Hi everyone, Just an update on our current status on POWER -
Backend test: On POWER, out of 5 backend tests failures, we got 3 failures fixed by upgrading LLVM to 3.8 and lz4 library. 1 failure in llvm-codegen-test is due to gcc 5 incompatibility with clang 3.8, for which LLVM community is already working on. And last failure is being debugged. Frontend test: On POWER, FE tests failures are still many (Total Tests run: 546, Failures: 226, Errors: 77, Skipped: 3 ). Most of these failures are due to test data not being generated and loaded correctly. Test data shared by Cloudera community is also not working, for which discussion is in progress. Test data generation/loading: On POWER, there are still errors generating/loading test data. We've also setup Impala on x86 environment and could successfully generate and load test data on it. So, now we'll try using x86 data that we generated on POWER just to verify frontend tests. Thank you for all your cooperation. Thanks, Nishidha From: Silvius Rus <[email protected]> To: Nishidha Panpaliya/Austin/Contr/IBM@IBMUS Cc: Sudarshan Jagadale/Austin/Contr/IBM@IBMUS, David Clissold/Austin/IBM@IBMUS, Valencia Serrao/Austin/Contr/IBM@IBMUS, Andrea D Franklin/Austin/IBM@IBMUS, Manish Patil/Austin/Contr/IBM@IBMUS, Zhi Zhi NY Yang <[email protected]> Date: 04/20/2016 12:00 PM Subject: Re: Impala follow up On Tue, Apr 19, 2016 at 6:02 AM, Nishidha Panpaliya <[email protected]> wrote: We are sorry that we cannot create gerrit code review as it will need us to rebase again from cdh5-trunk and then again porting new code on ppc64le. To avoid the loop of porting/rebuilding/testing/patch creation, I've created a patch between our code and cdh5-trunk (Commit ID: 00c1d6217cd2284fe81a442e288ba737e71ef99c) when we rebased last time. Thank you for the writeup, Nishidha. It's useful to see the changes broken down. It would be painful to review the whole patch outside gerrit especially if we iterate a few times as we would need to highlight just deltas from the previous review. Have you considered starting a branch and applying your changes on this branch as multiple commits? That way it would be easier to stay close to cdh5-trunk by rebasing frequently. Also, we could review your changes incrementally and we can set up continuous testing to make sure that your branch stays green on x86 throughout your porting effort to ppc64le and can be merged easily to the cdh5-trunk. I'm concerned that with the current approach it would be very hard to merge back to cdh5-trunk as your code is diverging. So even when you're "done", you'd be unable to upgrade and it would be hard even to apply Impala maintenance patches. Let me discuss this with a few Cloudera developers to get more opinions on how to do this right and with as little pain as possible.
