Qumulo LogoQumulo Logo

Blog

Oh, the Places Your Data Will Go (without Ever Leaving Its Container)!

Your data is already in Azure Blob Storage. That is the easy part. The hard part is all the things you want to be able to do with that data.

All your research files, your telemetry data, and the training sets sitting in a Blob container are tied to that container. Another region, another cloud, an on-premises cluster or a GPU in a different Azure region can't use them without a copy. File-based apps can't read them at all. Every repeat read is billed, whether or not the data has changed.

Today we are announcing the Qumulo Blob Accelerator, which presents your existing Azure Blob containers as file systems to local or remote applications, without moving, refactoring, or altering any of it. 

No migration, no copies

A Blob Accelerator runs on the Azure Native Qumulo (ANQ) or Cloud Native Qumulo (CNQ) cluster and attaches to an existing Blob container. Your container remains the only authoritative copy at all times. No data within the Blob container is altered or modified; nothing is duplicated, and nothing needs to be synced. Your existing applications keep writing to the container directly, as they do now.

The Blob Accelerator adds three additional features to that container: a cache, a file system, and a connection to the Qumulo Cloud Data Fabric.

Cache: stop paying for the same read twice

When it comes to capacity, scale, and durability, object storage beats the other options out every time. Where it doesn't win is the cost of reading the same data over and over.

Every request to a Blob container is billed, even if the data hasn't changed since the last time it was read. And the charge is flat: API requests are billed at a rate set by the type of object storage hosting the data, whether the container is busy or quiet. In fact, less-busy datasets are more likely to sit on colder storage, which carries higher API costs than hot object storage. For data that is read again and again, repeat-read API charges can run 100 to 500 percent of what it costs to store the data.

Qumulo NeuralCache runs on every Blob Accelerator, and serves more than 90% of read requests from NVMe or DRAM, rather than from the persistent data layer. Not only does a cache hit typically return in under a millisecond (compared to 100 to 300 milliseconds from Azure Blob Storage), but every request that never reaches the Blob Storage layer never shows up on your bill either. As for the container itself, its blob names, its tiering and lifecycle policies all stay as they are.

This caching feature pays off most in use cases where the same data is read again and again, e.g.:

  • Lakehouse and analytics. Repeated scans over the same tables.

  • AI training and inference. Training epochs that re-read the same data, and inference replicas that cold-start from the same weights.

  • Genomics, cryo-EM and research. Shared reference corpora that every team reads from.

Fabric: one container, accessible from anywhere, visible everywhere

This is what we are most excited about. Until now, bringing existing data into the Qumulo Cloud Data Fabric meant migrating it into ANQ or CNQ first. The Blob Accelerator turns the Blob container itself into a hub for the Fabric and makes it globally visible, rather than regionally. It leaves the data in place but projects it to any workload, anywhere, in real time.

Attach a Blob container to your ANQ or CNQ cluster, and within minutes Qumulo indexes it and presents it as a folder across the whole Fabric. That folder can be mounted from another Azure region, another cloud, an on-premises Qumulo cluster, an edge site, or a remote GPU, with every read in every location being strictly consistent every time. By comparison, moving 200 TB over a 10 Gb/s link takes a minimum of 44 hours, even at full line rate, before the first job can start. With an Accelerator? There's no copy to wait for. With a Blob Accelerator, there is one authoritative dataset, one namespace, and zero copies.

For example, a research organization can leave an imaging archive where it is, mounting it from everywhere it is needed. A training team can run GPUs wherever capacity is available and read the same data that every other concurrent workload sees.\

A file system, over your container

Qumulo Blob Accelerator also adds NFS and SMB connectivity to data within the Azure Blob container. File-based software that has never been able to read object data can now mount it like any other share, even while every native Blob tool keeps using the same container as before.

That matters more than it might sound. More than 80 percent of unstructured-data applications require a POSIX interface, and almost all AI tools do. Blob has always been a container, but the data inside it has not been usable-in-place by the majority of those applications. The Accelerator changes that, without migration or application refactoring.

It also changes the performance. Reading directly from Blob means 100 to 300 milliseconds of latency and a limited number of transactions per second. After the initial read, the Accelerator delivers sub-millisecond latency and nearly unlimited IOPS and throughput, scaling out and back in dynamically as needed, so file applications get the performance they expect. That is high-performance file system behavior at the price of object.

To be clear about the scope: file access is read-only to start, with full read/write (export) coming soon. Your applications keep writing to the container directly, and the container remains the single authoritative copy throughout.

Your Azure services keep working

Putting a file access layer in front of a Blob container has traditionally meant giving up compatibility with Azure services that require Blob access. A Blob Accelerator keeps your existing Azure services working against the container directly. For the inverse of this story, in which Azure services need to read file data on Qumulo without replicating it, see our [blog post on Blob Access Points].

Rethink where your data can go

The usual answer to a locked-in container is another copy: another dataset to secure, another bill to pay, and another sync window during which the two will disagree. A Blob Accelerator is the alternative. It keeps your container exactly where it is, makes it faster and cheaper to read, and lets the same data show up anywhere your work needs it.

If your data is already in Blob, the question stops being "how do we move it?" and becomes "where do we want to use it?"