| Thread Previous • Date Previous • Date Next • Thread Next |
On 2011-02-28 13:49, Robert Collins wrote:
…do we know that in real life our request isn't naïvely fetching lots of individual objects by id, ones that in the test are fresh in the cache? Think cases of "foo.bar.zot" where the objects you reference were just created by the test. References like that are generally where the fast but numerous queries come from.Are we 100% sure? no. We can (probably should) add in a storm store reset to this. Some tests already do do this, will need to poke around and see why others are not.
Perhaps both perspectives have value. Given a tension between "lightning fast" and "providing useful information," I imagine we want both a not-slow worst case and a very-fast common case.
Since I've done some work on instrumenting Storm for profiling of real object fetches already, maybe we should invest in gathering and reporting those statistics. They would tell us something about how real-world cardinality and hit rates influence the performance tradeoffs.
On a loosely related note, why do we support a pattern of cached fetch-object-by-id but not one of cached fetch-multiple-objects-by-id?
Jeroen
| Thread Previous • Date Previous • Date Next • Thread Next |
