Search corpus

Flash memory that stacks cells vertically in layers to increase density without shrinking planar features alone.

A cross-point non-volatile memory technology announced by Intel and Micron, distinct from NAND flash in cell physics.

4Kn exposes 4096-byte logical sectors and 4096-byte physical sectors without 512-byte sector emulation.

512e exposes 512-byte logical sectors over 4096-byte physical sectors.

In Ceph, the acting set is the ordered OSD list responsible for a placement group at a particular map epoch.

The specified logical-block address interval available to a storage test or its preparation workload.

An active-zone limit caps the zones simultaneously in implicitly open, explicitly open or closed states in a zoned namespace.

Adaptive Read-Retry selects shorter NAND precharge timing from profiled conditions while reserving error-correction margin.

Adaptive Replacement Cache adjusts a recency/frequency balance using resident lists and nonresident histories.

Adaptive scrubbing adjusts the rate or schedule of proactive storage verification in response to observed conditions or estimated error risk.

The additional sense code (ASC) is an eight-bit SCSI sense-data field that refines the sense key and is interpreted together with its qualifier.

The additional sense code qualifier (ASCQ) is an eight-bit SCSI sense-data field interpreted with the ASC to identify a more specific condition.

Advanced Format is a disk-sector architecture expanded beyond legacy 512-byte sectors, with 512e and 4Kn as its familiar 4K variants.

Advanced Host Controller Interface, the PCI class driver model SATA SSDs and HDDs used before NVMe dominance.

Offline copies physically isolated from production networks to resist ransomware.

A storage array built entirely from flash modules without HDD tiers.

Asymmetric Logical Unit Access is the SCSI mechanism for reporting logical-unit access states through target port groups.

An ANA group is a set of NVMe namespaces whose ANA access states change together for a given controller.

Annual replacement rate (ARR) counts recorded device replacements relative to device exposure, scaled to one year.

Annualized failure rate (AFR) expresses observed failures relative to drive-time exposure, scaled to a year, and is not inherently a one-year failure probability.

Anti-entropy repair compares replica state over a selected scope and synchronizes differences to reduce divergence.

Application quiescing is application-specific preparation and temporary restriction of writes to establish a defined state for snapshot or backup capture.

An application-consistent snapshot captures an application's data through a coordinated procedure that satisfies that application's consistency requirements.

Bits stored per unit of disk surface area, the central metric driving HDD capacity growth.

Asymmetric Namespace Access is the NVMe mechanism for reporting namespace access states through different controllers.

Asynchronous commit is an acknowledgment mode that can report a transaction as successful before its recovery log has reached durable storage.

Asynchronous DRAM Refresh is a platform power-fail mechanism that drains covered memory-controller write-pending buffers to persistent memory without protecting dirty CPU caches.

Asynchronous replication allows primary-side completion without waiting for the corresponding update to reach a specified processing stage on a replica.

ATA SMART is the ATA interface for retrieving drive self-monitoring attributes, status and logs from supporting devices.

Atomic rename is a filesystem namespace operation whose successful replacement of an existing destination changes the file reached through that pathname without exposing an intermediate missing destination to concurrent lookups.

Authenticated encryption combines confidentiality with cryptographic integrity protection for the protected message.

Non-interpolated average precision summarizes ranked prediction scores by weighting precision at each threshold by the corresponding increase in recall.

Background fsck is a filesystem-specific consistency-checking mode that performs permitted repairs while the filesystem remains in use.

A backup is a retained copy of data or software intended to support later recovery.

A backup manifest is a machine-readable inventory of expected backup contents and associated verification metadata.

Backup verification checks stored backup material against specified integrity and completeness rules.

Firmware maps out factory and grown NAND defects so hosts see a contiguous logical address space.

Modern tape coatings enabling higher coercivity and capacity in LTO generations.

BD-R HTL is a recordable Blu-ray type whose written marks have lower optical reflectivity than the unrecorded material.

BD-R LTH is an organic-dye recordable Blu-ray type whose written marks have higher optical reflectivity than the unrecorded material.

Bit error rate (BER) expresses erroneous bits relative to the total bits considered at a specified observation point.

A proposed HDD technology isolating magnetic islands lithographically instead of continuous grains.

Fixed-size addressable blocks presented to hosts or hypervisors without file semantics.

A compact approximate-membership structure that can reject absent keys but may report a false positive, without false negatives for correctly inserted keys in an intact standard filter.

A Blu-ray data archive stores retained digital content on recordable BD-R or rewritable BD-RE optical media.

Bounded crash testing systematically explores recovery behavior within an explicitly limited set of workloads, initial states and crash points.

Btrfs is a Linux filesystem built around copy-on-write B-trees, with integrated snapshots, checksumming and multi-device storage.

Btrfs NODATASUM marks file data as excluded from the filesystem's data checksumming.

A Btrfs read-only subvolume is a subvolume whose ro property prevents modification of its file contents and directory hierarchy.

The Btrfs tree log is a per-subvolume recovery structure that can persist file or directory changes for fsync without a full filesystem transaction commit.

Non-volatile magnetic memory using cylindrical domains; briefly shipped before cheap DRAM and flash.

Buffered I/O is file access through an operating-system cache, allowing transfers between an application and storage to be decoupled.

Byte-addressable storage exposes individual byte locations rather than only block-sized addresses, without by itself specifying atomicity or durability.

Cache pollution occurs when low-reuse entries displace entries that would otherwise serve useful cache hits.

CD BLER counts blocks containing detectable errors per second at the input to the C1 decoder.

A removable card standard using PCIe and NVMe for high-end cameras and embedded recorders.

Chain replication orders replicas so that updates enter at a head, propagate along the chain and complete at a tail that also serves queries.

A flash cell design that stores charge in localized traps within an insulating material rather than on a conductive floating gate.

Checkpoint starvation is sustained inability to finish or reset a database log checkpoint because the required progress conditions keep being obstructed.

A checksum is a value computed from specified data and later compared with a reference value to detect changes.

Chunk fragmentation is the dispersion of chunks needed by a deduplicated restore across storage containers or an order poorly matched to the restore sequence.

Count-key-data block layout used on classic mainframe disks unlike fixed SCSI sectors.

Class prevalence is the proportion of observations belonging to a specified class in a defined population or evaluation sample.

Clear sanitization uses logical operations across all user-addressable storage locations to resist simple, non-invasive recovery through the device interface available to the user, such as the host interface.

Close-to-open consistency is a file-sharing cache policy that combines writeback at close with validation at a later open to support visibility of updates between sequential users.

A closed-loop load model uses a fixed client population whose next request follows completion of a previous request, optionally after think time.

A storage ECC codeword is an encoded unit representing a data chunk together with error-correction redundancy.

Offline or very low-cost storage tier optimized for rare reads and long retention.

Compare-and-swap is an atomic conditional update that replaces a stored value only if it matches an expected value and reports the comparison outcome or observed value.

An ARIES compensation log record describes a rollback action for redo and records where undo processing should continue.

Completion latency in fio is the engine-dependent interval from submission of an I/O unit to observation of its completion.

An SSD or NVMe device that runs limited compute near data to reduce data movement.

A conditional write changes a resource only when a server-evaluated precondition, such as a matching entity tag or the absence of a current object, holds.

An SSD intended for personal computers and other client devices rather than shared server workloads.

Container capping limits the old containers referenced by a backup segment by storing selected duplicate chunks again.

Content-defined chunking divides a byte stream using content-dependent boundary rules rather than only fixed offsets.

Perpendicular magnetic recording with non-overlapping tracks so any sector can be rewritten in place.

Coordinated omission is a measurement bias in which a workload generator or sampler misses observations because its activity depends on the slow responses being measured.

copy_file_range is a Linux system call that copies a requested file range in the kernel and reports the number of bytes actually copied.

Copy-on-abundant-write is a research cloning policy that buffers changes to postpone unsharing until enough work accumulates to justify copying.

A technique that writes modified blocks to new locations instead of overwriting in place.

Crash consistency is the property that persistent state remains within an application's allowed recovery states after an unexpected interruption.

A crash-test oracle is a checker that decides whether a recovered state satisfies the specified correctness and durability contract.

CRUSH is a deterministic, pseudorandom placement algorithm that maps an identifier to storage targets using a weighted cluster hierarchy and placement rules.

Cryptographic erase sanitizes the keys protecting encrypted target data so that recovering the plaintext becomes infeasible.

A cryptographic hash function deterministically produces a digest and is designed to resist preimage, second-preimage and collision attacks.

A cryptographic nonce is an input value whose required uniqueness scope is specified by the cryptographic construction that uses it.

Compute Express Link attaches byte-addressable memory or accelerators on PCIe, blurring DRAM and storage tiers.

Digital Audio Tape format adapted for small-office digital backup in the 1990s.

Data deduplication replaces duplicate content with references to shared stored blocks or chunks within a defined matching scope.

Media sanitization is a process that makes recovery of target data infeasible for a specified level of effort.

Data scrubbing is proactive verification of stored data to discover errors before ordinary access requires the affected content.

A database checkpoint records engine-defined recovery state to bound or reduce the work needed after a crash.

Database byte write amplification is storage-engine output bytes written divided by logical input bytes over a stated interval and accounting boundary.

The interface-defined data or error returned when a host reads a deallocated logical block.

Degaussing uses a magnetic field to disrupt recorded magnetization and can sanitize suitable magnetic media when the process is matched to the medium.

A degraded array is a redundant storage array operating with less than its configured healthy redundancy because one or more members are unavailable or not fully synchronized.

A delete marker is a payload-free S3 version that makes a key appear deleted when the marker is current.

In delete-aware LSM research, delete persistence latency is the delay from inserting a tombstone to removing the entries it invalidates through compaction.

Destroy sanitization makes target-data recovery infeasible even with current state-of-the-art laboratory techniques and leaves the medium incapable of storing data again.

An SMR HDD that presents a conventional LBA interface while the firmware relocates writes across shingled bands.

A differential backup captures changes accumulated since its designated full-backup base, rather than only changes since the preceding differential backup.

Differential observability is a difference between entities' observations of the same system's health.

