Abstract:We describe a new deep learning approach to cardinality estimation. MSCN is a multi-set convolutional network, tailored to representing relational query plans, that employs set semantics to capture query features and true cardinalities. MSCN builds on sampling-based estimation, addressing its weaknesses when no sampled tuples qualify a predicate, and in capturing join-crossing correlations. Our evaluation of MSCN using a real-world dataset shows that deep learning significantly enhances the quality of cardinality estimation, which is the core problem in query optimization.
Submission history
From: Andreas Kipf [view email]
[v1]
Mon, 3 Sep 2018 18:05:12 UTC (214 KB)
[v2]
Tue, 18 Dec 2018 11:16:34 UTC (224 KB)