Generate graph edge costs based on measured performance

[kwmsmith]

The cost for odo's graph isn't generated on the actual runtime cost of the conversion, which leads odo to choose a conversion path that is sub-optimal.

To the extent that an edge's cost accurately reflects its actual runtime, the better odo can choose an optimal path from one format to another.

To generate these costs requires representative data sources for the different backends and a script that runs odo's basic converters for each backend pairing. To be more accurate, the script would need to run on differently sized data and on different platforms, but even just a single platform run with decently sized data sources will improve the current heuristic costs.

These generated costs would be generated by the script and stored for usage at runtime. Odo will need to be refactored to read in these generated costs and apply them to the graph edges.

@cpcloud thoughts?

[cpcloud]

I think this is a great idea. How difficult do you think it would be to use something like pandas vbench or airspeed velocity to generate a benchmark database that we'd use to measure the cost of various conversions of different sizes.

[kwmsmith]

We looked into vbench for a previous project, and at the time, we found it to be not up-to-snuff: it had a lot of hard-coded assumptions about its environment, etc. Perhaps that's changed, I'll be glad to see if it has.

I'll take a look at these and evaluate. I've never used ASV, but have been wanting to kick the tires.