Writing a HDFS clone in Go to learn more about Go and the nitty-gritty of distributed systems.

• MetaDataNode/DataNode handle uploads
• MetaDataNode/DataNode handle downloads
• DataNode dynamically registers with MetaDataNode
• DataNode tells MetaDataNode its blocks on startup
• MetaDataNode persists file->blocklist map
• DataNode pipelines uploads to other DataNodes
• MetaDataNode can restart and DataNode will re-register (heartbeats)
• Tell DataNodes to re-register if MetaDataNode doesn't recognize them
• Drop DataNodes when they go down (heartbeats)
• DataNode sends size of data directory (heartbeat)
• MetaDataNode obeys replication factor
• MetaDataNode balances based on current reported space
• MetaDataNode balances based on expected new blocks
• MetaDataNode handles not enough DataNodes for replication
• Have MetaDataNode manage the block size stuff (in HDFS, clients can change this per-file)
• Re-replicate blocks when a DataNode disappears
• Drop over-replicated blocks when a DataNode comes up
• Looking at DataNode utilization should take into account the DeletionIntents and ReplicationIntents
• Grace period for replicating new blocks
• MetaDataNode balances blocks as it runs!
• Record hash digest of blocks, reject send if hash is wrong
• DataNode needs to keep track of blocks it's receiving / deleting / checking so that the integrity checker can run only on real blocks
• Remove blocks if checksum doesn't match
• Run a cluster in a single process for testing
• Structure things better
• Resiliency to weird protocol stuff (run the RPC loop manually?)
• Command line parser doesn't work that well (try "main datanode -help")
• Events from servers for testing
• Better configuration handling (defaults)
• Allow decommissioning nodes
• Better logging, so warnings normally can be fatal for tests (two levels: warn that this process broke, and warn that somebody we're communicating with broke)
• Don't need to wait around to delete blocks, just prevent any new reads and we'll come back to them
• DataNode should do stuff on startup, and then spawn workers, not just spawn everybody (race conditions with address and data directories)
• Support multiple MetaDataNodes somehow (DHT? Raft? Get rid of MetaDataNodes and use Gossip?)
• Keep track of MoveIntents (subtract from predicted utilization of node), might fix the volatility when re-balancing
• HashiCorp claims heartbeats are inefficient (linear work aafo number of nodes). Use Gossip?
• Don't force a long-running connection for creating a file, give the client a lease and let them re-connect
• If a client tries to upload a block and every DataNode in its list is down, it needs to get more from the MetaDataNode.
• Keep track of blocks as we're creating a file, if the client bails before committing then delete the blocks.