Cardinality Estimator: Processing SQL with a Vertical Scanning Convolutional Neural Network

A BiLSTM cardinality estimator in complex database systems based on attention mechanism

“…The mean q ‐error of our model is 42.2 times smaller than that of IBJS [31]. All q ‐error measurements of BACE are better than that of VSCNN [23] on the Synthetic workload. Because the Synthetic workload is simple and BACE has an attention mechanism that can consider abstract information more efficiently, which consider the words in the whole WHERE clauses all at once.…”

“…Because our model has memory structure, it can deal with multiple joins of tables and complex predicates of the workload. All q ‐error measurements of BACE are better than that of VSCNN [23] on the JOB ‐ Light workload. VSCNN [23] uses three different convolution kernels to extract information from the WHERE clauses, but it is difficult to handle the order information of the words in the WHERE clauses.…”

“…All q ‐error measurements of BACE are better than that of VSCNN [23] on the JOB ‐ Light workload. VSCNN [23] uses three different convolution kernels to extract information from the WHERE clauses, but it is difficult to handle the order information of the words in the WHERE clauses. However, BACE uses the BiLSTM network structure, which can take the order of words from two directions into consideration.…”

“…We use q-error as an evaluation indicator by defining in Equation ( 22) and the loss function given in Equation (23). qerror can measure the relative gap between the estimated value and the true value.…”

“…Qiao et al. [23] propose a convolution estimation method for vertical scanning (called VSCNN). This method uses three different convolution kernels with different widths to scan the condition clauses from top to bottom to obtain the information in the query, but it is not compatible with discrete long string predicates because it cannot recognise the sequence of words in time well.…”