Direct Access is a Linux mechanism that lets suitable memory-like storage serve file data without an intervening page-cache copy.

Direct I/O is a file-access mode, such as Linux O_DIRECT, that requests I/O through user-space buffers while minimizing use of the local operating-system page cache.

Direct-I/O alignment is the per-file constraint on memory addresses, file offsets and I/O segment lengths used for direct transfers.

Directory fsync is synchronization of a directory descriptor to persist directory state, including changes to the names through which files are reached.

A block-storage request indicating that specified logical ranges are no longer needed by the caller.

Disk defect skipping formats a record around a known defective medium region so that normal reads bypass a gap instead of storing customer data there.

An enclosure providing power, cooling, and SAS/NVMe connectivity to many drives.

A distributed snapshot records a coordinated global state of participating processes and their communication channels.

A distributed spare is replacement capacity spread across multiple devices rather than reserved on a single dedicated spare device.

Digital Linear Tape, a mid-range backup cartridge family superseded by LTO in many sites.

dm-crypt is a Linux device-mapper target that transparently encrypts and decrypts block-device I/O using the kernel cryptographic API.

dm-integrity is a Linux device-mapper target that stores per-block integrity tags and can generate and verify them when configured with an internal hash.

dm-integrity recovery mode is a Linux device-mapper access mode that bypasses integrity checks and journal replay while disallowing writes through the mapping.

dm-verity is a Linux device-mapper target that verifies a read-only block mapping against a trusted cryptographic root hash.

A drive-day is one drive observed in operation for one day, used as a unit of fleet exposure.

Early rotating magnetic drum storage where data was recorded on the outer surface of a metal cylinder.

A hard drive with two independent actuator assemblies serving different platter zones to roughly double sequential throughput.

A drive with two independent data paths to the same media for multipath HA in arrays.

DVD PI Sum8 counts parity-inner rows containing errors over eight consecutive ECC blocks before correction.

Drive Writes Per Day rating expresses how many full-capacity writes an enterprise SSD can sustain daily for a warranty period.

Dynamic layout score is a workload-specific locality measure based on the fraction of traced filesystem block requests that are contiguous in logical block address space.

A short EDSFF SSD form factor aimed at 1U servers with power and thermal profiles between M.2 and U.2.

Early retention loss is the rapid initial charge-loss behavior observed after programming some NAND cells.

A family of ruler and short SSD form factors for dense NVMe deployment in servers and JBOD enclosures.

End-to-end data protection is an integrity scheme whose generation and verification points span an explicitly defined data path.

An enterprise SSD is a solid-state drive intended for server, storage-array or other multi-user storage environments.

An entity tag is an opaque HTTP validator assigned by a server to a selected representation for cache validation and conditional requests.

Erasure coding encodes data with redundancy so that missing fragments at known locations can be reconstructed from a sufficient, code-dependent set of surviving fragments.

Error-and-erasure decoding reconstructs a codeword while accounting for both known erased positions and incorrect values at unknown positions.

An error-correcting code adds structured redundancy to data so a decoder can detect and correct specified error patterns.

errseq_t is a Linux kernel datatype that records an error and sequence state so observers can detect failures since their saved cursor.

Earlier IBM optical channel link superseded by FICON for mainframe DASD.

Enhanced Small Disk Interface, a faster successor to ST-506 with integrated encoding on the drive.

Ext4 fast commit is a journaling mechanism that records metadata deltas for eligible updates instead of a full traditional metadata-block commit.

An ext4 orphan file is an internal metadata file that stores inode-number slots for orphan cleanup when the orphan_file filesystem feature is enabled.

An ext4 orphan inode is an inode tracked so recovery can finish deletion or remove excess allocations left by an interrupted truncation or extension.

Extended Asynchronous DRAM Refresh is a platform capability that extends persistent-memory power-fail protection to CPU caches.

F1 score is the harmonic mean of precision and recall for a specified positive class and decision threshold.

F2FS is a Linux log-structured filesystem designed for flash storage devices that expose a block interface.

An F2FS checkpoint records a consistent filesystem recovery point, including information selecting valid metadata-table versions.

F2FS roll-forward recovery reconstructs selected post-checkpoint file updates from recovery-marked node blocks.

Fail-slow describes a component that remains operational but performs more slowly than expected for its operating conditions.

Under an adjacent-gap rule, a failure burst is a maximal sequence of event starts in which every consecutive gap is no greater than a specified window.

A failure domain is a group of components that can be affected by the same failure event because they share a dependency or physical location.

A failure label is a recorded outcome assigned to an observation under an explicit rule defining which event counts as failure and when it occurs.

A failure operation ticket is a management-system record of a detected hardware problem and its handling status, rather than necessarily a unique failed component.

The false positive rate is the fraction of observed negative cases incorrectly classified as positive at a specified decision threshold.

FastCDC is a content-defined chunking approach that combines Gear-based boundary selection, sub-minimum skipping and chunk-size normalization.

Fault injection is a testing method that deliberately introduces specified faults into a system to observe detection, propagation and recovery under a defined workload.

Fibre Channel over Ethernet tunnels FC frames on lossless Ethernet for converged data center networks.

fdatasync is a file-synchronization operation that persists modified data and the metadata required to retrieve it, without requiring synchronization of all other file metadata.

A fencing token identifies an ownership generation that a receiving service checks to reject operations from an obsolete owner.

A high-speed serial fabric used to attach block storage in enterprise SAN environments.

Fibre Channel attachment for IBM mainframe storage channels.

A filesystem mapping for a contiguous range of file blocks and the storage blocks backing that range.

Reserving filesystem allocation for a file range before the application writes that range's data.

Shared hierarchical files and directories over NFS, SMB, or similar protocols.

Filesystem aging is the evolution of allocation layout through file operations that can degrade performance even when the logical contents remain comparable.

A filesystem consistency invariant is a rule relating filesystem metadata that must hold at a defined consistency boundary.

A filesystem freeze temporarily blocks modifications after completing ongoing filesystem transactions and flushing dirty filesystem data, metadata and logs to storage.

SCSI-like fixed sectors adopted on mainframe storage for simplification and FBA volumes.

Flash data retention is the ability of programmed NAND data to remain recoverable over time without refresh under specified wear, storage and read conditions.

A replaceable package of NAND dies and controller used in arrays and custom cards.

A flash translation layer maps host logical addresses to physical flash locations and manages flash-specific operations such as reclamation.

A flash cell design that stores charge on an electrically isolated conductive gate to alter the transistor's threshold voltage.

Removable flexible magnetic disk media used for software distribution and PC backup before optical and flash.

Spindle motor bearings using lubricant films to reduce vibration in modern HDDs.

Flying height is the physical clearance between a flying HDD head/slider region and the disk surface.

Force Unit Access is a write-request property requiring the flagged data to reach nonvolatile storage before that request is reported complete.

A forward assembly area buffers requested chunks in future restore-output order using advance knowledge from the backup recipe.

Free-space fragmentation is the division of a filesystem's available allocation space into separated contiguous runs.

fs-verity is a Linux filesystem facility that verifies the immutable contents of selected files against a cryptographic file digest.

fsync is a file-synchronization operation that requests persistence of a file's modified data and associated metadata before reporting completion.

An FTL translation page is a flash page containing address-mapping entries.

A full backup captures every data object within a defined backup scope, whether or not it has changed since an earlier backup.

A PostgreSQL full-page write records a recoverable page image in WAL to support recovery from an incomplete data-page write.

A parity-array write supplying replacement data for every data position of a stripe so its new parity can be computed without reading the prior stripe contents.

A fuzzy checkpoint records recovery information while transaction processing continues, rather than freezing all activity into one instantaneous database image.

Flash garbage collection reclaims a victim erase block by first preserving its valid data elsewhere and then erasing the block for reuse.

A ghost cache entry records the identity or history of an evicted item without retaining its cached payload.

Glass disks replace aluminum on high-capacity platters for better flatness and thermal stability.

Experimental archival storage that encodes data as structural changes inside glass for very long-duration, read-mostly retention.

Giant magnetoresistive heads enabled areal density gains in the late 1990s HDD era.

A gray failure occurs when an application observes a system as unhealthy while that system's failure observer regards it as healthy.

A flash block-cleaning policy that selects a candidate erase block with the fewest valid pages, minimizing the valid data copied for that selection.

Group commit is a logging optimization that shares a durable log synchronization across multiple concurrent transaction or write requests.

Grouped cross-validation evaluates a predictor using splits that keep all observations from each specified group together in either training or testing within a split.

A non-volatile storage device that records data on spinning platters with a moving read/write head.

A hard link is a directory entry naming an existing file's inode, so multiple hard links refer to the same file rather than to separate copies.

A hash collision occurs when two distinct inputs produce the same digest under the same hash function and parameters.

The hazard rate is the instantaneous event rate at a given time conditional on the event not having occurred before that time.

Head slap is a head-disk contact mechanism in which an HDD slider separates from the disk and then snaps back into it.

The head-disk interface is the interacting region between an HDD's head-carrying slider and rotating disk surface.

An HDD technology that briefly heats the media with a laser so bits can be written on high-coercivity material at higher density.

Tape heads scan at an angle across a slow-moving tape, used in DAT and some video formats.

A diagnostic port on helium drives to verify seal integrity over the product lifetime.

A sealed hard drive filled with helium to reduce platter drag and allow more disks in the same form factor.

Hinted handoff retains information about an update missed by an intended replica and attempts to deliver it later.

Deallocating whole filesystem blocks within a file range while preserving the file's logical length and making the affected range read as zeros.

Stores pages of data as interference patterns in a photosensitive volume, mostly experimental commercially.

A card or SoC port that connects servers to SCSI, SAS, Fibre Channel, or NVMe fabrics.

Host Memory Buffer is an optional NVMe feature through which a controller uses assigned host-memory ranges.

The host tracks which SMR bands are full and schedules sequential writes into empty zones.

An SMR HDD whose shingled bands are exposed to the host, which must write sequentially within zones.

Host filesystems or FTLs relocate valid pages before erasing flash blocks.

