Is this the same as data cleansing?

No. Cleansing addresses data defects; this QoD usage describes requirements for replicated versions.

Must all three dimensions be active?

The research allows policies using one or more dimensions, with implementation-specific details.

Does bounded divergence automatically mean linearizability?

No. A permitted divergence bound and linearizable operation semantics are different guarantees.

AI Glossary / Data engineering and architecture / Advanced

Quality of Data

Definition

Quality of Data, or QoD, here refers to a research approach for expressing required consistency of replicated data through limits on divergence. It uses dimensions such as time, pending updates and value difference to describe how far a replica may fall behind, rather than treating all data as equally urgent to replicate.

What it means

The supplied term points to the QoD work associated with vector-field consistency and geo-replicated stores, including HBase-QoD. This is a specific distributed-systems meaning. It should not be confused with general data quality measures such as correctness, completeness or validity.

A replica can contain accurate but outdated information. Conversely, perfectly synchronized replicas can all contain the same incorrect source value. QoD in this context addresses the relationship between replicated versions, not whether the underlying business fact is true.

Different data can tolerate different divergence. A slowly changing descriptive field may have different requirements from a rapidly changing value used for an immediate decision. The tolerances must come from the application's semantics rather than a universal quality percentage.

How it works

The research represents constraints using a three-dimensional vector, commonly written K = (θ, σ, ν). The dimensions concern time, sequence and value: how long updates can be delayed, how many changes can remain pending and how much a value may diverge under an appropriate metric.

A replication mechanism tracks the relevant changes and uses the configured bounds to decide when propagation is required. Constraints can be associated with data groupings rather than applying one identical policy everywhere.

HBase-QoD explored this approach as a research extension, including grouped propagation of related updates. Its published evaluation emphasized the sequence dimension. The paper should not be read as evidence that every current HBase installation includes this facility or that all combinations of guarantees have been independently demonstrated.

Enforcement also depends on communication and failure assumptions. A disconnected receiver cannot receive a fresh value by declaration alone. Applications need a defined response when the desired freshness cannot be achieved.

Three practical examples

A replicated activity dashboard distinguishes display data that can tolerate a short delay from status changes that need more urgent propagation.
A research prototype batches low-priority updates until a configured pending-update threshold is reached.
An analytical service evaluates a domain-specific value-divergence metric for a replicated measurement, while separately checking the measurement's real-world accuracy.

Think of it this way

Think of several offices holding copies of a changing notice board. Some notices can wait for the next scheduled delivery, while others become unacceptable after a small number of revisions or a significant change in content.

The analogy describes different tolerances for outdated copies. It does not make the notices true, and it cannot guarantee delivery when the transport link is unavailable. The communication mechanism and failure response remain essential.

Worked example

Checking three different bounds

Use an invented policy for a replicated numeric estimate: trigger attention at ten seconds since the relevant refresh, three pending updates or a value difference of five units, whichever threshold is reached first. These numbers are illustrative, not defaults from the research.

The replica holds 100. At two seconds the source becomes 102; there is one pending update and a difference of two. At four seconds the source becomes 104; there are two pending updates and a difference of four. At six seconds the source becomes 107; there are three pending updates and a difference of seven.

The sequence and value thresholds now trigger attention even though ten seconds have not elapsed. The system must initiate its defined propagation or failure-handling response. Transmission itself takes time, so a trigger rule must not be confused with proof that the remote value was already refreshed at that instant.

Illustrative case study

A fictional analytics team experiments with differentiated replication for a distributed dashboard. It initially labels its output “high-quality data” because its propagation policy meets a freshness target in a test.

A later review finds that one source measurement was entered incorrectly. Every replica receives that incorrect value promptly. The team revises its terminology and reporting: replica freshness and divergence are measured separately from source validation and measurement quality.

The lesson is that consistency quality answers a particular question about copies and updates. It does not replace the checks needed to establish that the data itself is appropriate for a business decision.

Common misconceptions and mistakes

QoD here is a complete measure of general data quality. Its stated scope concerns replica divergence and consistency requirements.
A small time delay also guarantees a small value difference. A value can change sharply in a short interval, so the dimensions express different constraints.
A research prototype establishes universal product guarantees. Availability, implementation scope and failure behavior must be verified for the actual system.

Business use

The QoD idea is useful when discussing how much replica staleness an application can tolerate and whether every update needs the same propagation urgency. It encourages requirements to be expressed in terms of data meaning rather than an undifferentiated label such as eventual consistency.

For an assessment, identify the authoritative update stream, the unit of replication, the divergence metric and the response to exceeded bounds. Explain whether a requirement concerns when sending begins, when receiving completes or what a reader is allowed to observe. These are different service promises.

For AI applications, fresh replicated features can still be incomplete, biased or incorrectly measured. Evaluate those dimensions separately. Any claimed performance benefit should be supported by the tested workload and configuration, not inferred from the existence of a tunable consistency model.

Check your understanding

Why does the example's policy trigger before the ten-second time threshold?

Answer: Three updates are pending and the value difference has reached seven units. The policy checks separate sequence and value thresholds as well as elapsed time.

Related terms

Vector-field consistency
Replication (statistics)
Divergence
Geo-replication
Data Quality
bounded staleness
replication
eventual consistency
service-level objective

Learn next

Read replication and consistency models. For one replicated value, define separate time, update-count and value-difference tolerances, then state what readers should experience if communication cannot satisfy them.

Sources

García-Recuero, Esteves and Veiga, Quality-of-Data for Consistency Levels in Geo-replicated Cloud Data Stores: https://www.dpss.inesc-id.pt/~sesteves/papers/cloudcom2013-qod-consistency-geo-replicated-data-stores.pdf. Primary paper; relevant architecture and implementation passages reviewed.

Esteves, Silva and Veiga, Quality-of-Service for Consistency of Data Geo-replication in Cloud Computing: https://www.dpss.inesc-id.pt/~lveiga/papers/25-2012-europar-sesteves.pdf. Primary research on divergence dimensions.

UK Government Data Quality Framework: https://www.gov.uk/government/publications/the-government-data-quality-framework/the-government-data-quality-framework. Primary contrasting general-quality framework. The threshold trace and dashboard case are original illustrations.

Previous: Qualitative probabilistic networks Next: Quantification