{
  "id": "data-backup",
  "title": "Data Backup",
  "lang": "en",
  "source_lang": "en",
  "summary": "Data backup creates recoverable copies of digital information to protect against deletion, corruption, equipment failure, and other losses.",
  "category": "computer-science",
  "subcategory": "software",
  "type": "concept",
  "aliases": [
    "Data back-up",
    "computer backup"
  ],
  "generated_at": "2026-11-02",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/data-backup/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/data-backup/",
      "json": "https://aiwiki.page/zh/data-backup.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Data backup is the creation and maintenance of copies of digital information that can be used to restore an earlier state after loss, damage, or unwanted change. A backup may contain individual files, application records, an entire system image, or a coordinated collection of data from several systems. Its defining property is recoverability: the stored information, together with the necessary metadata, software, credentials, and procedures, must support a usable restoration.\n\nBackup is a central component of information security and disaster recovery. It addresses failures ranging from accidental file deletion to storage-device failure, software defects, malicious destruction, and site-wide disruption. Backup engineering therefore concerns more than copying bytes. It includes choosing recovery points, preserving application consistency, separating copies from common hazards, controlling retention, and demonstrating that restoration works within operational constraints."
    },
    {
      "heading": "Scope, boundaries, and historical development",
      "text": "A backup preserves a recoverable state independently enough from the working data to survive specified failures. Independence is relative to the threat being considered. A second directory on the same disk may protect against an editing mistake but not disk failure; a second disk in the same building may survive a device failure but not a building-wide incident. Consequently, the number of copies alone does not describe protection. Their physical location, administrative control, accessibility, and dependence on shared components are equally important.\n\nBackup differs from several related techniques. Redundant arrays of independent disks, commonly called RAID, maintain availability despite some drive failures, but usually reproduce accidental deletion or malicious overwriting immediately. Data replication maintains copies across systems, often to support availability and read performance; without retained history, it can propagate corruption as effectively as valid changes. Synchronization makes selected locations converge, whereas backup intentionally preserves states that no longer exist in the source. These technologies can form parts of a protection system without individually constituting a complete backup arrangement.\n\nAn archive has a different primary purpose: long-term retention and retrieval of selected information, often after it is removed from active use. Backups are usually organized around restoring a system or a recent historical state; archives are commonly organized around enduring records and their interpretation. The distinction is functional rather than absolute. A retained backup may satisfy some archival needs, but it can be difficult to search, depend on obsolete software, or include many redundant versions that have no continuing records value.\n\nBackup practices developed alongside storage technology. Early computing installations used removable media, including punched cards and magnetic tape, to preserve programs and data. Mainframe operations established scheduled processing, tape rotation, and off-site storage. The spread of the personal computer brought floppy-disk and later external-drive backups, while networked systems encouraged centralized backup servers and automated libraries. From the 2000s onward, disk-based repositories, virtualization, and cloud computing expanded the range of destinations and recovery methods. Newer systems increasingly track changed blocks, retain snapshots, and integrate protection into storage platforms, but still depend on the older principle of preserving a usable state outside the immediate failure path."
    },
    {
      "heading": "Recovery objectives and protection design",
      "text": "Backup design begins with the consequences of losing data and the time required to recover service. The recovery point objective (RPO) specifies the maximum acceptable age of the state to which a system can be restored, expressed as a time interval before disruption. An RPO of one hour means that the design seeks to limit missing changes to approximately that interval. It does not mean that backups must merely start every hour: failed jobs, delayed transfers, and incomplete replication can leave the latest usable recovery point substantially older.\n\nThe recovery time objective (RTO) specifies the target duration for restoring a required service after disruption. RTO includes more than reading stored data. Incident assessment, infrastructure provisioning, credential recovery, application initialization, integrity checks, and traffic redirection may all lie on the recovery path. A repository with high advertised transfer speed may still support slow recovery if millions of small files must be reconstructed or if data must first be retrieved from an offline storage tier.\n\nRPO and RTO are related but independent. Frequent log capture may provide a recent recovery point while rebuilding a large database still takes hours. Conversely, a preconfigured standby system may resume service quickly but contain an older dataset. Protection design therefore identifies which business functions require rapid restoration, which data can tolerate reconstruction, and which dependencies must recover together. An order-processing application may depend on its database, identity service, message queues, configuration, and external connectivity; restoring only its main server does not necessarily restore the service.\n\nCapacity and bandwidth place physical limits on these objectives. For a data volume \\(D\\) and effective sustained transfer rate \\(B\\), the transfer component of recovery has a lower bound of approximately\n\n\\[T_{\\mathrm{transfer}} \\geq \\frac{D}{B}.\\]\n\nRestoring 10 decimal terabytes over an ideal 1-gigabit-per-second connection would require about 22.2 hours for transfer alone, before protocol overhead, contention, or processing. Real designs also account for backup windows, peak change rates, concurrent workloads, and growth. The relevant measure is end-to-end recovery under realistic conditions, not the speed of one device or network link in isolation."