A hot journal is a SQLite rollback journal that requires replay to recover an interrupted database transaction.

A hot spare is a standby device or reserved distributed spare capacity available to take over a failed member's role in a redundant array or pool.

Low-latency media and tiers serving active workloads with frequent reads and writes.

A data-placement model where hot tiers favor latency, warm tiers favor cost per TB on disk, and cold tiers favor tape or archive media.

A flash-placement strategy that groups data by update frequency so frequently overwritten pages and long-lived pages need not share the same cleaning pool.

A storage system mixing flash caches or tiers with HDD capacity pools.

I/O fencing enforces exclusion of a node from shared storage so that recovery cannot be undermined by its continued access.

I/O request size is the number of bytes requested by one input/output operation at a specified system boundary.

An I/O timeout is expiration of a layer's waiting limit for an input/output operation, which may initiate recovery without proving that the operation never executed.

The first commercial moving-head disk storage product, shipped with the IBM 305 RAMAC system.

Integrated Drive Electronics places the disk controller on the drive, the ancestor of PATA and SATA.

Idempotency is the property that repeating the same operation with identical inputs has the same intended effect as performing it once.

An idempotency key identifies attempts of one logical request so a service can recognize retries within its documented scope and retention rules.

The standards body maintaining SCSI command sets used by SAS, FC, and iSCSI block devices.

An incremental backup records changes relative to an earlier backup state and reuses retained backup data to represent unchanged content.

Informative censoring occurs when the observation-ending process conveys information about the event time beyond what the survival analysis accounts for.

An inode is a filesystem object that holds a file's metadata and has a filesystem-local number, distinct from the directory names and open file descriptions that refer to the file.

Intel's brand for storage products built on 3D XPoint non-volatile memory before the business was wound down.

An interarrival time is the elapsed interval between consecutive events in a specified event stream.

An invertible Bloom lookup table is a randomized data structure that supports updates and can list encoded entries with high probability under its sizing assumptions.

io_uring is a Linux asynchronous I/O interface that exchanges requests and completion results through shared submission and completion rings.

An io_uring completion queue entry is a result record containing an operation result, flags and application-provided identification data.

io_uring drain is a submission flag that delays a request until earlier requests in the same ring finish and delays later requests until the drain request finishes.

io_uring integrity metadata is a separately described protection-information buffer attached to a read or write request through SQE attributes.

io_uring linked requests are a chain of submission entries whose dependencies delay each successor until its predecessor completes.

Observed IOPS is the number of completed input/output operations divided by the measurement duration at a specified system boundary.

SCSI block commands carried over IP networks, allowing SAN-like access without dedicated FC hardware.

Just a Bunch Of Disks presents individual drives without RAID virtualization in the enclosure.

JEDEC publishes endurance and environmental specs referenced by enterprise SSD vendors.

In JBD2 terminology, a journal checkpoint writes committed journaled state to its home locations so the corresponding log space can be reclaimed.

A journal revoke record identifies a block whose older journaled images must not be replayed over newer uses of that block.

The Kaplan-Meier estimator is a nonparametric estimator of the survival function that multiplies conditional survival estimates at observed event times while accounting for right-censored observations.

Key-value separation stores values outside the main key index and retains references in that index.

A drive exposing key-value semantics instead of fixed-size LBAs for database and cache offload.

A latent sector error is an inaccessible sector whose failure remains undiscovered until that location is accessed.

Layer-to-layer process variation is fabrication-related variation in cell characteristics between vertical layers of a 3D NAND die.

An LDPC code is an error-correcting linear block code defined by a sparse parity-check matrix.

Least recently used replacement evicts the resident entry whose most recent access is oldest.

Left-truncation is selection into a time-to-event sample only after a unit has survived to its observation-entry time.

A storage legal hold prevents deletion or replacement of protected data until an authorized action removes the hold, rather than expiring on a preset date.

Leveled compaction organizes sorted data into levels with size targets and merges selected data with overlapping key ranges in a destination level.

Linearizability requires concurrent operations on an object to behave as a legal sequential execution that preserves the real-time order of non-overlapping operations.

The Linux block-device read-only flag is a software-controlled device state exposed through block-device ioctls to restrict write access.

Little's law relates time-averaged occupancy L to the operation rate lambda and mean residence time W within the same system boundary.

A locally repairable code is an erasure code that provides recovery of specified missing fragments from a small subset of surviving fragments.

A log sequence number identifies a position in an engine's ordered recovery-log address space.

A log-likelihood ratio is the logarithm of the relative likelihood of two alternatives, used in decoding to express a bit decision and its confidence.

A write-oriented index that buffers updates and merges sorted components over time, trading write batching against read and compaction work.

Logical decoding reconstructs database changes from a storage-level log into a representation that a consumer can interpret without the database's physical storage format.

Logical journaling records recoverable changes to logical objects or operations rather than simply reproducing their on-disk representation.

A logical sector is the addressable data unit selected by a logical block address at a storage interface.

A SCSI addressable block device presented by a target to initiators on a SAN.

Magnetic recording where bit magnetization lies in the plane of the disk, largely replaced by PMR for capacity.

A lost write is a write reported as completed whose intended update is absent at the required storage location.

LSM compaction reorganizes sorted storage components, commonly merging selected runs or SST files and removing obsolete entries when visibility rules permit.

An LSM tombstone is a versioned deletion marker that suppresses older values for a key without immediately erasing their stored bytes.

Linear Tape File System is a self-describing tape format that stores files and directory metadata for interchange between compatible implementations.

An LTFS data placement policy specifies which file data may be stored in the index partition alongside filesystem metadata.

An LTFS index is XML metadata describing files, directories and the mapping from file contents to recorded tape extents.

An open tape format standard for high-capacity cartridges, advanced by the LTO Program for backup and archive.

LUKS is a disk-encryption format that records cryptographic parameters and key-management metadata, including keyslots used to unlock encrypted storage.

The Linux Logical Volume Manager, a device-mapper framework that builds resizable virtual block devices (logical volumes) out of one or more physical disks or partitions.

M-DISC is a family of write-once optical media marketed for archival storage using an inorganic recording layer.

A small form-factor card standard commonly used for NVMe and SATA SSDs in laptops and desktops.

Sequential storage on linear magnetic tape in removable cartridges, used mainly for backup and cold archive.

Optical-magnetic media written with a laser and magnetic field, used for archival and removable storage.

Direct-access storage devices attached to IBM Z systems, historically count-key-data disks.

MAP_SYNC is a Linux mmap flag for supported DAX files that provides a file-offset persistence guarantee used with appropriate CPU persistence instructions.

A maximum distance separable code achieves minimum distance n - k + 1 for length n and dimension k, allowing recovery from any n - k known symbol erasures.

Mean time to data loss (MTTDL) is the expected time until a storage system first loses data that cannot be retrieved or reconstructed.

Mean time to detection is the arithmetic mean of elapsed times between event occurrence and detection for a specified event population.

Medium Auxiliary Memory is nonvolatile storage associated with an individual storage medium, used to hold auxiliary attributes about the medium and its recorded data.

Resistive switching devices researched as dense memory; not mainstream datacenter bulk storage.

A memtable is an in-memory structure that holds recent LSM-tree updates before they are flushed to sorted table files.

A Merkle tree combines hashes of data and child nodes into a root so that selected entries or subtrees can be checked without transferring the entire data set.

A metadata update dependency is an ordering requirement between persistent changes needed to preserve a filesystem invariant.

Earlier tape coatings using magnetic metal particles before barium ferrite advances.

An energy-assisted HDD approach that uses a spin-torque oscillator to lower write coercivity with microwaves instead of laser heating.

A mirror set is a group of storage members configured to maintain copies of the same logical data.

Storage mirroring maintains a copy of source data on another storage device or system.

A misdirected write stores an update at the wrong address or on the wrong device.

High-RPM enterprise HDDs for transactional workloads before flash adoption.

SSDs using two-bit-per-cell NAND, now largely replaced by TLC in consumer lines.

Monotonic reads is a session guarantee requiring successive reads to observe states that include the effects of writes relevant to the session's earlier reads.

Monotonic writes requires every replica incorporating a later session write to incorporate the session's earlier writes and order them before it.

msync is an operation that synchronizes a range of a shared file-backed memory mapping with its underlying file, with MS_SYNC waiting for completion.

Mean time between failures (MTBF) is the expected operating time between successive failures of a repairable system under specified conditions.

NAND flash that stores two bits per cell using four distinguishable threshold-voltage states.

Multipart upload is an object-transfer process that sends numbered parts independently and assembles them into one object when the upload is completed.

Multipath I/O manages multiple paths to the same storage device for failover and, where supported, load distribution.

A NAND channel is a communication path between an SSD controller and flash components sharing its transfer resources.

NAND copyback relocates page data through an on-chip buffer to a destination page without transferring the payload to an external controller.

The NAND cross-temperature effect is the change in read behavior associated with programming and subsequently reading a cell at different temperatures.

Stacking multiple flash dies in one package to increase capacity per socket.

A NAND erase block is a group of flash pages erased together by a block-erase operation.

A non-volatile memory type arranged in NAND strings, forming the storage media in most SSDs and flash cards.

A NAND page is a group of data bits addressed together for a page read or program operation within an erase block.

NAND page pairing is the grouping of pages whose bits occupy the same physical memory cells.

NAND program interference is an unintended shift in a cell's threshold voltage caused by programming neighboring cells.

A NAND read-reference voltage is a sensing voltage used to distinguish stored cell states during a read operation.

NAND read-retry rereads a flash page with adjusted sensing settings after an unsuccessful error-correction attempt.

A NAND wordline is a shared control-gate connection used to address a group of flash memory cells.

Network-attached storage serves files over IP to clients without block-level host attachment.

Native Command Queuing (NCQ) allows multiple tagged ATA commands to be outstanding so a supporting device can schedule their execution.

A class of high-capacity, lower-RPM hard drives intended for always-on but not primary-tier storage.

Lower-RPM SAS drives tuned for capacity and always-on use in servers.

Network File System is a family of protocols for accessing files and directories on a remote server over a network.

