Glossary

Hybrid flash array

A hybrid flash array is a storage system that combines flash and hard disk drives, using flash to serve latency-sensitive I/O and disks to hold most of the capacity. Software decides which data sits on which medium, either by caching copies on flash or by moving data between tiers.

Why hybrid designs matter for petabyte estates

In a large estate only a small share of stored data is active at any moment. Recent writes, metadata and a working set of frequently read files or objects receive most of the requests, while the bulk of the data is read rarely. A hybrid design spends flash on the active part and disk on the rest, which is why hybrid systems still hold much of the capacity in data lakes, backup targets, archives and media libraries.

The cost of that economy is that performance becomes conditional. Requests served from flash complete in microseconds; requests that fall through to disk take milliseconds. How often each happens decides whether the system behaves like flash or like disk.

Caching and tiering

Hybrid systems use one or both of two mechanisms.

• Caching: flash holds copies of hot data while disk keeps the authoritative copy. Placement reacts within seconds to what is being read. Cache capacity does not add to usable capacity. The mechanism is covered under flash cache.
• Tiering: data moves between flash and disk and lives on one of them at a time. Policies based on access frequency or age relocate data in the background, typically over hours or days. Flash adds to total capacity, but placement depends on predicting what will be hot. See storage tiering.

Writes usually land on protected flash first and are destaged to disk later in larger, sequential batches. Large sequential reads often bypass flash entirely, because disks stream efficiently once positioned and caching the stream would push useful data out.

Hit ratio and effective latency

Average read latency follows from the hit ratio h, the share of reads served from flash: average = h × flash latency + (1 − h) × disk latency. With an assumed 0.1 ms for flash and 8 ms for disk:

A fall from 99% to 95% almost triples the average, and the slowest requests run at disk speed whatever the average says. The misses set both the mean and the tail.

Write destaging and sustained ingest

The flash write area absorbs bursts at flash speed. When ingest is sustained above the rate at which disks can accept destaged data, the write area fills and new writes slow to the pace of the disk layer. A hybrid system's sustained write rate is therefore the aggregate streaming rate of its disks, while its burst write rate is set by its flash.

What hybrid flash means for multi-petabyte operations

Flash is sized to the working set. A 10 PB data lake whose active data is about 200 TB can be served mostly from flash equal to 2% of its capacity. If new projects grow the working set to 1 PB, the same flash covers a fifth of it and the system behaves much more like disk, with no change in total capacity to signal the shift.

Some workloads defeat caching outright. AI training jobs read an entire dataset in shuffled order every epoch; when the dataset exceeds the flash, each block is evicted before it is read again and the hit ratio collapses. Full analytic scans, large restores and migrations behave the same way. For these jobs throughput comes from the disk layer spread across many drives and servers, and the drive count matters more than the flash.

Tiering lags demand. A tiering engine that relocates data on a nightly or weekly schedule leaves a newly active dataset on disk until the next cycle, so a project that starts reading archived data sees disk performance for hours or days before the policy catches up.

Metadata benefits most consistently. In object and file systems holding billions of items, a metadata lookup precedes every data read, and keeping metadata on flash removes a disk seek from every request while bulk data stays on disk.

Failures expose the design. After a node replacement or failover the new cache starts cold, and until it warms the workload sees the performance of the disk layer underneath it.

Hybrid servers in Scality RING

Scality RING is software on standard x86 servers, so the balance of flash and disk is a property of the servers chosen for each deployment. The large RING customers Scality describes run hybrid storage servers that mix flash and hard drives; one bank's fraud detection data lake in that account reads at 80 GB/s per site. Where the capacity layer itself needs to be flash, RING XP runs the same software on NVMe-only servers.