Some ideas about using multiple threads to run a query.
== Position at N% of table/index ==
If we want to run these with N threads, we need to give 1/Nth of table to each thread. (An alternative is to run one "reader" thread and distribute work to multiple compute threads. The problem with this is that reading from the table won't be parallel. This will put a cap on the performance.)
In order to do that, we will need storage engine calls that do
- "position at N% in the table"
- "position at N% in the index range between [C1 and C2]".
these calls would also let us build equi-height histograms based on sampling.
== General execution ==
There are many works about converting SQL into MapReduce jobs. Are they relevant to this task? The difference seems to be in the Map phase - we assume that source data is equi-distant to all worker threads.
== Evaluation ==
It would be nice to assess how much speedup we will get. In order to get an idea, we could break the query apart and run the parts manually. The merge step could also be done manually in some cases (by writing to, and reading from temporary tables).