An NFS attribute cache is a client-side copy of file or directory metadata reused until refresh or revalidation is required.

NFSv3 COMMIT requests stable storage of previously buffered writes and associated metadata for a specified file range.

An NFS lookup cache retains directory-name resolution results, including found and missing entries, subject to revalidation.

An NFSv3 write verifier is a server cookie returned by WRITE and COMMIT that lets clients detect state changes that may have lost uncommitted writes.

An NFSv4 file delegation is a recallable server grant allowing a client to manage specified opens and cached file state locally while conflicting access is coordinated by the server.

no_path_retry is a DM Multipath setting that controls queueing behavior when every path to a mapped device is unavailable.

Noise-predictive maximum-likelihood (NPML) detection incorporates noise prediction or whitening into sequence-detector branch metrics for a partial-response recording channel.

A random-access flash memory type often used for firmware execute-in-place rather than bulk SSD storage.

Normalized magnitude of data loss (NOMDL) is the expected number of bytes lost per usable terabyte within a stated mission time.

A DIMM form factor combining DRAM with flash backup to persist memory contents across power loss.

An NVMe system containing one or more controllers, domains and ports, with zero or more namespaces.

A host controller interface and command set optimized for PCIe-attached non-volatile memory such as NAND SSDs.

NVMe Available Spare is the normalized percentage of spare capacity remaining in the NVM subsystem.

Midplane wiring routing PCIe signals from server motherboards to U.2 or EDSFF drive bays.

The core NVM Express document defining admin and I/O command sets for PCIe SSDs.

NVMe Composite Temperature is an implementation-specific temperature value in kelvins representing the controller and its associated namespaces in the SMART / Health Information log.

NVMe Controller Busy Time is the cumulative number of minutes during which the controller has an outstanding I/O command, as defined by the SMART / Health Information log.

NVMe Critical Composite Temperature Time is a cumulative SMART / Health log duration in minutes of controller operation with Composite Temperature at or above the critical composite-temperature threshold.

NVMe Critical Warning is the SMART / Health Information log bitmask identifying controller warning conditions at the time the log is read.

NVMe Data Units Written is the SMART / Health Information log counter of host-written data, excluding metadata, in rounded-up groups of 1,000 units of 512 bytes.

NVMe Device Self-test is an optional controller-executed diagnostic capability with short and extended operations whose namespace scope is selected by the initiating command.

The NVMe Device Self-test log is log page 06h, which reports current operation progress and twenty slots for results of completed or aborted self-test operations.

An NVMe Endurance Group is a scope within which an NVM subsystem manages endurance.

NVMe Error Information Log Entries is the SMART / Health Information log counter of Error Information log entries accumulated over the controller's life.

An optional NVMe capability that lets a host guide write placement through placement identifiers while retaining conventional logical-block reads and writes.

NVMe Host Controlled Thermal Management is a controller feature through which a host configures temperature targets for automatic transitions to lower-power active power states or vendor-specific thermal actions.

NVMe Media and Data Integrity Errors is the SMART / Health Information log counter of unrecovered data-integrity errors detected by the controller.

Using multiple routes to the same NVMe namespace for path failover and, where supported, load distribution.

An NVMe multipath I/O policy is a host-side rule for choosing among eligible routes to one namespace.

An NVMe resource collection that a host can address through a namespace identifier, such as formatted non-volatile storage.

An extension that carries NVMe commands across RDMA, Fibre Channel, or TCP networks to remote namespaces.

NVMe Percentage Used is a vendor-specific estimate of the percentage of NVM endurance consumed by actual use.

An FDP value combining a reclaim group identifier with a namespace-scoped placement handle that maps to a reclaim unit handle.

An FDP storage abstraction within a reclaim group that the controller can physically erase independently of other reclaim units.

An FDP controller resource that references one reclaim unit in each reclaim group of an endurance group.

NVMe Sanitize starts a controller-managed sanitization operation using a supported Block Erase, Overwrite or Crypto Erase action.

The NVMe SMART log is the standardized log page containing controller health indicators and cumulative usage, endurance and error counters.

NVMe Unexpected Power Losses is the SMART / Health Information log counter of main-power losses that meet the specification's incomplete-shutdown conditions.

NVMe Warning Composite Temperature Time is a cumulative health-log duration in minutes governed by composite-temperature thresholds and, in supporting revisions and configurations, temperature hysteresis.

NVMe Management Interface for out-of-band monitoring and firmware management of SSDs and enclosures.

O_DSYNC is an open-file status flag requiring write operations to complete with synchronized I/O data integrity, covering data and the metadata needed to retrieve it.

O_SYNC is an open-file status flag requiring write operations to complete with synchronized I/O file integrity, covering data and associated file metadata.

An object lock enforces retention restrictions on a specified object version through a retention policy or hold.

Object storage organizes data as independently addressed objects with associated metadata, accessed through object identifiers or keys rather than a block-device interface.

An object version ID is a service-assigned identifier used with a bucket and key to select an S3 object version.

Object versioning retains separately identifiable revisions of an object under the same key.

An open file description is a system-wide open-file object referenced by one or more file descriptors and holding state such as the current file offset and file status flags.

An open-channel SSD exposes media parallelism and placement controls so host software can manage data placement and I/O scheduling.

An open-loop load model schedules request arrivals independently of earlier request completions.

An open-zone limit caps the zones simultaneously in implicitly open or explicitly open states.

Optical archival storage preserves data on removable laser-readable media as part of a long-term retention and migration workflow.

Robotic libraries swapping optical discs for archival retrieval.

An optical-disc lifetime estimate predicts time to a defined error threshold for a stated population percentile and storage environment under an aging model.

SSD over-provisioning is flash capacity withheld from host-addressable user storage to support controller management such as garbage collection and wear leveling.

The page cache is the kernel-managed memory that holds cached file contents for access and writeback.

A page LSN records the log sequence number of the latest logged change reflected in a database page under the engine's recovery protocol.

Parity declustering places parity groups across a larger device pool to distribute reconstruction work.

Parity pollution is propagation of incorrect data into redundancy information through parity calculations.

Parity RAID uses computed redundancy across striped data to reconstruct missing chunks without storing complete mirrored copies.

Partial Parity Log is a Linux MD RAID5 mechanism that records the XOR of untouched stripe chunks before dispatching data and parity updates.

A partial quorum configuration allows a read response set and a write acknowledgement set that need not overlap.

Partition alignment places a partition's starting offset on boundaries appropriate to the storage topology reported by the underlying stack.

An SST membership-filter layout split into independently loadable partitions located through a top-level index.

A parallel attachment standard for storage devices, dominant in PCs before SATA.

Fourth-generation PCIe roughly doubles per-lane bandwidth used by NVMe SSDs.

Fifth-generation PCIe further increases lane bandwidth for next-wave NVMe drives.

A PCIe lane is one pair of transmit and receive signal paths between two PCI Express ports; a link bundles 1, 2, 4, 8 or 16 lanes, so the lane count (x1 to x16) together with the PCIe generation sets the peak bandwidth available to an SSD controller.

A signal conditioner extending PCIe trace length in servers with many NVMe slots.

Use of PCI Express lanes as the electrical transport for modern high-performance SSDs, usually with NVMe as the protocol.

Ceph peering establishes agreement among OSDs about a placement group's object state and associated metadata.

A magnetic recording orientation where bits stand perpendicular to the platter, enabling higher areal density than longitudinal recording.

Recording orientation on tape analogous to PMR on disks for higher density.

A persistence domain is the set of components whose stored contents are guaranteed to survive a specified failure under the platform's operating contract.

Persistence ordering is a constraint on the order in which updates become durable, rather than merely the order in which an application issues them.

Persistent memory is nonvolatile storage exposed through a memory-addressable interface, with durability governed by the platform's persistence contract.

Memory that switches between amorphous and crystalline states; Intel 3D XPoint marketed a variant.

Physical journaling records modified on-disk representations, such as metadata buffer contents or changed regions, for recovery.

A physical sector is the device-level data unit whose reported size describes physical read/write granularity rather than logical addressing.

Pipelined Read-Retry overlaps NAND retry sensing with transfer and decoding of earlier attempts.

In Ceph, a placement group is a logical grouping of objects within a pool that share storage placement and are managed together for replication and recovery.

Planar (2D) NAND is flash memory in which the cells are built in a single layer on the silicon surface, the layout that 3D NAND replaced as a way to raise bit density.

NAND storing five bits per cell; higher density with challenging endurance and read margins.

Point-in-time recovery restores a database to a selected historical state by applying retained transaction-log history to a suitable backup.

POSIX_FADV_DONTNEED is a file-access advice value requesting release of cached data that the application does not expect to access soon.

POSIX_FADV_NOREUSE is a file-access advice value indicating that the application does not expect to reuse the specified data.

Power-loss protection is a set of hardware and firmware measures intended to preserve specified SSD data and internal state when external power is unexpectedly removed.

Powersafe overwrite is the storage property that an interrupted file write leaves bytes outside its requested range unchanged.

Precision is the fraction of positive predictions that match observed positive outcomes at a specified decision threshold.

A prediction horizon is the future interval after a prediction time in which a specified target event is evaluated.

A Bloom filter built over extracted key prefixes rather than only whole keys, allowing negative membership tests for those represented prefixes.

A key transformation with a defined input domain that produces the prefix used for prefix-based indexing or filtering.

A recovery contract in which surviving updates form a prefix of the specified write order, including all writes covered by the last successful flush.

Partial-response maximum-likelihood (PRML) detection recovers a recorded symbol sequence by equalizing readback to a chosen partial-response target and evaluating sequence likelihoods under a channel model.

One complete program and erase of a NAND erase block, a core endurance metric.

Protection information is integrity metadata accompanying storage I/O, with checks determined by the protocol, protection format and participating layers.

Purge sanitization uses suitable logical or physical techniques to make target-data recovery infeasible even with current state-of-the-art laboratory techniques while potentially allowing the medium to be reused.

Solid-state drives built primarily from quad-level cell NAND for capacity-centric workloads.

NAND flash that stores four bits per cell using sixteen charge levels, optimizing cost per GB over endurance.