    },
    {
      "heading": "Full, incremental, differential, and continuous methods",
      "text": "A full backup captures all selected data for a recovery point. “Full” refers to the protected scope rather than necessarily the entire machine: a full backup may cover one directory, one database, or one virtual disk. Full backups simplify some restoration paths because they provide a self-contained baseline, but reading and transferring the entire dataset can impose substantial load. A new full backup also need not occupy the full logical dataset size if the repository shares unchanged blocks with existing copies.\n\nAn incremental backup captures changes since a preceding backup or another defined checkpoint. A conventional restore may require an initial full backup followed by every necessary incremental set in sequence. For example, if a full backup is taken on Sunday and daily incrementals thereafter, restoring Thursday’s state may require Sunday’s full copy and the Monday-through-Thursday increments. The dependency chain reduces repeated copying but increases the importance of catalog integrity and the availability of intermediate data. Modern systems can use indexes and shared-block maps so that users do not manually apply each increment.\n\nA differential backup records changes since the most recent full backup. In the same weekly schedule, Thursday’s differential includes all qualifying changes since Sunday, so restoration ordinarily requires the full backup and Thursday’s differential rather than each intervening set. Differential backups tend to grow as the baseline ages. Incremental and differential describe reference points, not a universal implementation: products can detect changes at file, block, or application-record level, and may treat deletion, renaming, and metadata changes differently.\n\nSynthetic full backups construct a new full recovery point from an existing baseline and later changes, mainly within the backup repository. This avoids rereading the entire source but shifts work to the storage system. “Incremental forever” designs typically take an initial baseline and subsequently preserve changes while maintaining metadata that presents complete historical views. Their apparent simplicity depends on reliable indexing, dependency management, and reclamation of data no longer referenced by retained recovery points.\n\nContinuous data protection records changes continuously or at very short intervals, often in a journal that supports selection of recovery times. It differs from merely scheduling frequent backups, although product terminology varies. The journal may allow recovery to a point just before an unwanted operation, but its value depends on journal retention, ordering, and application consistency. A corrupted write can be faithfully captured; continuous capture preserves change history rather than deciding which historical state is correct."
    },
    {
      "heading": "Consistent capture of files, databases, and distributed applications",
      "text": "Copying data while an application is modifying it creates a consistency problem. Different files or blocks may be read at different times, yielding a combination that never existed as a valid application state. A database’s data pages might be copied after a transaction while its associated control information is copied before it. Even if every byte is readable, the result may not be recoverable. This distinguishes physical readability from logical consistency and explains why copying open files is not automatically a sound application backup.\n\nA snapshot provides a point-in-time view of a storage object or filesystem. Copy-on-write implementations preserve older blocks when changes occur, while redirect-on-write implementations place new data elsewhere and retain references to the previous state. A snapshot can make the backup input stable without requiring the application to remain stopped throughout the copy. However, a snapshot retained only on the original storage system may share its failure risks. Snapshot creation and independent backup transfer are separate operations, even when a product combines them.\n\nCrash-consistent capture aims to preserve a state comparable to sudden power loss. Applications with robust recovery mechanisms may reconstruct valid state from it. Application-consistent capture goes further by coordinating with the application, for example by flushing buffers, establishing a checkpoint, or temporarily quiescing writes. Operating systems and application frameworks provide mechanisms for this coordination; Microsoft’s Volume Shadow Copy Service is one example. The degree of consistency depends on participating components, not simply on the presence of a snapshot feature.\n\nA database often uses write-ahead logging, in which recovery information is made durable before the corresponding data-page changes. A supported online backup process preserves a base copy and enough log records to make it consistent. Point-in-time recovery then replays changes up to a selected transaction boundary or time, subject to the database’s semantics. A missing segment in the required log sequence can prevent recovery beyond that gap, even if later log files remain available.\n\nApplications built using distributed computing introduce coordination across independent components. Capturing each server at the same wall-clock time does not necessarily produce a globally consistent state, because messages may be in transit and transactions may cross system boundaries. Backup procedures may use coordinated barriers, application-level export facilities, transaction identifiers, or replayable event logs. The correct consistency boundary follows the application’s invariants: inventory, payment, and order records may need reconciliation even when each underlying database individually recovers without error."
