Velocity • Durability

VDURA — Resiliency
AI SOLUTIONS / RESILIENCY

Hardware fails. Your training run doesn't.

Built on the assumption that drives and nodes fail daily at scale, client-side dual-parity erasure coding protects every file individually across VPODs, so a failure is a ticket and a dashboard, not a war room.

✓ Up to 12 nines durability ✓ 6+ nines availability ✓ 1,000+ production deployments
ONE FILE, TEN SHARDS, TEN SEPARATE NODES
DirectFlow client erasure-codes 8D+2P from the page cache D0 D1 D2 D3 D4 D5 D6 D7 P Q STORAGESET · FAILURE DOMAIN · ONE SHARD PER VPOD, EACH ON A SEPARATE NODE NODE 1 NODE 2 NODE 3 NODE 4 NODE 5 NODE 6 NODE 7 NODE 8 NODE 9 NODE 10 12 ×VPOD 12 ×VPOD 12 ×VPOD 12 ×VPOD 12 ×VPOD 12 ×VPOD 12 ×VPOD 12 ×VPOD FAILED FAILED rebuild, many-to-many, into spare space Any 2 VPODs, or 2 full nodes, can fail with zero data loss. Metadata takes a separate path: three copies on three VPODs, each on a different node.
Up to 12 nines
DURABILITY
Software-defined, well past what hardware RAID can guarantee.
6+ nines
AVAILABILITY
Durability and availability are different promises. Both are stated, both are measured.
n+2
DUAL PARITY ERASURE CODING
Any two VPODs, or two full nodes, can fail with zero data loss.
1,000+
PRODUCTION DEPLOYMENTS
Twenty-five years of parallel file system heritage behind the architecture.
THE REBUILD PROBLEM

A traditional array reconstructs drives. VDURA reconstructs files.

That distinction is the whole argument. Rebuilding a drive means rebuilding its entire raw capacity, empty blocks included, through a fixed set of spindles. Rebuilding files means touching only the data that was actually there, using the write bandwidth of every node in the StorageSet.

REBUILD TIME
Degraded mode is the risk window
The odds of a drive failing this hour are the same as last hour. The longer a system sits degraded, the higher the chance a second failure lands before the first is repaired. Rebuild speed is the resiliency plan.
VDURA: Reconstruction reads from every node in the StorageSet and writes into spare space spread across all of them, so it uses the combined write bandwidth of the whole set.
BLAST RADIUS
A server is too big to be a failure domain
When the whole node is the unit of failure, a single bad device takes the entire server out of the cluster and puts all of its capacity into rebuild.
VDURA: Each storage node hosts 12 VPODs. A device failure is isolated to one VPOD, and reconstruction touches only that VPOD’s component objects rather than the node.
WASTED WORK
RAID rebuilds empty space
A traditional array reconstructs the raw contents of a drive, including blocks that never held anything. You wait for capacity, not for data.
VDURA: VDURA reconstructs files. Only the data that was actually there is rebuilt, using each file’s own erasure code.
MULTI-LEVEL ERASURE CODING

Protection across every node, and a second level inside the mixed fleet.

MLEC is VDURA's patent-protected approach to software-defined durability. Every file is protected by client-side dual-parity erasure coding across VPODs on separate nodes, on every node type. On mixed-fleet nodes, MLEC adds a second level of protection inside each VPOD. Neither level depends on RAID controllers or HA hardware.

LEVEL ONE · ACROSS VPODS · EVERY NODE TYPE
Dual parity (n+2) striping across separate nodes
Large files stripe across VPODs with P and Q parity, one shard per VPOD on separate nodes
Any two VPODs, or two full storage nodes, can fail with zero data loss
Metadata takes a separate path: three copies on three VPODs, each on a different node
Directories are triplicated with no two copies on the same Director Node
LEVEL TWO · INSIDE THE VPOD · MIXED-FLEET NODES
Local protection per device group
Each storage node hosts 12 Virtualized Protected Object Devices, each a self-contained OSD
On mixed-fleet nodes, device-level failures are absorbed inside a single VPOD instead of failing the node
Every node stores both data and metadata, shared-nothing throughout
No RAID controllers and no HA hardware anywhere in the design
The failure domain is as small as one VPOD, and as survivable as two whole servers. A device failure is contained to a single VPOD's component objects; dual-parity striping means two full nodes can fail with zero data loss.
CLIENT-SIDE ERASURE CODING

Data is protected before it leaves the compute node.

DirectFlow computes the erasure code on the client, in the page cache, then streams each shard to a different VPOD in parallel. There is no server tier in the middle computing parity, which is why a checkpoint burst gets a straight line to every SSD instead of queueing at a gateway.

The client computes parity, so no server tier sits in the data path adding latency to every write
Shards flush asynchronously and in parallel, which is what lets a checkpoint burst run at full flash speed
Every SSD runs flat out during a burst, with no gateway ceiling on aggregate throughput
Erasure coding happens before data reaches flash, so it is already protected on arrival
Standard POSIX durability semantics: write() returns fast, fsync() returns durable and erasure-coded
HOW A FILE IS PROTECTED, BY SIZE
Large files DUAL PARITY (n+2)
Striped across a set of VPODs with two parity shards, each on a separate node. Striping is also where a parallel file system gets its throughput.
Small files TRIPLE REPLICATED
Copied across three VPODs. For small objects this is both faster and more space efficient than computing parity.
Files that grow TRANSPARENT TRANSITION
A small file that grows is converted to the erasure-coded layout by the Director Node at the point where erasure coding becomes the more efficient choice. No operator action.
Erasure coding is per file, not per volume, so protection follows the data rather than the container it happens to sit in.
WHAT ACTUALLY HAPPENS WHEN IT BREAKS

Three failures, three non-events.

