SHA-3 crypto-checksums, as well as, likely, Edon-R and Blake2 to name a few.

Those are the distinct properties that allows us to say the following: two docker images that have the same sha256 digest are bitwise identical; the same holds for layers and manifests or, for that matter, any other sha256 content-addressable “asset”.

This simple fact can be used not only to self-validate the images and index them locally via Graph’s in-memory index. This can be further used to support global container/image namespace and global deduplication. That is:

Global Namespace
Global Deduplication

  • for image layers. Hence, this Proposal.

3. Docker Cluster

Rest of this document describes only the initial implementation and the corresponding proof-of-concept patch:

The setup is a number (N >= 2) of hosts or VMs, logically grouped in a cluster and visible to each other through, for instance, NFS. Every node in the cluster runs docker daemon. Each node performs a dual role: it is NFS server to all other nodes, with NFS share sitting directly on the node’s local rootfs. Simultaneously, each node is NFS client, as per the diagram below:


Blue arrows reflect actual NFS mounts.

There are no separate NAS servers: each node, on one hand, shares its docker (layers, images) metadata and, separately, driver-specific data. And vice versa, each node mounts all clustered shares locally, under respective hostnames as shown above.

Note: hyper-convergence

Often times this type of depicted clustered symmetry, combined with the lack of physically separate storage backend is referred to as storage/compute “hyper-convergence”. But that’s another big story outside this scope..

Note: runtime mounting

As far as this initial implementation (link above) all the NFS shares are mounted statically and prior to the daemon’s startup. This can be changed to on-demand mount and more..

Back to the diagram. There are two logical layers: Graph (image and container metadata) and Driver (image and container data). This patch patches them both – the latter currently is done for aufs only.

4. Benefits

  • An orchestrator can run container on an image-less node, without waiting for the image to get pulled
  • Scale-out: by adding a new node to the cluster, we incrementally add CPU, memory and storage capacity for more docker images and containers that, in turn, can use the aggregated resource
  • Deduplication: any image or layer that exists in two or more instances can be, effectively, deduplicated. This may require pause/commit and restart of associated containers; this will require reference-counting (next)


It’s been noted in the forums and elsewhere that mixing images and containers in the Graph layer is probably not a good idea. From the clustered perspective it is easy to see that it is definitely not a good idea – makes sense to fork /var/lib/docker/graph/images and /var/lib/docker/graph/containers, or similar.

6. What’s Next

The patch works as it is, with the capability to “see” and run remote images. There are multiple next steps, some self-evident others may be less.

The most obvious one is to un-HACK aufs and introduce a new multi-rooted (suggested name: namespace) driver that would be in-turn configurable to use the underlying OS aufs or overlayfs mount/unmount.

This is easy but this, as well as the other points below, requires positive feedback and consensus.

Other immediate steps include:

  • graph.TagStore to tag all layers including remote
  • rootNFS setting via .conf for Graph
  • fix migrate.go accordingly

Once done, next steps could be:

  • on demand mounting and remounting via distributed daemon (likely etcd)
  • node add/delete runtime support – same
  • local cache invalidation upon new-image-pulled, image-deleted, etc. events (“cache” here implies Graph.idIndex, etc.)
  • image/layer reference counting, to correctly handle remote usage vs. ‘docker rmi’ for instance
  • and more

And later:

  • shadow copying of read-only layers, to trade local space for performance
  • and vice versa, removal of duplicated layers (the “dedup”)
  • container inter-node migration
  • container HA failover
  • object storage as the alternative backend for docker images and layers (which are in fact immutable versioned objects, believe it or not).

Some of these are definitely beyond just the docker daemon and would require API and orchestrator (cluster-level) awareness. But that’s, again, outside the scope of this proposal.

7. Instead of Conclusion

In the end the one thing that makes it – all of the above – doable and feasible is the immutable nature of image layers and their unique and global naming via crypto-content-hashes.

Interoperable Image Storage


The idea is probably not new, bits and pieces of it exist in multiple forms for many years and likely constitute some Platonic truth that is above and beyond anyconcrete implementation. This applies to image storage systems(*), object storage systems, file archival storage systems and even block storage. There’s often a content checksum that accompanies the (image, object, file, block) as part of the corresponding metadata – checksum of the corresponding content that is used to validate the latter AND, at least sometimes, address it – that is, find its logical and/or physical location when reading/searching.

Historically those content checksums were 16bit and 32bit wide CRCs and simlar; nowadays, however, more often than not storage systems use cryptographically-secure, collision- and pre-image-attack resistant SHA-2 and SHA-3, possibly Edon-R and Blake2 as well.

Now.. example. Let’s say, Vendor A and Vendor B both store a given popular image, say, one of those Ubuntu LTS as per For instance. Vendor A would store this content as a file over blocks, while Vendor B would use object backend to store the same image as bunch of distributed chunks.

When storing the image, in addition to (and preferably, in parallel with) checksumming blocks – A, or chunks – B, both vendors would compute the whole-image checksum: SHA-2 and/or SHA-3 for starters (NIST’s final report provides other viable alternatives). The latter then must become the stored-image metadata – alias on the image’s name, or more exactly, a specific version of the latter. Further, users may request this crypto-alias, or aliases, to be returned as part of the operation of storing the image, or later, in “exchange” for the stored version name.

Let’s see what it gives..

  1. User can now perform
    Get(hash-type, crypto-hash alias)

    – on any vendor that supports the above, to uniquely retrieve and validate a given version of the image

  2. The like/dislike and trust/distrust feedback can be metadata-attached to the image’s cryptohash and thus becomes globally cumulative and available to all users
  3. Vendors can begin partnering to deduplicate common content, independently of their respective vendor-specific and often-proprietary backend mechanisms
  4. The content can be uniquely identified and end-to-end validated with 100% certainty

The long and short of it is that secure crypto-hash on a given content can be used for vendor-neutral  access to trusted images which can further enable cross-vendor interoperability, global internet-wide deduplication, peer-to-peer networking with torrent-like service to distribute the same trusted self-validating content, and beyond.

See related: Global Namespace for Docker