One more brief clarification. The ZK problem is this: at any time, if two tests use the same local ZK on the same port, the tests will run fine until whoever finishes first attempts to shut down the ZK, leaving the other test(s) without a connnection and failing in the middle of their run. So its not like all integration tests can't share a ZK right now. Its the lifecycle of all the local ZK's we fire up that is the problem.
On Thu, Mar 28, 2013 at 10:27 AM, Eli Reisman <[email protected]>wrote: > Sorry about the delay. You need to change the names to something > test-specific for all "_bspZookeeper" type folders (one for ZK znodes and > one for ZKManager to use if I rememebr right, see InternalVertexRunner) and > then a port for your local ZK to use. The issue is that during tests, > multiple local ZK's can exist at once (during any integration tests that > execute concurrently) and in that case they clean up (erase) the 2nd one to > finish's data from the temp dirs if they are the same. The ZK's get > similarly confused. > > So it breaks down like this: Giraph's local ZK by default uses 22181. The > IVR uses 22182. The YARN Miniclustwer from giraph 13 is using 22183. It > might be smart to take a look at factoring all integration tests that > requrie ZK into one test that can start just one ZK for all, and one common > Configuration (with the zkLIst set)? But the YARN stuff has to > conditionally compile and I would need to munge those tests into our normal > tests. > > Anyway, hope that helps. > >
