Correctly count number of items in LRU

[koenbeuk]

Fixes #8741

Microsoft Reviewers: Open in CodeFlow

[modabas]

I think, moving increment count into static add function builders may be a problem, and may cause incrementing count way above real value at startup with many concurrent operations.
Also does LRU.GetOrAdd need a fix? It seems to work properly and does not increment count at update. For LRU.AddOrUpdate, if intent is not to increment count on update, perhaps following code may work (didn't test)

    public TValue AddOrUpdate<TState>(TKey key, Func<TState, TKey, TValue> addFunc, TState state)
    {
        var generation = GetNewGeneration();
        var storedValue = cache.AddOrUpdate(
            key,
            static (key, state) =>
            {
                var (_, outerState, addFunc, generation) = state;
                return new TimestampedValue(addFunc(outerState, key), generation);
            },
            static (key, existing, state) =>
            {
                var (self, outerState, addFunc, generation) = state;
                return new TimestampedValue(addFunc(outerState, key), self.GetNewGeneration());
            },
            (Self: this, State: state, AddFunc: addFunc, Generation: generation));

        var result = storedValue.Value;

        if (storedValue.Generation == generation)
        {
            Interlocked.Increment(ref count);
            AdjustSize();
        }

        return result;
    }

[modabas]

I have put together a quick test app. Added a CacheCount property to LRU class that returns actual item count in cache to demonstrate what i meant.

[koenbeuk]

I think, moving increment count into static add function builders may be a problem, and may cause incrementing count way above real value at startup with many concurrent operations.

Thanks for catching that, this indeed seems to be an issue. I've updated this PR to use the logic as you proposed.

[mikescandy]

It seems this is the cause of an issue we're having in our production deployment that I haven't noticed until recently

Aug 23 is when we upgraded from 7 to 7.2.1, which includes the live grain migration and the new lru logic.
For context, I see the count exploding only for a stateless worker grain that we call for an health check
It would be awesome if this release could be release asap :)

[rwkarg]

@mikescandy Release is out with the updates:
https://github.com/dotnet/orleans/releases/tag/v7.2.4

It seems this is the cause of an issue we're having in our production deployment that I haven't noticed until recently

Aug 23 is when we upgraded from 7 to 7.2.1, which includes the live grain migration and the new lru logic. For context, I see the count exploding only for a stateless worker grain that we call for an health check It would be awesome if this release could be release asap :)