[BUG] Performance degradation in v5.0.18 about litedb

[alexbereznikov]

So I decided to profile this repro to understand what leads to this performance degradation. Profiling for 5.0.19 gives the following results:

As we can see, CommitAndReleaseTransaction takes 4x more time than upsert itself, and for some reason we're doing writing to disk in the same thread.

And here is how it looks for v5.0.17:

As we can see, in v5.0.17 CommitAndReleaseTransaction takes little to no time, and all the writing is offloaded to a separate thread.

After some investigation, turns out that there is an issue in AsyncManualResetEvent - when Set is being called, execution continues in DiskWriterQueue.ExecuteQueue after await _queueHasItems.WaitAsync(). I would say, we don't need this separate class at all, as .NET provides a way to await any WaitHandle using thread pool (see https://github.com/StephenCleary/AsyncEx/blob/v5.0.0/src/Nito.AsyncEx.Interop.WaitHandles/Interop/WaitHandleAsyncFactory.cs, DoFromWaitHandle). So I changed that event to ManualResetEventSlim in conjunction with this method.

Now repro runs in 42 seconds. 4x faster than 5.0.19 but still ~3x slower than 5.0.17. Further profiling results:

As we can see, for some reason we do call BasePage.ToString a lot, as well as formatting strings. I'm 100% sure that this should not happen on the hot path. Here is the issue (BasePage.Get):

ENSURE(this.IsValidPos(position), $"invalid segment position in index footer: {ToString()}/{index}");
ENSURE(this.IsValidLen(length), $"invalid segment length in index footer: {ToString()}/{index}");

So I just commented these 2 lines to check and 1 in IndexService, and it runs in 13 seconds, pretty close to v5.0.17.

[alexbereznikov]

Created PR to fix the first part - #2453

[alexbereznikov]

Created PR to fix the second part - #2454

[alexbereznikov]

@mbdavid Sorry but I believe this is a critical issue, in the current state v5.0.19 is not usable. I did all the research and fixes, please consider merging it.

[viveleltsi]

I agree. This fix will help me too. thank you for your work @alexbereznikov . I hope it will be merged