Hi, My understanding about the RocksDB state backend is as follows:
When using a RocksDB state backend, it the checkpoints are backed up locally (to the TaskManager) using the backup feature of RocksDB by taking snapshots from RocksDB which are consistent read-only views on the RockDB database. Each checkpoint is backed up on the task manager node and this checkpoint is asynchronously backed up to the remote HDFS location. When each checkpoint is committed, the records are deleted from RocksDB, allowing RocksDb data folders to remain small. This in turn allows each snapshot to be relatively small. If the Task node goes away due to failure, I assume the RocksDB database is restored from the checkpoints from the remote HDFS. Since each checkpoint state is relatively small, the restoration time from HDFS for the RocksDB database on the new task node is relatively small. The question is, if using really long windows (in hours) if the state of the window gets very large over time, would size of the RocksDB get larger? Would replication to HDFS start causing performance bottlenecks? Also would this need a constant (at checkpoint interval?), read from RocksDB, add more window elements and write to RocksDB. Outside of the read costs, is there a risk to having very long windows when you know you could collect a lot of elements in them. Instead is it safer to perform aggregations on top of aggregations or use your own custom remote store like HBase to persist larger state per record and use windows only to store the keys in HBase. I mention HBase because of its support for column qualifiers allow elements to be added to the same key in multiple ordered column qualifiers. Reading can also be throttled in batches of column qualifiers allowing for the better memory consumption. Is this approach used in practice? Thanks, Sameer
