ByteBufferSlicePool uses a ThreadLocal-based ByteBuffer caching scheme that pretty much results in uncollectable ByteBuffers. That's ok, if the number of ByteBufferSlicePool instances is limited, which I believe was the design intent.

But HttpUpgradeConnectionProvider is not limiting the number of instances. One is created per channel creation, in ConnectionAdaptorImpl.adapt(). This is the cause of WFLY-4881.

Possible solution is to cache the ByteBufferSlicePool in the outer HttpUpgradeConnectionProvider and reuse it for each ConnectionAdaptorImpl.adapt call.

RemoteConnectionProvider.connect also creates a pool for each connect call. Perhaps it could be cached in RemoteConnectionProvider and reused.

David in chat mentioned switching to using global pools as the solution.