[IMPALA-6118] Assertion failure in mem-tracker when releasing runtime filter memory

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 2.11.0
• Fix Version/s: Impala 2.11.0
• Component/s: Backend
• Labels: crash
• Target Version: Impala 2.11.0
• Epic Color: ghx-label-5
• Assignee: Tianyi Wang
• Reporter: Philip Martin
• Created: 27/Oct/17 16:09
• Updated: 06/Nov/17 18:55
• Resolved: 06/Nov/17 18:55

Description

In a build on jenkins.impala.io, I hit an assertion in mem-tracker.h. The log file is https://jenkins.impala.io/job/ubuntu-16.04-from-scratch/506/artifact/Impala/logs_static/logs/ee_tests/impalad.FATAL/*view*/ and https://jenkins.impala.io/job/ubuntu-16.04-from-scratch/506/artifact/Impala/logs_static/logs/ee_tests/impalad.ip-172-31-13-161.ubuntu.log.ERROR.20171025-012412.38084/*view*/ . The ERROR log is attached to this ticket.

The stack trace was:

F1025 02:20:43.786911 82485 mem-tracker.h:231] Check failed: tracker->consumption_->current_value() >= 0 (-1052615 vs. 0)
Runtime Filter (Coordinator): Total=-1.00 MB Peak=1.00 MB
*** Check failure stack trace: ***
    @          0x2f1e11d  google::LogMessage::Fail()
    @          0x2f1f9c2  google::LogMessage::SendToLog()
    @          0x2f1daf7  google::LogMessage::Flush()
    @          0x2f210be  google::LogMessageFatal::~LogMessageFatal()
    @          0x17425fb  impala::MemTracker::Release()
    @          0x1fa7e8b  impala::Coordinator::UpdateFilter()
    @          0x186e3cf  impala::ImpalaServer::UpdateFilter()
    @          0x18d824f  impala::ImpalaInternalService::UpdateFilter()
    @          0x1dda35a  impala::ImpalaInternalServiceProcessor::process_UpdateFilter()
    @          0x1dd8308  impala::ImpalaInternalServiceProcessor::dispatchCall()
    @          0x15410ea  apache::thrift::TDispatchProcessor::process()
    @          0x171042b  apache::thrift::server::TAcceptQueueServer::Task::run()
    @          0x170c307  impala::ThriftThread::RunRunnable()
    @          0x170da13  boost::_mfi::mf2<>::operator()()
    @          0x170d8a9  boost::_bi::list3<>::operator()<>()
    @          0x170d5f5  boost::_bi::bind_t<>::operator()()
    @          0x170d508  boost::detail::function::void_function_obj_invoker0<>::invoke()
    @          0x171bdfc  boost::function0<>::operator()()
    @          0x19f3393  impala::Thread::SuperviseThread()
    @          0x19fbf26  boost::_bi::list4<>::operator()<>()
    @          0x19fbe69  boost::_bi::bind_t<>::operator()()
    @          0x19fbe2c  boost::detail::thread_data<>::run()
    @          0x20a7c9a  thread_proxy
    @     0x7fe6536186ba  start_thread
    @     0x7fe65334e3dd  clone
r.java:81)

tarmstrong suggested:

Yeah it's probably another consequence of
https://issues.apache.org/jira/browse/IMPALA-5789. Maybe your patch changed
the timing enough to trigger it.

I think the bug might be related to using directory.capacity() as the
argument to Release(). Calling directory.clear() after releasing the memory
in FitlerState::Disable() won't necessarily deallocate the memory so we
could end up releasing it twice.

Issue Links

• is broken by: IMPALA-5789 Prune all partitions if an always-false runtime filter is received