cache: reserve memtable memory from block cache about pebble HOT 3 CLOSED

@Ryanfsdf this is the issue I mentioned yesterday while we were chatting.

from pebble.

RocksDB provides this functionality via WriteBufferManager. The way it works is that WriteBufferManager pins a series of "virtual" entries in the cache. These entries actually contain no data, but the are marked as being 256KB in size. I believe 256KB entries are used, rather than the actual memtable size, so that space can be reserved by each shard in the cache.

Pebble currently has no support for entries like this in its cache. An alternative approach is to directly model reserved space in the cache, which can then distributed that reservation to each of the shards.

from pebble.

Reserving space in the cache is easy, but we don't currently have a signal as to when a memtable is unused. Instead we rely on GC to reclaim the memtable memory. I suppose we could use a finalizer, but those are generally frowned upon and touch rarely used code paths in the GC. It might be better to explicit track memtable references. There are not too many of them: DB.mu.mem.queue and readState.flushables. I can't think of any others right now.

Explicit tracking would also open up the possibility of recycling the memtable memory, rather than waiting for the GC to reclaim it.

from pebble.