An SSD dies
AUTOMATIC
Cross-VPOD erasure coding absorbs it. I/O continues with no client impact, and rebuild starts immediately at the affected-stripe level only, not across the whole dataset.
WHAT YOU SEE
Alert in under two minutes via SMART and telemetry. A P1 ticket opens automatically at the VDURA NOC, with rebuild progress live in Grafana.
TIMING
An NVMe single-drive rebuild typically completes in minutes. Replacement hardware is dispatched next business day.
A node fails
AUTOMATIC
Client-side erasure coding keeps every file readable and writable. Shards from the failed node’s 12 VPODs rebuild in parallel across the surviving nodes.
WHAT YOU SEE
The node is marked failed in the VDURA Management Console. Availability is unchanged, and per-volume latency telemetry lets you verify tenant impact, of which none is expected.
TIMING
Rebuild is many-to-many, so it gets faster as the cluster grows. Availability climbs with scale.
Two nodes at once
AUTOMATIC
Dual-parity (n+2) stripes survive any two VPOD failures, or two full node failures, with zero data loss. Metadata remains triplicated across nodes throughout.
WHAT YOU SEE
The same workflow: alerts, an automatic ticket, and a live rebuild. HDD failures in a capacity tier stay inside that customer’s domain.
TIMING
Doubly-degraded stripes are rebuilt first. VDURA field engineering handles the physical swap end to end.
A failure is a ticket and a dashboard, not a war room. Self-healing VPODs, no manual rebuild steps, no downtime windows.
RELIABILITY THAT INCREASES WITH SCALE

The bigger the cluster, the shorter the exposure.

Most systems get less reliable as they grow, because more time in degraded mode means higher odds of a second failure landing while the first is still rebuilding. Recovering to full protection quickly is the entire resiliency plan.

VDURA reconstruction is many-to-many. Every node in the StorageSet contributes read bandwidth, and rebuilt VPODs are written into spare space spread across all of them. Add nodes and the degraded window gets shorter, not longer.

01
Only the affected VPODs rebuild
The system reconstructs the VPODs that were on the failed node, not the node’s entire raw capacity.
02
Every surviving node contributes reads
Erasure-coded fragments for each affected file are pulled from all the other storage nodes at once.
03
Spare space is spread across the set
Each StorageSet reserves a configurable amount of spare space on every node. Rebuilt VPODs land on a randomly chosen node in the set.
04
The degraded window shrinks
Combined write bandwidth cuts total reconstruction time, which cuts the odds of an additional failure landing during it.
PARITY VS DURABILITY

Durability is measured in nines, not parity symbols.

More parity always sounds safer, and it always costs the same way: extra capacity and write amplification on every file, forever, to shorten a risk window that rebuild speed already shrinks. Durability is a function of parity, rebuild time, failure-domain size, and scrubbing, which is why we publish the output, up to 12 nines, rather than one input. Two questions cut through any parity claim: what are your measured nines, and what exactly happens one failure past your limit? Our answers are on this page.

THE UNCOMFORTABLE QUESTION

What if three nodes fail at once?

Dual parity (n+2) protects against two simultaneous failures in a StorageSet, and the realm recovers transparently from more than two as long as no more than two are down at any one time. Past that, VDURA has one more line of defense, and more importantly, an exact answer.

It transitions to read-only
If a third storage node fails in a StorageSet while two others are being reconstructed, that StorageSet immediately goes read-only rather than accepting writes it cannot protect.
The loss is bounded and named
Only files that had VPODs on all three failed nodes lose data. Every other file in the StorageSet is either unaffected or recoverable from its erasure coding.
Bigger sets lose less
The number of files that could sit on all three failed nodes falls as the StorageSet grows, so scale works in your favor here too.
Most architectures leave you with significant uncertainty about the extent of a loss. VDURA tells you exactly which files were affected, and the number shrinks as the StorageSet grows.
WHAT THIS MEANS FOR THE BUSINESS

An SLA you can underwrite, not one you hope holds.

The SLA Is A Number
VPOD-isolated failure domains, up to 12 nines durability, 6+ nines availability, and continuous scrubbing make the commitment something finance and SRE can defend.
No HA Hardware To Buy
Reliability comes from software in a shared-nothing design: no RAID controllers, dual-ported drives, or redundant enclosures in the bill of materials.
Failures Do Not Need People
At ten thousand drives, failures are daily. HYDRA detects, isolates, rebuilds, and rebalances automatically. No tickets to work, no downtime windows to schedule.
Contained Tenant Exposure
A StorageSet is the failure domain, so an HDD problem in a capacity tier never reaches a performance tier or another customer.
Silent Corruption Gets Caught
Continuous background scrubbing reads every file and validates its erasure codes against the stored data. Rare among HPC-class storage systems.
Certainty After An Incident
You get an exact list of affected files rather than an open question. That is the difference between a disclosure and an investigation.

See what VDURA resiliency means for your uptime, and for the SLAs you offer tenants.