Queue depth is the number of outstanding I/O requests at a specified boundary, while a configured depth limit is the maximum or target allowed in flight.

Quorum intersection is the property that every relevant pair of allowed quorums shares at least one replica.

Research memory moving magnetic domain walls along nanowires; not a shipping bulk store.

Redundant Array of Independent Disks combines multiple drives for performance or fault tolerance.

RAID 10 combines replicated data copies with striping across storage members.

RAID 5 stripes data and distributed single parity across members so that one missing member can be reconstructed when the remaining required information is readable and consistent.

RAID 6 uses dual parity protection to retain data after any two member failures, provided the remaining required data and redundancy are readable and consistent.

A RAID rebuild reconstructs the required contents of a replacement or out-of-sync member from the array's surviving redundancy.

In The Tail at Store, a RAID tail hour is a group-hour in which at least one data drive has an hourly average I/O latency at least twice its group's median.

The RAID write hole is the risk that interrupted member updates leave inconsistent parity and corrupt reconstruction of missing data.

A range tombstone is a versioned deletion marker that suppresses covered older keys across an ordered key interval.

The rate of occurrence of failures (ROCOF) is the derivative with respect to time of the expected cumulative failure count of a repairable system.

Rateless IBLT is a set-reconciliation scheme that streams additional coded symbols until the receiver can recover the set difference without first estimating its size.

Raw bit error rate is the number of erroneous bits divided by the total bits read at a specified measurement point before error correction.

Remote direct memory access reduces CPU overhead for NVMe-oF and fast SAN traffic.

Read amplification expresses storage-read work per logical read request or per useful byte returned, under an explicitly stated accounting convention.

Read disturb is an unintended change in NAND cell thresholds caused by reads to other wordlines in the same block.

A read quorum is a protocol-approved set of replicas whose responses satisfy the requirement for completing a read.

Read reclaim relocates valid NAND data to fresh pages to limit accumulated read-disturb errors.

Read repair reconciles divergent replicas as part of processing a read request.

A storage update that reads existing data needed to preserve unchanged information, combines it with replacement data, and writes the modified data and any required redundancy.

A read-only mount restricts filesystem modification through Linux mount state, whose scope can be a VFS mount node, the filesystem superblock, or both.

Read-pass voltage is the bias applied to unselected NAND wordlines to make their cells conduct while another wordline is sensed.

Read-your-writes is a session guarantee requiring subsequent reads to use a state that incorporates the session's prior successful writes.

Recall is the fraction of observed positive outcomes identified by a classifier at a specified decision threshold.

A parity-array update method that computes new parity from incoming replacement data and unchanged data from the same stripe instead of updating old parity from old overwritten data.

The recovery point objective specifies the point in time to which data must be recoverable after a disruption, commonly expressed as the maximum tolerable data-loss interval.

A recovery target is the configured condition at which database log replay should stop.

The recovery time objective sets the tolerable duration of resource unavailability before its effect on supported processes becomes unacceptable.

Reed–Solomon codes are maximum-distance-separable error-correcting codes over finite-field symbols that add n − k check symbols to k data symbols and can rebuild missing (erased) symbols, which is why storage systems use them for RAID 6 parity and erasure-coded object storage.

Filesystem support for sharing physical extents between files until one copy is modified.

Repair locality is the number of helper fragments needed to recover a missing fragment under a specified erasure-code repair procedure.

Repair-read amplification is the ratio of bytes read during recovery to the explicitly stated quantity of lost data or lost encoded fragments being restored.

A replay attack reuses previously valid data to induce acceptance in an unauthorized context.

Replication lag is a difference in replication progress between a source and a replica, measured at a named processing stage in a stated unit.

In PostgreSQL, a replication slot tracks a consumer's progress and retains WAL and other resources needed for that consumer to continue.

A resilver is ZFS's process of bringing a newly attached, replacement or out-of-date device into agreement with the data it must hold.

A restore test exercises recovery from backup material and checks whether the recovered data or selected system functions meet defined acceptance criteria.

Retention interference is dependence of a flash cell's charge-retention behavior on the state of a neighboring cell.

Retry deduplication recognizes repeated submissions of one logical request so that an already completed mutation is not executed again.

A static approximate-membership filter built using banded Boolean linear systems, offering a space and construction-cost tradeoff relative to Bloom filters.

Right-censoring records that a time-to-event exceeds the last observed event-free time without observing the event time itself.

Mechanisms that move tape cartridges between slots and drives in a library.

RocksDB atomic flush installs the results of flushing selected column families into the database as one logical all-or-nothing MANIFEST group.

A RocksDB column family is a separately configurable logical key space within one database, with its own memtables and SST files but a shared write-ahead log.

RocksDB MANIFEST is the version-edit log mechanism used to recover the database's recorded state, with CURRENT identifying the active manifest log.

RocksDB pending compaction bytes is an estimate of rewrite work needed to bring LSM levels within their size targets under level-based compaction.

A RocksDB write slowdown is the rate-reduction state of write backpressure, distinct from a full write stop.

A RocksDB write stall is backpressure that delays or stops foreground writes when configured resource or background-work limits are reached.

RocksDB WriteBufferManager is a shared controller that accounts for memtable memory across participating column families or database instances and can trigger flushing, optional write stalls and block-cache charges.

A rollback journal is an undo log that preserves original database content so an incomplete transaction can be reversed.

Rollback protection rejects protected state older than an enforced version or freshness boundary using trusted state that cannot be rolled back with the protected object.

A rolling hash updates a fingerprint incrementally as the input window or stream advances, instead of recomputing it over the entire input each time.

Average time for the platter to rotate the requested sector under the head after a seek.

Collective term for HDDs, optical disks, and tape that depend on moving surfaces.

Rotational vibration is angular oscillatory motion that can disturb an HDD's ability to keep its heads aligned with recording tracks.

A magnetic recording code limiting consecutive same-polarity bits to increase linear density.

RWF_ATOMIC is a Linux pwritev2 flag requesting torn-write protection for the data range of a supported direct-I/O write.

RWF_DONTCACHE is a Linux per-operation flag requesting best-effort removal of page-cache content newly instantiated by a buffered read or write.

Self-Monitoring, Analysis and Reporting Technology exposes drive health metrics to the host.

Storage area network provides block access over Fibre Channel or iSCSI to multiple servers.

Sanitization validation decides whether to accept or reject a sanitization result after evaluating its evidence, suitability and remaining risk to data confidentiality.

Sanitization verification inspects the outcome of a sanitization operation to determine whether the selected technique completed successfully.

A switch-like component that fans out SAS ports so many drives share fewer host links.

SAS PHY event counters are per-physical-link-endpoint counts of specified link events, including invalid DWORDs, disparity errors, synchronization loss and PHY reset problems.

A connector standard that multiplexed SATA and PCIe lanes for faster client storage before M.2 NVMe.

SATA PHY event counters record distinct interface events in General Purpose Log 0x11 rather than measuring a disk failure rate.

Small Computer System Interface (SCSI) is a family of standards for device commands, logical units, status reporting and the transports that carry them.

CHECK CONDITION is a SCSI command status used to report a condition described by sense information.

In SCSI, the initiator is the endpoint that sends commands and the target is the endpoint that waits for them and performs the requested data transfers; the roles apply across parallel SCSI, Fibre Channel, iSCSI and other storage protocols.

SCSI persistent reservations control initiator access to a logical unit using registered keys and a selected reservation type.

A SCSI sense key is a four-bit field identifying the broad category of a condition reported in sense data.

A SCSI command that requests deallocation of specified logical block ranges.

Secure Digital cards with a second row of pins for PCIe/NVMe attachment alongside legacy SD modes.

Secure erase is a context-dependent label for a storage erasure procedure whose command, scope and assurance must be specified.

Time for the actuator to move the head stack to the correct track on a hard disk.

A self-encrypting drive encrypts stored data internally using integrated cryptographic and key-management capabilities.

A serial bus interface and protocol family for attaching HDDs and SSDs, widely used in PCs and servers.

An enterprise storage interface that extends SCSI semantics over a serial point-to-point or expander topology.

Serializability is a transaction-isolation property under which the read results and writes of committed concurrent transactions are equivalent to some serial execution of those transactions.

Session guarantees are consistency contracts for a session's operations, including read-your-writes, monotonic reads, monotonic writes and writes-follow-reads.

Set reconciliation is the task of identifying items present in one of two sets but not the other.

A shingled band is a group of overlapping magnetic tracks separated from neighboring bands by guard regions that limit overwrite propagation.

An HDD recording method that overlaps tracks like roof shingles to raise areal density at the cost of rewrite behavior.

NAND flash that stores one bit per cell, trading capacity for endurance and typically faster writes than denser cell types.

A portion of TLC/QLC SSDs programmed in pseudo-SLC mode to absorb bursty writes.

SSDs built with single-level cell NAND for high endurance industrial or cache use.

A sloppy quorum allows reachable fallback nodes outside a key's normal replica set to participate when intended replicas are unavailable.

In Perseus, the slowdown ratio is observed drive latency divided by the workload-dependent upper prediction bound for that latency.

A SMART normalized value is an ATA drive firmware's vendor-specific score for an attribute, interpreted against that attribute's threshold.

A SMART raw value is the vendor-defined attribute payload reported by an ATA drive, distinct from its normalized score.

The smartctl exit status is a process result whose normal ATA interpretation combines independent command, health-threshold and log-history flags in an eight-bit mask.

An SMB byte-range lock coordinates access to a specified region of a file through shared or exclusive locking associated with a file open.

An SMB durable handle allows a client to attempt to recover an existing file open after a network disconnect.

Server Message Block is a network protocol family used to access shared files and other server resources.

SMB FLUSH is an SMB2/SMB3 request to propagate cached information for an open file to its backing persistent store, or to wait for a named-pipe reader to consume previously written data.

An SMB lease is a server-granted arrangement that permits selected client-side read, write, or handle caching while coordinating access to a shared object.

An SMB opportunistic lock is a revocable server-granted caching arrangement that lets a client reduce network operations while coordinating access to a file.

