sigmod-2018

Results: http://sigmod18contest.db.in.tum.de/leaders.shtml

Team: vsb_ber0134

Institution: VSB-TUO

Jakub Beránek, BER0134, VSB-TUO, FEI (department 460), Master's (graduate, last year), [email protected]

Advisors:

• Ing. Radim Bača, Phd., FEI, VSB-TUO, [email protected]
• Ing. Petr Lukáš, FEI, VSB-TUO, [email protected]

Description:

My solution heavily relies on indices, which are built for all join columns. Columns are grouped into small blocks and then sorted with radix sort.

After the indices are built, the joins are made with a combination of merge sort join and nested loop join with index accesses. The created plan is a left deep tree with small order optimizations based on sorted columns or expected result sizes. All joins are then divided into tens of small tasks which are computed in parallel.

Queries are heavily rewritten. Primary/foreign key pairs are checked and joins based on them are removed if possible. Filters for individual join columns are expanded for the whole join component. Information taken from filters and min/max of columns are used to determine whether the query can return any results. Filters are also optionally JIT compiled to x64 assembly.

Third-party code:

• radix sort (https://github.com/voutcn/kxsort), MIT