The work by INESC TEC and University of Minho researchers examines whether Artificial Intelligence (AI) can already replace traditional methods for deciding how a database responds to queries.
Luís França and José Pereira, researchers at INESC TEC and the University of Minho, received the Best Paper award at INForum 2026 for their paper – Evaluating the Practicality of Learned Cardinality Estimators in Query Optimizers.
Whenever we run a query on a database, the system must choose the fastest route to reach the answer. To do so, it needs to estimate how many results it will obtain at each intermediate step. This is known as cardinality estimation, and a poor estimate can lead to inefficient choices and slower query responses.
In traditional systems, these estimates rely on statistics about the data. In large distributed repositories, known as data lakes, where information is stored in files, keeping these statistics up to date is considerably more difficult. For this reason, the scientific community has proposed using AI models, which, in many cases, provide more accurate estimates.
But is it worth it? A more accurate model can also be slower, and the time required to query it may cancel out the gains in accuracy. This was the question the INESC TEC team set out to answer. Although the study was essentially experimental, the paper assessed, in practical terms, whether traditional algorithms could feasibly be replaced by AI methods inside databases.
To this end, the researchers integrated a state-of-the-art AI model into an open-source query optimiser and tested five approaches to speeding up its operation, using both conventional processors (CPUs) and graphics processing units (GPUs).
The results indicate that, with current technology, these models are still not a viable alternative to traditional methods in a general-purpose query optimiser, mainly due to the time they take to produce results.
The paper proposes an intermediate solution: a hybrid architecture in which traditional methods handle simple queries quickly, while the AI model is reserved for more demanding analyses, where accurate estimates can make a real difference.
“Receiving this award is a very important milestone that recognises the value of the work we have carried out and confirms that we are on the right track, providing a strong incentive to take on new challenges in this field of study”, said Luís França, one of the authors of the work at INESC TEC and a master’s student at the School of Engineering of the University of Minho.
The research was developed as part of the ADAPQO project, funded through the CMU Portugal programme, with support from FCT – Foundation for Science and Technology.
INForum is a national event dedicated to sharing and discussing research and innovation in Computer Science, fostering closer ties between academia and industry. The 2026 edition took place at the Faculty of Sciences and Technology of NOVA University Lisbon on 3 and 4 September.
News, current topics, curiosities and so much more about INESC TEC and its community!
