The same IO-bound scenario was run on 2 identical machines, one with RocksDB and the other with TokuDB. In a Zope application (ERP5), using NEO/MariaDB as storage, we filled the disks with a few data streams. The memory leak happened when reading a single whole data stream.

The result is that ERP5/NEO does 2 queries in loop. Here is an example from general log:

SELECT tid, compression, data.hash, value, value_tid FROM obj FORCE INDEX(`partition`) LEFT JOIN data ON (obj.data_id = data.id) WHERE `partition` = 6 AND oid = 94294686 AND tid < 271213171275484455 ORDER BY tid DESC LIMIT 1

SELECT tid FROM obj FORCE INDEX(`partition`) WHERE `partition`=6 AND oid=94294686 AND tid>271177614835290848 ORDER BY tid LIMIT 1

More about NEO's architecture is described at https://www.nexedi.com/blog/NXD-Document.Blog.Optimising.MariaDB.Big.Data

The data stream is 1.5TB uncompressed, taking about 650MB on-disk. RocksDB got OOM-killed when 1TB was read. TokuDB respected its memory constraints until the end.

Each machine has 2 disks of 2TB, each one with a MariaDB DB that holds half of the Zope DB. IOW, we have 2 RocksDB databases of about 2TB. 8GB of RAM for the 2 mysqld (rocksdb_block_cache_size = 2G).