An SMB persistent handle extends durable-open recovery to server failover on a supporting continuously available share.

SMOTE is an oversampling method that creates synthetic minority-class feature vectors by interpolating between a minority example and selected minority-class nearest neighbors.

SMR garbage collection relocates staged updates and rewrites affected shingled regions to reclaim reusable space.

An SMR persistent cache is a nonvolatile staging area that holds updates before they are incorporated into their target shingled regions.

Snapshot isolation gives a transaction a stable committed snapshot plus its own writes and prevents concurrent transactions that write the same data from both committing.

A research FTL that recovers a flush-defined disk snapshot instead of an arbitrary later write prefix.

Storage Networking Industry Association publishes education and terminology for data infrastructure.

Soft updates is a filesystem metadata-consistency mechanism that delays unsafe on-disk changes until their update dependencies are satisfied.

Soft-decision decoding uses reliability information about observed symbols rather than only their hard symbol decisions.

A solid-state drive is a storage device that retains data in nonvolatile semiconductor memory without moving recording media.

Storage without macroscopic moving parts, including NAND, 3D XPoint, and NVRAM.

An immutable table of sorted key-value entries used as a persistent component in storage engines such as LevelDB and RocksDB.

Database space amplification quantifies storage overhead relative to logical data size under an explicitly stated total-space or excess-space convention.

A file with unallocated ranges that read as zeros, allowing its logical size to exceed its allocated storage.

A SQLite super-journal is a coordination file that links recovery outcomes for a qualifying rollback-mode transaction modifying multiple attached databases on one connection.

A SQLite WAL end mark is the committed WAL boundary that determines which page versions a read transaction may observe.

SQLITE_BUSY_SNAPSHOT is SQLite's extended result code for a WAL read transaction that cannot become a writer because its snapshot is obsolete.

ASIC or SoC managing NAND channels, ECC, and host protocols on an SSD.

SSD DRAM is volatile working memory that can hold address mappings, controller metadata or buffered user data, depending on the design.

An SSD mapping cache retains selected logical-to-physical address mappings in memory.

Preparatory writing used to establish a specified SSD state before performance measurement.

A workload-specific SSD performance condition in which a named metric satisfies stated stability criteria over a defined measurement window.

SSD thermal throttling is a controller response that limits activity or performance to manage device temperature.

A hard drive with a small NAND cache to accelerate frequently read data while retaining rotating media.

Seagate's follow-on full-height 5.25-inch HDD interface companion to the ST-506 ecosystem.

An early HDD interface where the drive exposed raw heads and the controller lived on a host adapter card.

Staggered scrubbing is a scan-order strategy that visits segments across multiple disk regions before returning to subsequent segments in each region.

STATX_WRITE_ATOMIC is a Linux statx request for a file's reported atomic-write size limits and maximum vector count.

A storage cache flush requests persistence of writes buffered in the addressed storage stack, with its coverage and ordering defined by that interface.

Storage immutability prevents specified data from being overwritten or deleted under an enforced retention rule or hold.

Storage latency is the elapsed time between specified start and completion events for an I/O operation, including any delays inside those measurement boundaries.

A storage snapshot is a point-in-time representation of a filesystem or volume, whose implementation may share blocks with its source.

Storage throughput is the amount of data transferred per unit time at a specified system boundary, usually expressed in bytes per second.

A layer abstracts multiple backends into pooled logical volumes presented to hosts.

Media with latency between DRAM and NAND, often byte-addressable, used for persistence tiers.

Splitting sequential data across drives to aggregate bandwidth.

Submission latency in fio is the elapsed time from initialization of an I/O unit to an engine-dependent submission boundary.

An approximate point-and-range membership structure based on a truncated order-preserving Fast Succinct Trie.

The grain-size floor where longitudinal bits become thermally unstable, motivating PMR and energy-assisted recording.

The survival function S(t) = P(T > t) is the probability that a time T to a defined event exceeds t, measured from a specified time origin.

A symbol erasure is a codeword position whose value is unavailable or deliberately discarded and whose location is known to the decoder.

sync_file_range is a Linux-specific operation that initiates or waits for page-cache writeback over a file range without synchronizing file metadata or flushing volatile device write caches.

syncfs is a Linux system call that synchronizes pending file data and metadata for the filesystem selected by an open file descriptor.

Synchronous replication makes successful completion depend on acknowledgement that a specified set of replicas has reached the processing stage required by the replication contract.

A synthetic full backup reconstructs a complete backup for a selected recovery point from retained full and incremental backup data without rereading the entire protected dataset from production.

A systematic code includes the original data symbols unchanged among its encoded symbols.

Tail latency describes the high-latency part of an operation response-time distribution, usually summarized using specified high percentiles.

Encrypting data as it is written to tape cartridges to protect offline media.

Automated systems inserting LTO or enterprise tape cartridges into drives.

TapeAlert is a SCSI status-reporting mechanism for tape devices that exposes condition flags through LOG SENSE page 2Eh.

Terabytes Written rating estimates total host writes an SSD warranty covers before wear exhaustion.

A Trusted Computing Group spec for managing hardware encryption on self-encrypting storage devices.

Temporal validation evaluates a predictor on later observations using training information available before the evaluation boundary.

Thermal flying-height control uses an embedded heater to adjust HDD head-disk clearance through local thermal protrusion.

Presenting logical storage capacity while allocating physical backing from a shared pool as writes require it.

PCIe tunneling over Thunderbolt enables external NVMe enclosures at near-internal link speeds.

Tiered compaction allows multiple sorted runs to accumulate and merges groups of runs into typically larger runs instead of maintaining one run at every tier.

A proposed patterning approach grouping grains into magnetic tiles for higher density.

SSDs using triple-level cell NAND as the mainstream balance of cost and endurance.

Tunnel magnetoresistance sensors succeeded GMR heads for higher signal on dense tracks.

A torn write is an interrupted write that persists only part of its intended update, potentially leaving a database page with mixed old and new contents.

An HDD touchdown sensor is a temperature-sensitive element used to observe thermal changes near the head-disk interface during clearance and contact studies.

The ATA DATA SET MANAGEMENT Trim function tells a device that specified logical block ranges are no longer needed.

NAND flash that stores three bits per cell using eight charge levels, common in consumer and many enterprise SSDs.

Enterprise SSD bay and connector family derived from SFF-8639, supporting NVMe and legacy SAS/SATA pinouts in U.3.

USB Attached SCSI Protocol (UASP) transports SCSI commands over USB and supports multiple outstanding commands, without by itself guaranteeing throughput, SMART pass-through or durable writes.

Ultra Density Optical was a 30 GB-class professional WORM optical cartridge format.

Uncorrectable bit error rate is a storage reliability metric that normalizes unrecovered data-error outcomes by a specified bit count after the stated correction and recovery procedures.

Unlink is the removal of a directory entry that names a file, without invalidating the file's other hard links or existing open references.

The SCSI UNMAP command requests deallocation of logical blocks.

An unserializable write is a storage failure in which recovered contents cannot be reconciled with an operation order allowed by the known write history.

An allocated filesystem extent whose data is not initialized and is exposed as zeros when read through the filesystem.

In Ceph, the up set is the ordered OSD list assigned to a placement group by CRUSH for a particular map epoch.

The USB device class that presents block storage to hosts using BOT or UASP transports.

Samsung's brand for its charge-trap 3D NAND vertical stacking architecture.

A value log is an append-oriented record store whose values are located through a separate index.

Host or array software that groups physical drives into logical volumes and snapshots.

WAL archiving preserves write-ahead-log segments for later recovery before normal log recycling can remove them.

Capacity-oriented online storage balancing cost and occasional access latency.

A flash management technique that spreads erase cycles across blocks so no single region wears out prematurely.

Winchester disk architecture combines low-mass, lightly loaded flying heads with lubricated disks in a protected head-disk assembly; it does not require non-removable media.

Write-once-read-many storage preserves recorded content against subsequent alteration while allowing repeated reads.

Device-level byte write amplification is the ratio of bytes programmed to the storage medium to bytes written by the host over the same accounting interval.

Merging adjacent small writes into fewer large operations to reduce amplification on flash or SMR.

A write quorum is a protocol-approved set of replicas whose acknowledgements satisfy the requirement for completing a write.

Write skew is a concurrency anomaly in which transactions read overlapping data and make disjoint writes whose combined outcome can violate an invariant that each transaction would preserve alone.

A write-ahead log is a recovery log whose relevant records must become durable before the corresponding changes are persisted in database data files.

Write-back caching completes a write after updating the cache and defers updating its backing store until later.

A write-intent bitmap tracks regions of a redundant array that may need synchronization after interrupted writes.

Storage that physically or logically prevents rewriting committed regions.

Write-through caching propagates writes to the backing store as part of completing them at the cache layer.

A writeback error is a failure encountered while transferring modified cached data to its backing storage.

Writes-follow-reads requires every replica incorporating a session write to incorporate and order before it the updates relevant to that session's earlier reads.

The XFS Active Item List tracks logged items whose journal records remain needed until later retirement.

The XFS Committed Item List tracks in-memory logged changes awaiting a journal checkpoint.

An XFS deferred operation postpones metadata work to a later transaction to satisfy locking or reservation constraints.

XFS delayed logging aggregates metadata changes in memory before committing them to the filesystem journal.

An XFS intent log item describes deferred work that log recovery may need to complete.

XFS relogging records an object's outstanding metadata changes again in a newer transaction.

An XFS rolling transaction commits one step of a transaction chain and continues with a fresh transaction using a permanent log reservation.

XTS-AES is a tweakable AES encryption mode for storage data units that provides confidentiality without authenticating the data or its source.

Zoned ATA Commands are the SATA counterpart to ZBC for host-managed SMR HDDs.

Zoned Block Commands standardize host-managed SMR access for SCSI/SATA devices.

A ZFS error scrub is an integrity scan restricted to files already recorded with data errors.

The ZFS Intent Log is a per-dataset recovery log that can persist filesystem operations before their transaction group reaches stable storage.

A ZFS separate log is dedicated pool log-device storage used for intent-log blocks instead of their default allocation in the main pool.

