lucenenet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Shad Storhaug <>
Subject RE: TestSearcherManager_Mem
Date Fri, 24 Mar 2017 02:48:43 GMT

FYI - TestSearcherManager_Mem() succeeds much more frequently, but still randomly fails.

Also, although I was able to make the error message change for TestCRTReopen(), I didn't manage
to get it working. The problem is pretty obvious - the WaitForGeneration() method (
is WAY too slow. Even if I increase the wait period from 20 to 60 seconds it still doesn't
finish in time. I played with a few of the variables in ControlledRealTimeReopenThread, but
couldn't get the behavior to change. I verified that PulseAll() gets called, but it doesn't
seem like it is having any effect on the Wait().

For TestEarlyTerminationDifferentSorter(), I reviewed all of the code under test in the Misc
project line by line, but there was nothing significant to fix. So, still broken (sometimes).

Shad Storhaug (NightOwl888)

-----Original Message-----
From: Shad Storhaug [] 
Sent: Thursday, March 23, 2017 8:25 PM
To: Van Den Berghe, Vincent;
Subject: RE: TestSearcherManager_Mem

Parallel universe or not, I think you are making progress. I found a similar IncrementAndGet
issue in the ThreadedIndexingAndSearchingTestCase that I have already corrected. However,
it only mattered in one case, in all other cases the result of IncrementAndGet was not being

It is like someone intentionally wanted to make it impossible to get all of the bugs out of
this code...

Anyway, stupid is as stupid does...I went through and scanned the entire codebase for IncrementAndGet
and compared it against Lucene. Sure enough, the Core.Util.RefCount class was refactored from
its original. I changed it back to the original code (backed by an AtomicInteger/AtomicInt32),
and the TestCRTReopen() test no longer fails almost immediately - after a couple of minutes
it now fails with the message "waited too long for commit generation". I don't know if I fixed
it or broke it more, but it is definitely behaving differently now.

Now, let me see if I can bring your other changes into my universe...perhaps the new failure
has something to do with the reset event.

Shad Storhaug (NightOwl888)

From: Van Den Berghe, Vincent []
Sent: Thursday, March 23, 2017 7:04 PM
Cc: Shad Storhaug
Subject: TestSearcherManager_Mem

Even though I seem to live in a parallel universe where 42 isn't 42 and 4.8 isn't 4.8, I'll
have a stab at resolving TestSearcherManager_Mem.

First, there is a method in TrackingIndexWriter:

        public virtual long GetAndIncrementGeneration()
            return indexingGen.IncrementAndGet();

The implementation calls the wrong indexGen method: it should call GetAndIncrement(), which
doesn't exist in the .NET version. You can add the method to the AtomicLong class.
Too bad there's no Interlocked.PostIncrement, but it's easy enough:

              public long GetAndIncrement()
                     return Interlocked.Increment(ref value) - 1;

And adjust the call accordingly:

        public virtual long GetAndIncrementGeneration()
            return indexingGen.GetAndIncrement();

Next we turn our attention to ControlledRealTimeReopenThread<T>.
There's an event defined as follows:

        private ManualResetEvent reopenCond = new ManualResetEvent(false);

This is not correct, since the remainder of the implementation only Sets and Waits, but never
resets. Once the event Set, the wait will never ... uh... wait the second time around. Change
this as follows:

        private AutoResetEvent reopenCond = new AutoResetEvent(false);

Next, for some reason, time is counted in nanoseconds, but since Environment.TickCount is
in milliseconds, we need to convert it by multiplying by 1000000.
Unfortunately, this is done by multiplication:

Environment.TickCount * 1000000

Since Environment.TickCount is an int and 1000000 is an int, the result is negative unless
you just rebooted your computer in a Tardis doing a polka backwards.

                private const long MS_IN_NS = 1000000;

... and change all other references to 1000000 except one (see below) with MS_IN_NS: this
should solve the overflow problem using C#'s promotion rules.
Next, notice that 64-bit integers are sometimes read outside locks:

                searchingGen = refreshStartGen;
                if (targetGen > searchingGen)
                    while (targetGen > searchingGen)

This isn't guaranteed to be atomic, and I'm a curmudgeon when it comes to parallelism and
atomicity. Change all these lines by:

                Interlocked.Exchange(ref searchingGen, refreshStartGen);
                if (targetGen > Interlocked.Read(ref searchingGen))
                    while (targetGen > Interlocked.Read(ref searchingGen))

In my own spacetime continuum, the test now passes.

Added bonus points: dispose of the waitable event in ControlledRealTimeReopenThread<T>
Dispose method by adding:


(after the Monitor.PulseAll(this); statement)

Extreme added bonus points: the following statement is incorrect, but it works anyway:

                  reopenCond.WaitOne(new TimeSpan(sleepNS / 1000000));//Convert NS to Ticks

(the 1000000 should not be replaced by MS_IN_NS in the new version) The reason why it's incorrect
is because the argument to TimeSpan in the call accepts 100-nanoseconds units, and dividing
nanoseconds by 1000000 yields milliseconds instead. So you will wait a delay of a factor 10000
shorter. It turns out that the correction (divide by 100) will cause timeouts in the tests,
so I left it as-is.

But all of this might be wrong, I may not even exists.


View raw message