pg_shard performance improvement

[onderkalaci]

This branch implements a cache in front of GetOperatorByType function, so that it prevents unnecessary calls to GetDefaultOpClass function.

With this implementation, we get the same performance improvement that we observed with @ozgune. I may send the details of the results via an email.

fixes #50

Review tasks:

• [x] Replace appearances of OperatorType with OperatorId (also check comments)
• [x] Short-circuit without caching if GetOperatorByType returns InvalidOid
• [x] Clean up the if statement with the three boolean expressions
• [x] Ensure function declaration order matches definition order

[coveralls]

Coverage increased (+0.06%) when pulling 2a49c31696b3008a3f19a277c771597c90884a27 on feature-issue#50 into 16079449a57a4797624cabb60fe0704004ae29bd on develop.

[jasonmp85]

Overall looks good. I think this is a :shipit: after a few changes that I'll add up top.

[jasonmp85]

If you're wondering "why the heck does he add the checklists to the top-level comment rather than down below?" it's because adding them up top gives a little progress bar on the Pull Requests page:

[coveralls]

Coverage increased (+0.11%) to 93.19% when pulling 863f3ecefad5ccb960ee238f12d79c095be605f8 on feature-issue#50 into 16079449a57a4797624cabb60fe0704004ae29bd on develop.

[onderkalaci]

@jasonmp85 I didn't merged this branch and let you do so. I applied check boxes, only not sure about "Short-circuit without caching if GetOperatorByType returns InvalidOid" item.

You suggested me to see analogous code in distribution_metadata.c for a pattern. However, we cannot apply it directly as far as I see. Could you check [my implementation][2], is it OK for you?

[2]: https://github.com/citusdata/pg_shard/pull/57/files#diff-777ad11485b48996843ea21a64147014R281

[jasonmp85]

I said analogous, not identical! You did exactly what I was looking for.

I'm going to close this out myself but need to make one change before I do so: our code style is to always initialize variables. The one place it might be OK to do it on another line is when declaring a local struct variable, but even then we make sure to memset them with zeros as soon as possible (i.e. after all other declarations are out of the way). So when you declare operatorId it should be set to InvalidOid and oldContext should initially be set to NULL.

[jasonmp85]

Looks good. Pushing up the small change I mentioned and merging.

[coveralls]

Coverage increased (+0.12%) to 93.2% when pulling ddd57b647df7384121a93cf4682adb8f1435b036 on feature-issue#50 into 16079449a57a4797624cabb60fe0704004ae29bd on develop.