CRAWD works well on most queries of WDBench. While increasing the
budget would eventually be enough to solve all issues, some of the
problematic queries only need to find a better join ordering:
query_407 and
query_637.
Currently, CRAWD exploits a basic heuristic based on each triple
pattern cardinality. By adding a --force-order parameter, we are
able to force the join order of the count-distinct query and
subsequent count queries to that of the provided --query or
--file.
To run the experiment, do not forget to symlink
rawer-jar-with-dependencies.jar and wdbench-blaze.jnl in this directory. Then, we use
snakemake:
snakemake -c1 just_do_it.dat.
The results are located in the results directory.
Results are much better, even on the smallest 10k-scans budget:
- For
query_407, the expected value is 66053832. CRAWD's estimate is 65254568 with a sublimit of 4 scans (1.2% mean relative error (vs 96% without force order)). - For
query_637, the expected value is 27994204. CRAWD's estimate is 27814624 with a sublimit of 4 scans (0.64% mean relative error (vs 100% without force order)).