Not sure "bug" is the right description, because I can't say for sure that the large number of SSTables is the cause of the memory issues. I'll share my research so far:
Under high read-load with a very large number of compressed SSTables (caused by the initial default 5mb sstable_size in LCS) it seems memory is exhausted, without any room for GC to fix this. It tries to GC but doesn't reclaim much.
The node first hits the "emergency valves" flushing all memtables, then reducing caches. And finally logs 0.99+ heap usages and hangs with GC failure or crashes with OutOfMemoryError.
I've taken a heapdump and started analysis to find out what's wrong. The memory seems to be used by the byte backing the HeapByteBuffer in the "compressed" field of org.apache.cassandra.io.compress.CompressedRandomAccessReader. The byte are generally 65536 byes in size, matching the block-size of the compression.
Looking further in the heap-dump I can see that these readers are part of the pool in org.apache.cassandra.io.util.CompressedPoolingSegmentedFile. Which is linked to the "dfile" field of org.apache.cassandra.io.sstable.SSTableReader. The dump-file lists 45248 instances of CompressedRandomAccessReader.
Is this intended to go this way? Is there a leak somewhere? Or should there be an alternative strategy and/or warning for cases where a node is trying to read far too many SSTables?
Searching through the code I found that PoolingSegmentedFile keeps a pool of RandomAccessReader for re-use. While the CompressedRandomAccessReader allocates a ByteBuffer in it's constructor and (to make things worse) enlarges it if it's reasing a large chunk. This (sometimes enlarged) ByteBuffer is then kept alive because it becomes part of the CompressedRandomAccessReader which is in turn kept alive as part of the pool in the PoolingSegmentedFile.