A ZFS transaction group is an ordered batch of changes whose synchronization advances the pool's consistent persistent state.

A removable cartridge drive format from Iomega offering tens to hundreds of megabytes per disk in the 1990s.

Zone Append writes at a controller-selected position within a specified zone and returns the assigned starting logical block address on successful completion.

Zone capacity is the maximum number of logical blocks that may be written within a zone.

Zone Finish changes an eligible sequential zone to Full, ending further writes until reset without discarding the data already written.

Zone Reset makes an eligible sequential zone empty for reuse, returns its write pointer to the zone start and invalidates access to its previous data.

A zone write pointer identifies the next writable logical block address in a sequential zone when its current state makes that pointer valid.

A zoned block device exposes a logical block address space divided into zones with zone-specific access rules.

A zoned namespace is an NVMe namespace whose logical address space follows the ZNS command set's zone organization and write constraints.

Zoned Namespaces (ZNS) is an NVMe command set that organizes a namespace into zones with sequential-write constraints and explicit reset before reuse.

Bairavasundaram et al., FAST 2008, report that 3,088 of about 358,000 nearline disks (0.86%) and 767 of 1.17 million enterprise-class disks (0.065%) developed at least one silent checksum mismatch in 41 months of support logs covering 1.53 million disks, about 400,000 mismatched 4 KB blocks in all. The median affected disk had 3 mismatches, the mean 104, the maximum 33,000. Only detected corruption is counted. A CERN IT report from 2007 gives an independent but different measure: 22 of 33,700 files failed a checksum comparison.

Gunawi et al., SoCC 2016, Table 3, mark 355 of 597 unplanned outages as UNKNOWN. The study covers 32 popular Internet services and 1,247 public reports from January 2009 through December 2015. The paper states that only 40% of those descriptions reveal a root cause and only 24% reveal a fix. Among known root-cause tags, the three largest counts are UPGRADE 54 (16%), NETWORK 52 (15%), and BUGS 51 (15%). STORAGE is 13 occurrences on 4 services (4% of known causes). HARDWARE is 5 occurrences on 4 services (1% of known causes). One outage can carry more than one tag, so the occurrence column is not a partition of the 597 outages. In the yearly column, M means at least 10, not exactly 10.

Narayanan et al., SYSTOR 2016, studied over half a million SSDs in one operator's datacenters over nearly 3 years. 62% of failed devices had shown at least one of four SMART symptoms and 38% none; devices with a symptom failed 3 to 20 times as often. A classifier reached 87% precision and 71% recall with 5-fold cross-validation. Cross-check rows come from NetApp (FAST 2020) and Backblaze (2016).

An example BCH code corrects raw bit error rates up to 1.0e-3 and soft LDPC up to 5.0e-3 (Cai et al., 2017; Zhao et al., FAST 2013 find about 3 times). Google's field median raw bit error rate was 5.8e-10 to over 3e-8, and 2 to 6 per 1,000 drive days had an uncorrectable error (Schroeder et al., FAST 2016).

At 128K and 100 percent sequential read, one three-node VMware vSAN cluster reached 10,366 MB/s on KIOXIA CD6 NVMe, 4,566 MB/s on 12 Gbps SAS, and 2,776 MB/s on 6 Gbps SATA. Principled Technologies reported those cluster totals at each configuration's own 100 percent Vdbench IOrate. Hands-on testing ended 21 May 2021, the configurations were finalized on 14 April 2021 or earlier, and the KIOXIA-commissioned report is dated September 2021. The same rows show latency of 13.88 ms on NVMe and 3.51 ms on SATA, so the higher throughput is not a lower-latency result. The 6 Gbps and 12 Gbps figures are the printed link rates and are not converted to MB/s.

Xu et al., USENIX ATC 2018, built a disk-error predictor for Azure. At a 0.1% false-positive rate it caught 36.50% of faulty disks on one test set (29.67% to 41.09% across three), against 15.51% to 30.51% for SVM and random forest baselines on SMART data; a random split gave 91.64% on the same data. Rows from Backblaze (2016) show the SMART gap in another fleet.

Annualized failure rate by drive power-on age, computed from Backblaze's raw daily Drive Stats files for Q4 2025 (October 1 to December 31, 2025) with the exposure counted per age bucket. Sample: hard drives of 3 TB or more with a power-on-hours value, 30,575,671 drive days and 926 failures (1.11% pooled AFR). The rate is 0.36% at 1 to under 2 years, 2.59% at 6 to under 7 and 7.08% at 7 to under 8, from 44 failures in 0.7% of drive days. Association only, one fleet and one quarter, with model mix that changes by age.

Side-by-side published annualized failure rates for 33 hard-drive models in Backblaze's Q1 2026 (January 1 to March 31) and Q2 2026 (April 1 to June 30) Drive Stats model tables. 31 models appear in both. Fleet AFR in the published totals rows rose from 1.24% (1,030 failures, 30,203,180 drive days) to 1.73% (1,498 failures, 31,553,350 drive days). Vendor claim; Backblaze's own fleet, not an industry sample.

Yearly annualized failure rate of Backblaze's hard drives by nominal capacity group, computed from the raw daily Drive Stats files for 2022 to 2025 (sixteen quarterly zips) with the drive-day exposure of each group and exact 95% Poisson intervals. Sample: drives of 3 TB or more, 388,670,883 drive days and 16,051 failures over the four years. In 2025 the 20 to 22 TB group was at 0.51% on 16.0% of drive days, 16 to 18 TB at 1.03%, 12 to 14 TB at 1.84% and 8 to 10 TB at 2.19%. One fleet, and capacity groups mix models and ages, so the file does not show an effect of size.

Yearly annualized failure rate of Backblaze's hard drives for Seagate, HGST, Toshiba and WDC, computed from the raw daily Drive Stats files for 2022 to 2025 (sixteen quarterly zips) with exact 95% Poisson intervals. Sample: drives of 3 TB or more, 388,670,883 drive days and 16,051 failures over the four years. Pooled AFR was 1.39% in 2022, 1.70% in 2023, 1.59% in 2024 and 1.36% in 2025. HGST went from 0.63% in 2022 to 3.01% in 2024 and 2.89% in 2025. One fleet, and the file does not say why rates moved.

Month-by-month annualized failure rate of Backblaze's hard drives in 2025 (January 1 to December 31), computed from the raw daily Drive Stats files with exact 95% Poisson intervals, for all drives of 3 TB or more and for three capacity classes. Sample: 115,936,784 drive days and 4,327 failures, a 1.36% pooled AFR that matches Backblaze's published annual figure. Monthly AFR runs from 0.54% in December to 1.72% in September. One fleet, one year, and the file does not say why months differ.

Which hard-drive models Backblaze added or retired between January 1, 2022 and December 31, 2025, from the raw daily Drive Stats files (sixteen quarterly zips). Sample: all 66 model strings of drives of 3 TB or more, 388,670,883 drive days and 16,051 failures, with drives on December 31 of each year, exposure, failures and the 2022 to 2025 AFR with exact 95% Poisson intervals per model. Models added after 2022 Q1 held 79,349 of the 334,844 drives on December 31, 2025. One fleet, and model strings are used as they appear in the files.

Yearly drive days, failures and annualized failure rate with exact 95% Poisson intervals for the 22 Backblaze hard-drive models that had at least 100,000 drive days in every year from 2022 to 2025, computed from the raw daily Drive Stats files (sixteen quarterly zips). Sample: 88 model-years. Some models stay flat, such as the Toshiba MG07ACA14TA at 1.01% to 1.15%, while others move a lot: HGST HUH721212ALN604 went from 0.74% in 2022 to 5.64% in 2024, and Toshiba MG08ACA16TEY from 1.58% in 2024 to 6.30% in 2025. One fleet, and the file does not show why a rate moved.

Published Q1 and Q2 2026 Backblaze AFRs for 33 hard-drive models, each with an exact 95% Poisson interval computed from the published failures and drive days. Sample: Q1 2026 (January 1 to March 31) 30,203,180 drive days and 1,030 failures; Q2 2026 (April 1 to June 30) 31,553,350 drive days and 1,498 failures. Of 31 models in both quarters, the intervals do not overlap for 4 and overlap for 27. The Poisson assumption is a modelling choice made here, not something Backblaze publishes.

Pooled annualized failure rates for Backblaze's published Q1 2026 (January 1 to March 31) and Q2 2026 (April 1 to June 30) model tables, grouped by manufacturer (4 groups) and nominal capacity (10 groups). Pooled over 33 model rows: 30,203,180 drive days and 1,030 failures in Q1, 31,553,350 drive days and 1,498 failures in Q2. Pooled AFR rose in all four manufacturer groups and in 8 of 9 capacity groups present in both quarters. Vendor claim data from Backblaze's own fleet; the grouping and pooling are computed here.

Backblaze's published Q2 2026 quarterly hard-drive annualized failure rate was 1.73% (vendor claim): 1,498 failures across 31,553,350 drive days on 354,415 drives. This download is the 31-model table for April 1 through June 30, 2026 inclusive (drive count above 100 and drive days above 10,000), not the daily SMART archive.

Exposure and failures for the SSD models in Backblaze's raw Drive Stats files for 2025 (January 1 to December 31, 2025). Sample: 20 SSD models, 1,075,258 drive days, 2,972 drives on December 31, and zero failures. With no failures the file gives each model's exact 95% upper bound on the annualized failure rate: 0.13% for all SSD models pooled, and 0.34% to 4.30% for the models with 10,000 or more drive days. One fleet, mostly boot drives, so it does not describe SSDs in general.

Fourteen deterministic cases compare equal bits per entry with optimized allocation across three invented runs. Expected probes, not measured latency.

Fourteen reported-value rows from three historical studies, with distinct counting units and source qualifiers. The separate synthetic burst example illustrates adjacent-gap grouping, not an empirical failure rate.

Six synthetic cold-cache LRU scenarios isolate repeated container reads from unused fetched bytes. No device timings or real backup records.

Lu et al. (FAST 2020) and two SMART-warning sources: 13 reported figures with units, bounds and validation limits. A source extract, not a model reproduction.

