Hi David, Thank you for taking a look at that. I think the test already uses just one tablet server, so no replicas would be possible. I see the following code in the test:
StartCluster(1); // Start MiniCluster with a single tablet server. TestWorkload workload(cluster_.get()); workload.set_num_replicas(1); workload.Setup(); // Convenient way to create a table. Did I miss something? I.e. should I toggle just another control knob somewhere? Thanks, Alexey On Wed, Oct 12, 2016 at 1:43 PM, David Alves <[email protected]> wrote: > Hi Alexey > > Thanks for going down the rabbit hole. > Could you try your patch without tablet replication for that test? If the > problem persists it's unlikely that it's related to the current consistency > gaps we have. > I'm a bit suspicious in that it seems to be doing snapshot scans without > retrying, which is what we're doing pretty much everywhere else to work > around our gaps. > > -david > > On Wed, Oct 12, 2016 at 1:06 PM, Alexey Serbin <[email protected]> > wrote: > > > One small update: the issue might be not in GC logic, but some other > > flakiness related to reading data at snapshot. > > > > I updated the patch so the only operations the test now does are inserts, > > updates and scans. No tablet merge compactions, redo delta compactions, > > forced re-updates of missing deltas, or moving time forward. The updated > > patch can be found at: > > https://gist.github.com/alexeyserbin/06ed8dbdb0e8e9abcbde2991c6615660 > > > > The test firmly fails if running as described in the previous message in > > this thread, just use the updated patch location. > > > > David, may be you can take a quick look at that as well? > > > > > > Thanks, > > > > Alexey > > > > On Wed, Oct 12, 2016 at 2:01 AM, Alexey Serbin <[email protected]> > > wrote: > > > > > Hi, > > > > > > I played with the test (mostly in background), making the failure > almost > > > 100% reproducible. > > > > > > After collecting some evidence, I can say it's a server-side bug. I > > think > > > so because the reproduction scenario I'm talking about uses good old > > > MANUAL_FLUSH mode, not AUTO_FLUSH_BACKGROUND mode. Yes, I've modified > > the > > > test slightly to achieve higher reproduction ratio and to clear the > > > question whether it's AUTO_FLUSH_BACKGROUND-specific bug. > > > > > > That's what I found: > > > 1. The problem occurs when updating rows with the same primary keys > > > multiple times. > > > 2. It's crucial to flush (i.e. call KuduSession::Flush() or > > > KuduSession::FlushAsync()) freshly applied update operations not just > > once > > > in the very end of a client session, but multiple times while adding > > those > > > operations. If flushing just once in the very end, the issue becomes > 0% > > > reproducible. > > > 3. The more updates for different rows we have, the more likely we > hit > > > the issue (but there should be at least a couple updates for every > row). > > > 4. The problem persists in all types of Kudu builds: debug, TSAN, > > > release, ASAN (in the decreasing order of the reproduction ratio). > > > 5. The problem is also highly reproducible if running the test via > the > > > dist_test.py utility (check for 256 out of 256 failure ratio at > > > http://dist-test.cloudera.org//job?job_id=aserbin.1476258983.2603 ) > > > > > > To build the modified test and run the reproduction scenario: > > > 1. Get the patch from https://gist.github.com/alexeyserbin/ > > > 7c885148dadff8705912f6cc513108d0 > > > 2. Apply the patch to the latest Kudu source from the master branch. > > > 3. Build debug, TSAN, release or ASAN configuration and run with the > > > command (the random seed is not really crucial, but this gives better > > > results): > > > ../../build-support/run-test.sh ./bin/tablet_history_gc-itest > > > --gtest_filter=RandomizedTabletHistoryGcITest > > .TestRandomHistoryGCWorkload > > > --stress_cpu_threads=64 --test_random_seed=1213726993 > > > > > > 4. If running via dist_test.py, run the following instead: > > > > > > ../../build-support/dist_test.py loop -n 256 -- > > > ./bin/tablet_history_gc-itest --gtest_filter= > > > RandomizedTabletHistoryGcITest.TestRandomHistoryGCWorkload > > > --stress_cpu_threads=8 --test_random_seed=1213726993 > > > > > > Mike, it seems I'll need your help to troubleshoot/debug this issue > > > further. > > > > > > > > > Best regards, > > > > > > Alexey > > > > > > > > > On Mon, Oct 3, 2016 at 9:48 AM, Alexey Serbin <[email protected]> > > > wrote: > > > > > >> Todd, > > >> > > >> I apologize for the late response -- somehow my inbox is messed up. > > >> Probably, I need to switch to use stand-alone mail application (as > > iMail) > > >> instead of browser-based one. > > >> > > >> Yes, I'll take a look at that. > > >> > > >> > > >> Best regards, > > >> > > >> Alexey > > >> > > >> On Mon, Sep 26, 2016 at 12:58 PM, Todd Lipcon <[email protected]> > > wrote: > > >> > > >>> This test has gotten flaky with a concerning failure mode (seeing > > >>> "wrong" results, not just a timeout or something): > > >>> > > >>> http://dist-test.cloudera.org:8080/test_drilldown?test_name= > > >>> tablet_history_gc-itest > > >>> > > >>> It seems like it got flaky starting with Alexey's > > >>> commit bc14b2f9d775c9f27f2e2be36d4b03080977e8fa which switched it to > > >>> use AUTO_FLUSH_BACKGROUND. So perhaps the bug is actually a client > bug > > and > > >>> not anything to do with GC. > > >>> > > >>> Alexey, do you have time to take a look, and perhaps consult with > Mike > > >>> if you think it's actually a server-side bug? > > >>> > > >>> -Todd > > >>> > > >>> -- > > >>> Todd Lipcon > > >>> Software Engineer, Cloudera > > >>> > > >> > > >> > > > > > >
