Maintaining a unified predictive framework under telemetry space evolution in distributed data storage systems

Abstract

The article addresses the problem of maintaining a unified predictive framework under the evolution of the telemetry space of software-defined distributed data storage systems. The relevance of the problem stems from the fact that, during long-term cluster operation, the available telemetry changes due to platform upgrades, the introduction of new diagnostic metrics, changes in the composition and semantics of system log messages and service states, and modifications to the hardware configuration. Consequently, constructing a separate model for each new configuration leads to fragmentation of historical data and loss of continuity in failure prediction. The telemetry space is represented as an open structure comprising an invariant feature core and variable extensions. The invariant core acts as a compatibility layer between different periods of system operation, whereas the variable extensions reflect the evolution of the platform, infrastructure, and available telemetry. To describe feature evolution, typed compatibility relations are introduced, including preservation, renaming, normalization mapping, aggregation, decomposition, and semantic correspondence. The scientific novelty lies in interpreting the invariant core not as the mechanical intersection of identical metrics across different operational periods, but as a layer of semantic compatibility in which the operational meaning of features can be preserved through admissible transformations. This approach makes it possible to maintain continuity of the predictive model even when metric names, telemetry sources, and data representation formats change. The proposed approach is illustrated using the evolution of the telemetry space of a Ceph cluster from Kraken to Reef. It is shown that changes in telemetry composition can be treated as controlled extensions of a unified predictive framework provided that the invariant core is preserved or can be reconstructed. If this condition is violated, constructing a separate predictive model becomes a justified necessity.