Interpretable early warning of statistics induced query plan regression in real-time analytical workloads
2026, vol.18 , no.4, pp. 35-46
Article [2026-04-04]
Analytical databases under continuous ingest accumulate a divergence between the data distribution assumed by the query optimiser and the data actually stored, and plans degrade silently when that divergence crosses a plan-selection boundary. This paper asks whether such regressions can be predicted before execution and attributed to a repairable cause. Three PostgreSQL databases receive identical drifting data under different statistics policies, so regression is measured as a latency ratio against a fresh-statistics reference rather than assumed. Across 1,620 query instances the optimiser’s own cost estimate predicts nothing, whereas a two-predicate rule reading the catalogue and the unexecuted plan matches a gradient-boosted model, identifying all thirty held-out regressions without false positives. Repairing the attributed column resolves every regression, while repairing a second drifted column resolves none.
query optimisation, data drift, cardinality estimation, explainable artificial intelligence, PostgreSQL
https://doi.org/10.59035/VJZK9124
Maithili S. Deshmukh, Sohel A. Bhura. Interpretable early warning of statistics induced query plan regression in real-time analytical workloads. International Journal on Information Technologies and Security, vol.18 , no.4, 2026, pp. 35-46. https://doi.org/10.59035/VJZK9124