    },
    {
      "heading": "Storage media, repositories, and geographical separation",
      "text": "Backup destinations include magnetic tape, hard disks, solid-state devices, removable media, remote servers, and cloud storage services. Their characteristics differ in access latency, throughput, durability mechanisms, power consumption, portability, and cost. Disk repositories provide convenient random access and support rapid incremental ingestion. Solid-state storage can accelerate metadata operations and concurrent recovery, but its performance advantage does not eliminate bottlenecks elsewhere in the restoration path.\n\nMagnetic tape remains significant for large sequential datasets and copies that can be physically removed from network access. Robotic libraries automate cartridge selection, while off-site vaulting separates stored copies from the production location. Linear Tape-Open technology, introduced commercially in 2000, provides an example of a tape format family used for backup and archival storage. Tape retrieval involves mounting and positioning media, making access latency quite different from disk access. Successful long-term use also depends on compatible drives, readable catalogs, appropriate handling, and migration between technology generations.\n\nCloud repositories commonly use object storage, where data is stored as objects addressed through service interfaces rather than ordinary local disk blocks. They can provide geographic separation and avoid operating a local tape library or large disk array. Nevertheless, storage durability is not equivalent to application recoverability. Costs may include requests, retrieval, network egress, and temporary recovery infrastructure. Some storage classes require a retrieval process before reading, so the chosen tier affects achievable recovery time.\n\nA widely used planning heuristic is the 3-2-1 rule: maintain three copies, including the working copy, on two different media types, with one copy off-site. It is a useful reminder of redundancy and separation rather than a complete technical specification. Two differently branded disks attached to the same administrator-controlled server may remain vulnerable to one destructive action. Conversely, protection can depend more on independent control planes and failure domains than on a literal distinction between physical media technologies.\n\nGeographical separation must be evaluated against correlated failures. Two facilities can share a floodplain, electrical infrastructure, telecommunications routes, or operational credentials. Storage systems may distribute data internally through replication or erasure coding, but these mechanisms generally protect against specified infrastructure failures rather than preserve every historical state. An authenticated deletion request can be propagated through a highly durable service. Backup architecture therefore distinguishes protection inside a repository from protection against loss or misuse of the repository itself."
    },
    {
      "heading": "Change detection, compression, deduplication, and indexing",
      "text": "Efficient backup systems avoid repeatedly processing unchanged information. File-based tools may compare paths, sizes, modification times, and other metadata before deciding what to copy. More detailed methods read file content or use filesystem change journals. Block-based systems track altered regions of disks or virtual disks, reducing the amount read and transferred when small portions of large files change. Every approach has assumptions: metadata can be misleading, change-tracking state can be invalidated, and some applications rewrite large regions despite small logical changes.\n\nLossless compression represents data more compactly while preserving exact reconstruction. Its effectiveness depends on the source. Text and repetitive records may compress well, whereas encrypted files and already compressed media often yield little further reduction. Compression consumes processing time, so the best throughput is not always obtained with the smallest output. The placement of compression also matters: source-side compression reduces network traffic, while repository-side compression may reduce source load but requires transferring the original input.\n\nData deduplication stores repeated content only once and maintains references from recovery points to shared data. Whole-file deduplication identifies identical files; block-level deduplication recognizes repeated regions within otherwise different files. Fixed-size chunking divides input at regular boundaries, while content-defined chunking chooses boundaries from local data patterns. The latter can preserve matches after insertions shift subsequent bytes, although it introduces additional computation and metadata.\n\nA cryptographic hash function is commonly used to identify candidate duplicate chunks or verify integrity. Hash-based lookup must be designed with collision handling and the relevant security assumptions in mind; some implementations perform additional comparisons or checks. Deduplication also creates dependencies: damage to one shared chunk can affect many recovery points. Repository integrity checks, redundancy, and reliable reference tracking consequently become important as storage savings increase.\n\nThe backup catalog records what exists, when it was captured, and how to reconstruct it. It may map file names and versions to objects, chunks, media identifiers, encryption information, or application recovery metadata. A repository can contain intact data yet be difficult to restore if its catalog is lost. Catalog protection, reconstruction procedures, and compatible software are therefore part of the backup system. Retention cleanup must also distinguish unneeded recovery-point references from chunks still required by other backups; premature deletion can silently break otherwise apparently valid histories."