An 18-row extraction of reported replacement rates and fleet event gaps, with source qualifiers and population limits; not a disk-lifetime dataset.

Historical source figures and reproducible fixed-window arithmetic: distinguish Poisson event counts from distinct-shard binomial probabilities, with explicit storage-class scope.

Maneas et al., FAST 2020, mined NetApp support messages for about 1.4 million SSDs (3 manufacturers, 18 models, January 2017 to May 2019). The average annual replacement rate was 0.22%, from 0.07% to nearly 1.2% by model. Early firmware versions of some families were replaced 2 to more than 10 times as often as later ones, and the chance of another replacement within a week in a RAID group was 9.39% against 0.0504% in a random week. Cross-check rows come from Google (FAST 2016).

Ford et al., OSDI 2010, studied tens of Google storage cells over one year. Node MTTF was 4.3 months against 10 to 50 years for a disk, and 37% of failures were part of a burst of at least 2 nodes. Cross-check rows come from NetApp (FAST 2020) and HPC1 (FAST 2007) and show clustered drive replacements.

Maximum and average rotational latency of a hard disk drive for six spindle speeds from 3,600 to 15,000 rpm, computed as 60 / rpm and half of that, and checked against the values printed in the cited Wikipedia article. At 7,200 rpm one revolution takes 8.33 ms and the average wait is 4.17 ms, and at 15,000 rpm they are 4.00 ms and 2.00 ms. Arithmetic from a definition, no measurements and no seek times.

Han et al., FAST 2021, studied nearly one million Alibaba SSDs over two years. 12.9% of failures were intra-node and 18.3% intra-rack within 30 minutes, and the strongest SMART attribute had a rank correlation of only 0.23 with them. Cross-check rows come from NetApp (FAST 2020) and Google (OSDI 2010).

Talagala and Patterson (UC Berkeley, 1999) logged 18 months of a prototype with 368 SCSI data disks. 7 data disks failed (1.9%), against 13 of 46 enclosures, and disks were about 90% of parts but around 4% of errors. Cross-check rows come from Schroeder and Gibson (FAST 2007) and Ford et al. (OSDI 2010).

How IOPS, bandwidth, queue depth and mean latency relate, as a worked table of 100 rows: four block sizes (4 KiB to 1 MiB), queue depths of 1 to 128 and mean latencies of 100 microseconds to 10 milliseconds. The rules are IOPS = queue depth / mean latency (Little's law) and bandwidth = IOPS x block size, in decimal MB/s and in MiB/s as fio prints them. Pure arithmetic, no measurement. For example, queue depth 1 at 100 microseconds is 10,000 IOPS and 40.96 MB/s at 4 KiB, and queue depth 32 at 1 millisecond is 32,000 IOPS.

Thirty deterministic byte-accounting cases compare inline records with separated values, including relocation-driven index writes. Invented inputs, not engine measurements.

Linux switching-sched names 4 per-device blk-mq schedulers: mq-deadline, none, bfq, and kyber. I/O priority is named for bfq and mq-deadline, with 3 classes, 8 RT class-data levels, and best-effort levels 0 (highest) through 7 (lowest); none and kyber are not named. Kyber has 2 tunables, read and synchronous-write targets in nanoseconds, which are throttle targets rather than a measured latency. The none scheduler does not reorder, and completion order is not guaranteed. No throughput was measured.

In Jiang et al., FAST 2008, Table 1, low-end NetApp systems recorded 4,338 physical-interconnect failure events and 3,230 disk failure events from January 2004 through August 2007 (22,031 systems and 264,983 FC disks ever installed). For Figure 4(b), which excludes subsystems that used Disk H, that class has a stated disk AFR of 0.9% and a storage-subsystem AFR of 4.6%. In Maneas et al., FAST 2020, Table 2, SCSI error is 32.78% of SSD replacements (ARR 0.055%) on almost 1.4 million NetApp SSDs; that signal is a hardware error the SSD reports (the example is drive DRAM ECC), while aborted commands are 13.56% (ARR 0.023%) and include a host or connection case where write data never reaches the device.

Five deterministic whole-fragment examples distinguish data repair, all-fragment averages and parity-inclusive accounting. Not cluster measurements.

Native and advertised compressed capacity, release year and maximum native throughput of LTO Ultrium generations 1 to 10 (11 rows, with the 30 TB and 40 TB LTO-10 cartridges), taken from the LTO Program pages for generations 8, 9 and 10 and the Wikipedia generations table for the rest, plus arithmetic: the hours to write one full cartridge at the maximum native rate and the capacity step from one generation to the next. Native capacity goes from 0.1 TB in LTO-1 to 40 TB in LTO-10, and the 40 TB cartridge takes 27.78 hours to fill at 400 MB/s. Two published speed figures do not match their own arithmetic and are flagged.

Selected chip parameters and simulation results from 2008-2018, with explicit source scope. These are not current SSD specifications or a controlled comparison of NAND generations.

Over 56% of failed drives in the Google fleet studied by Pinheiro, Weber, and Barroso (FAST '07; more than 100,000 consumer PATA and SATA drives; SMART counts from December 2005 through August 2006) had no count on scan errors, reallocation count, offline reallocation, or probational count. Adding every other SMART parameter except temperature still left over 36% of failed drives at zero on all of those variables. That 56% figure is an upper bound on coverage for the four signals, not a false-positive-aware accuracy.

Twenty-one exact survivor rows compare fixed and balanced width-four groups with a width-eight control. Synthetic accounting, not storage measurements.

Theoretical one-direction payload ceilings for 14 storage-link generations: SATA 1.0 to 3.0, SAS-1 to SAS-4 (one lane), and PCIe 1.0 to 7.0 at four lanes. Each value is the published line rate times the line-code efficiency, and nothing is measured. SATA 3.0 gives 600 MB/s, SAS-4 gives 2,400 MB/s, PCIe 4.0 x4 gives 7,877 MB/s and PCIe 7.0 x4 gives 60,500 MB/s. Real drives and protocols stay below these ceilings.

Reported Perseus drive counts and benchmark scores, with population limits and a separate comparison to operator incident reports.

Per-model power-on age of the hard drives that failed in Backblaze's Q4 2025 (October 1 to December 31, 2025) against the age of the fleet on December 31, computed from Backblaze's raw daily Drive Stats files. Sample: 34 hard-drive models, 334,392 drives on December 31, 30,534,025 drive days, 943 failures (1.13% pooled AFR), of which 924 carry a power-on-hours value. Pooled medians are 4.65 years for failed drives and 3.38 years for the fleet, but within single models the two medians are close.

56 deterministic RAID5/P+Q examples compare read-modify-write and reconstruct-write with explicit byte denominators. No device measurements or timing predictions.

32 synthetic states distinguish inconsistent parity, recovery of an untouched missing chunk and preservation of an interrupted update. Not hardware measurements or failure probabilities.

A finite example contrasts a narrow blocking repair, response-only reconciliation and a full-partition read, without running a Cassandra cluster.

Four constructed counter-delta cases preserve zero-host-write intervals, distinguish ratios from rates, and show why sequential windows must be combined before division.

Ma et al., FAST 2015, analysed about 1 million SATA disks in 6 models in one vendor's backup systems (logs of 21 to 60 months from June 2008). Replacing a disk once its reallocated-sector count passed 200 avoided about 88% of triple-disk RAID failures, which were 80% of RAID failures, or about 70% of all disk-related incidents. In simulation, thresholds below 200 caught 52% to 70% of failures within 60 days with 0.8% to 4.5% false positives. Cross-check rows come from Google (FAST 2007) and Backblaze (2016).

Selected published figures for drive-week prediction, simulated detection time and two other fleets; these are different endpoints, not a pooled estimate.

A slow-drive hour, a slow-RAID hour and a request percentile count different events. This source review separates them and shows why fan-out estimates need explicit dependence and workload assumptions.

Failure-day SMART association in Backblaze Q4 2025: 690 of 926 failed drives reporting at least one selected attribute have a positive raw value (74.5%), versus 8.1% in the December 31 nonfailed snapshot. Of 236 failed rows with reported values but no positive, 66 have all five zero and 170 have partial coverage; another 19 have none. Capacity filter: at least 3 TB. Not a prediction test.

Botezatu et al. report 98% Seagate recall; their Hitachi table gives 79%, while the prose says 81%. These random-split results are not a prospective deployment estimate.

Twelve constructed flat-map cases distinguish TB from TiB, mapping granularity from host sectors, and cache address coverage from cached user data.

Seven computed reserve-to-write-amplification cases with explicit logical and physical capacity denominators. Analytical model evaluation, not SSD measurements.

Six synthetic five-round IOPS examples separate variation, drift and local stability. These are arithmetic counterexamples, not SSD benchmark measurements.

390 synthetic request records compare independent and completion-driven arrivals during one pause; downloadable timestamps reproduce the p99 calculation.

How to convert between TBW (terabytes written) and DWPD (drive writes per day) for an SSD, with a worked table of 60 rows: six capacities from 0.25 to 8 TB, warranty periods of 3 and 5 years, and DWPD ratings of 0.1, 0.3, 1, 3 and 10. The formulas are TBW = DWPD x capacity x 365 x years and DWPD = TBW / (capacity x 365 x years). Pure arithmetic from the public definitions, with no measured data and no product ratings. For example, 1 TB at 1 DWPD over 5 years is 1,825 TBW, and 600 TBW on a 1 TB drive over 5 years is 0.33 DWPD.

Splits the older power-on age buckets (6 to under 7, 7 to under 8, 8 or more years) of Backblaze's Q4 2025 fleet (October 1 to December 31, 2025) by drive model, from the raw daily Drive Stats files. Sample: hard drives of 3 TB or more with a power-on-hours value; 1,424,836, 226,691 and 1,383,760 drive days in the three buckets, with 101, 44 and 70 failures. In the 7 to under 8 year bucket (7.08% AFR), two models hold 27 plus 13 of the 44 failures: HGST HUH728080ALE600 and Toshiba MG07ACA14TA. Association only, one fleet and one quarter.