Founded in March 2012 in Seattle but really launched in March 2015, Qumulo has raised $130M from top investors and has of today around 200 employees. The firm was created by several formers Isilon leaders and has probably one of the best file system team in the storage industry having participate to the revolution early 2000 with Isilon, still the reference in Scale-Out NAS.
Qumulo, in addition to a few other strong technology players in file storage, participate to push object storage in the corner, where that approach should stay, I mean capacity and long-term data preservation. Some players have made some tries to offer a file system mode, in fact file storage, but it only works on the paper having some difficulties, both as a team and as the company management, and it fails with data integrity problem and high latency issues. Just think about 2 core functions of the file system: rename() and link() and you got an idea of challenges to solve not to say again the need for a strict consistency model. These points are important to confirm that building a file system on top of an object storage is still a dream or an utopia but the reverse is easier, offering a object storage API on top of a file system. It explains why object storage had real market difficulties in 2017.
Known for its innovative approach in distributed file system, Qumulo has recently repositioned its message around Qumulo File Fabric aka QF2. Funny to see that many companies use the term Fabric to finally replace and extend what we name a few years ago FAN i.e File Area Network.
And Qumulo has forgotten a key fundamental player in file system and I’m pretty surprise they didn't list Veritas. For people who know, build and play with file system, they know the role of Veritas in that space with VxFS both a technology and market presence. And if you consider the companion volume manager aka VxVM and some file system accelerator options, you get the whole picture, in other words, Veritas has invented everything in that space. Just realize that snapshot existed in 1992, resize file system in both direction – shrink or grow – dynamically again in 1992 among many capabilities. And some interesting flavors like SGI XFS could be also listed when it got introduced with IRIX 5.3 in 1993. I need to probably build the same map I created for object storage and CAS, I'm sure you remember this famous article.
Back to Qumulo, the philosophy is to provide a highly scalable file system cluster deployed in various flavors: on-prem with Qumulo appliances, on-prem software model on commodity servers such HPE Apollo with their dense servers and finally within the cloud in AWS running in EC2. It illustrates perfectly the SDS approach and advantages giving flexibility to deploy on preferred users' models and evolve with it.
Qumulo has chosen to build independent clusters glued together with a data propagation method. Imagine a global environment with a local cluster with Qumulo appliances, a second cluster deployed on Apollo servers and a third running on AWS. The company has developed an Asynchronous Automatic Continuous Replication (AACR) method to distribute data across clusters. AACR is a file-based model that copies the entire file today without deduplication and not block-based yet. With such data copies techniques, Qumulo is able to run jobs in various places on-demand, a pretty clever approach especially for some vertical use cases.
This design invites us for the next remark about the data consistency, Qumulo QF2 is strong consistent within a cluster and eventual consistent across data centers or WAN. Qumulo is also an adept of Paxos in term of consensus protocol.
In term of data protection, started with replication the company has offered for a few quarters now erasure coding and relies on the Intel ISA-L service, pretty good choice finally. With billions of files stored on the platform potentially, modern file systems had to design a new additional element to satisfy operations on metadata. Qumulo built QumuloDB distributed across all cluster nodes, a similar side metadata database model is also used for RozoFS but it is one central database per file system. Imagine a recurrent backup task that wishes to select and protect only the modified files since a certain date, let's in an incremental manner. With small volumes, walking the tree is acceptable but with huge volume and tons of files, this step is a suicide as the task will almost never finished taking too much time and you have to do this protection pretty often. Worse case, similar tasks will be added on the system as tasks take longer than the time interval between backups. Now imagine if all this create/update operations on files are stored on a side fast database you can query to find out the list of file you need to backup, it will be almost magic and super fast, you got it back path files and submit to the backup job. Same remark for archiving, tiering or migration that you need to integrate as a viable solution. This is just an illustration of this kind of service QumuloDB offers, historical and current metadata tracking and storage, with the ability to freeze a version of the database to provide a snapshot mechanism. In other words, a version of the DB is a version of the file system.
And Qumulo appeared recently for the first time in the bizarre Gartner Magic Quadrant about distributed file system and object storage, I wrote a long analysis of it recently in StorageNewsletter, you can read it here and here. Funny analysts associate object storage and distributed file system why not with secondary platform…
Clearly Qumulo is one of the few gems in file storage business with Avere Systems, now a Microsoft company, Elastifile, Panasas, Quantum with StorNext-based offering so now the last Xcellis scale-out NAS iteration, Rozo Systems and WekaIO. They all demonstrated the superiority of their native file-based approach, sometimes with a parallel mode, over things like object storage that is good finally for capacity and long-tem retention but not for high demanding files environments. Some dream about it but the market invite them to consider the business reality.
We expect Qumulo to introduce a geo-dispersed approach even with restrictions and we hope a tiering feature across cluster, cloud… and a new iteration of AACR capability explained above. The company prepares to land in Europe in Q1 2018 and I anticipate a pretty rapid growth there. Honestly the product is strong so no doubt the Seattle company will recruit top guns to rapidly gain market share and penetrate the old continent. We hope to meet them again next year during a future tour to measure progress and to confirm development directions.
0 commentaires:
Post a Comment