    },
    {
      "heading": "Security, isolation, and resistance to destructive attacks",
      "text": "Backups concentrate sensitive information and often contain historical material no longer present in active systems. They therefore require confidentiality controls as well as protection from deletion. Encryption in transit reduces exposure during transfer, while encryption at rest protects stored content against specified forms of unauthorized access. Neither measure alone prevents an authorized account from deleting the encrypted objects. Confidentiality, integrity, and availability remain separate design concerns.\n\nKey management is inseparable from recovery. If the only copy of a decryption key is lost with the production system, an intact encrypted backup may be unusable. Conversely, storing unrestricted keys beside the protected data can undermine separation. Backup environments may use dedicated key-management services or a hardware security module, but recovery still depends on access policies, retained key versions, and procedures for operating when normal identity infrastructure is unavailable. Key rotation must preserve access to older backups for their required lifetime.\n\nRansomware and other destructive attacks can target backup servers, repository credentials, catalogs, and retention settings. An attacker who controls both production and backup administration may erase recovery options before disrupting the primary systems. Separate administrative identities, limited privileges, strong authentication, and independently controlled repositories reduce some common paths of compromise. Their effectiveness depends on actual trust relationships: apparently separate accounts may still rely on the same compromised identity service or management host.\n\nOffline and immutable copies address different mechanisms. An offline copy is not reachable through the ordinary operational network while disconnected. An immutable copy resists alteration or deletion for a defined period under the storage system’s enforcement rules. Write-once media and retention-locked object storage are examples, but the scope of protection varies. Some modes permit privileged changes; others enforce retention even against administrators. A network-reachable immutable repository can still face credential theft, confidentiality breaches, denial of access, or exhaustion of capacity.\n\nSecurity also affects restoration. A recent backup may contain malicious software, altered configurations, or compromised credentials introduced before detection. Recovery selection may therefore require identifying a sufficiently early state and rebuilding parts of the environment from trusted installation material. Backups can support recovery after a security incident, but they do not establish that the recovered system is trustworthy. They also do not reverse disclosure: restoring files after data theft does not remove copies already obtained by an attacker."
    },
    {
      "heading": "Retention, rotation, privacy, and long-term preservation",
      "text": "Retention policy specifies how long recovery points remain available and how their density changes with age. A system may retain frequent recent points, daily points for a longer period, and monthly points for longer still. This balances storage consumption with the likelihood of discovering older errors. A deletion noticed immediately may require yesterday’s version; a slowly developing corruption may require a recovery point from several months earlier. Retention is therefore tied to detection delays, not just backup frequency.\n\nRotation schemes organize reuse of media or expiration of recovery points. Grandfather–father–son arrangements typically distinguish daily, weekly, and monthly sets, although implementations vary. Modern disk and cloud systems can express similar schedules through policy rather than physical cartridge rotation. Retention must respect technical dependencies: deleting a baseline that retained incremental backups still require can destroy their usefulness. Systems may avoid this through consolidation, synthetic full creation, or reference-aware deletion.\n\nStorage growth depends on change patterns as well as total dataset size. A database that rewrites large portions of its files may consume more backup capacity than a much larger but mostly static document collection. Deletion at the source does not immediately free backup storage if retained historical points still reference the deleted content. Likewise, removing a recovery point may not reclaim much space when most of its chunks are shared with other retained points. Capacity planning therefore considers unique retained changes, indexing overhead, temporary processing space, and delayed reclamation.\n\nBackups create special questions for data privacy and records management. Retained copies may contain personal information, expired credentials, or records that have been removed from active systems. Policies must account for applicable retention duties, deletion obligations, access restrictions, and legal preservation requirements, which vary by jurisdiction and context. A restored system can reintroduce previously deleted records unless deletion and suppression processes are reconciled after recovery. Backup retention and active-data deletion are related but technically distinct operations.\n\nLong retention also raises the concerns of digital preservation. A medium can remain readable while its file format, application, or licensing environment becomes unavailable. Preserving scientific datasets, design files, or historical databases may require documentation, schemas, software dependencies, and provenance in addition to the original bytes. Periodic media migration and format validation address different risks from ordinary short-term backup rotation. An old system image may reproduce its original environment, but continued usability can still depend on compatible emulation, hardware interfaces, or external services."
    },
    {
      "heading": "Restoration procedures, validation, and operational testing",
      "text": "Restoration is the process that gives backup its practical value. Depending on the failure, it may recover a single file, selected application records, an entire machine, or a complete service environment. These operations have different prerequisites. A file restore may need only repository access and a destination, whereas bare-metal recovery may require boot media, disk layouts, device drivers, encryption keys, network settings, and compatible firmware. An apparently comprehensive data copy can omit configuration essential to rebuilding the system that interprets it.\n\nIntegrity verification operates at several levels. Checksums can detect altered stored blocks; repository checks can confirm that referenced chunks exist; file reconstruction can test whether backup metadata is usable. None of these alone proves that an application will start or produce correct results. A database may require recovery replay and consistency checks, while an application may need validation of permissions, relationships, scheduled tasks, and external integrations. Successful job completion reports only that the backup process met its own completion criteria.\n\nRecovery tests range from sample restores to full disaster-recovery exercises. Sample testing can expose missing files, inaccessible keys, or damaged media with relatively little disruption. A full exercise can reveal dependencies that remain hidden until systems are unavailable: name resolution, identity providers, certificates, installation repositories, and communications arrangements. Tests also measure whether documented recovery order and staffing assumptions are realistic. Exercises are often isolated from production to avoid duplicate scheduled jobs, unintended messages, or conflicting identities.\n\nRecovery performance depends heavily on data shape and workflow. Millions of small files can be limited by metadata operations rather than streaming bandwidth. Large databases may require lengthy log replay after transfer. Deduplicated storage may need to assemble data from widely scattered chunks, and multiple simultaneous restores can compete for the same devices. Instant-recovery systems can start a workload directly from backup storage and migrate it later, shortening initial downtime while potentially imposing performance limits during the transition.\n\nOperational measurement distinguishes backup success, recovery-point age, verified restorability, and service restoration time. These are different quantities and should not be inferred from one another. A missed backup may be tolerable if other valid points satisfy the objective; a sequence of nominally successful jobs may be inadequate if restoration has never been demonstrated. Backup catalogs, restoration instructions, repository configuration, and emergency credentials also require protection because losing the control information can delay access to otherwise intact data."
    },
    {
      "heading": "Workload-specific protection and emerging architectures",
      "text": "Backup scope varies substantially by workload. On a smartphone, a device backup may preserve settings and selected application data while synchronized photos, messages, or documents reside in separate services. Some applications exclude information, use their own export mechanisms, or store secrets in device-bound security components. Restoring a device backup therefore does not necessarily reproduce every aspect of the original device. The boundary between locally protected data and service-maintained data must be understood separately for each application.\n\nVirtual machines can be protected through hypervisor-level snapshots and image backups, allowing restoration of whole guest systems. Application-aware coordination remains necessary where a crash-consistent image is insufficient. Containers and orchestrated applications shift the emphasis: executable images may be reproducible, while persistent volumes, configuration, secrets, and control-plane state contain the irreplaceable information. In systems built around microservices, recovering each component independently may still leave cross-service workflows inconsistent.\n\nFor software as a service, provider-operated redundancy and customer-accessible backup are not necessarily the same capability. A service may preserve availability while offering only limited recovery from user deletion or unwanted configuration changes. Customer-controlled exports and external backups depend on the service’s application programming interface, including rate limits, data coverage, permissions, and version semantics. Exported records may omit relationships, access controls, or history needed for faithful reconstruction, so the existence of a downloadable copy does not itself establish full recoverability.\n\nLarge analytical and scientific environments often contain a mixture of reproducible and non-reproducible assets. Derived datasets may be regenerated from source data and processing code, but regeneration can require substantial computing time or inaccessible external inputs. Machine-learning workloads may need original observations, labels, source code, configuration, and model checkpoints. A reproducibility strategy can reduce what must be backed up, but only if the inputs and build process remain complete and available.\n\nInfrastructure automation can similarly reduce the need to preserve every installed system file. Versioned configuration and deployment code can rebuild replaceable servers, leaving protection focused on authoritative data, secrets, and state. Nevertheless, a repository managed with Git is not automatically an independent backup of every dependency, artifact, or access credential. Recovery architectures increasingly combine reconstruction of software environments with restoration of persistent state; their success depends on preserving the boundary between what can be recreated and what exists only in retained copies."
    }
  ],
  "links": [
    "information-security",
    "disaster-recovery",
    "data-replication",
    "personal-computer",
    "cloud-computing",
    "database",
    "write-ahead-logging",
    "distributed-computing",
    "lossless-data-compression",
    "cryptographic-hash-function",
    "encryption",
    "ransomware",
    "authentication",
    "data-privacy",
    "digital-preservation",
    "smartphone",
    "virtual-machine",
    "software-as-a-service",
    "application-programming-interface",
    "source-code"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}