Fix table RPC to not be interruptible

TODO

@@ -1,29 +1,16 @@
-Object table
+Replication
-------------
-
-
-Rename version table to object table
-In value handle the different versions
-
-So that the table becomes bucket + Sort key = object key -> CRDT(list of versions)
-
-CRDT merge rule:
-- keep one complete version (the one with the highest timestamp)
-- keep all incomplete versions with timestamps higher than the complete version
-
-Cleanup rule: remove incomplete versions after a given delay (say 24h)
-
-
-Block table
 -----------
 
-Table is version_UUID -> BTreeMap<(offset, block hash)> OR Deleted (= CRDT top)
+- for each interval of tokens, we know the list of nodes that are responsible
+- every node watches the current ring and state of the network
+- and thus determines the interval of tokens for which they are responsible
 
 
-Block reference table
----------------------
 
-Table is block_Hash + Sort key: version_UUID -> boolean (true when deleted)
+To do list
+----------
 
-Since the hash key is the same as for the blocks themselves,
+- important: check block values on read and repare corrupted block contents
-we can simply consider the updates to this table as events that increase/decrease a reference counter.
+- less a priority: hinted handoff
+- FIXME in rpc_server when garage shuts down and futures can be interrupted
+  (tokio::spawn should be replaced by a new function background::spawn_joinable)

src/rpc_server.rs

@@ -76,22 +76,26 @@ async fn handler(
 			// and the request handler simply sits there waiting for the task to finish.
 			// (if it's cancelled, that's not an issue)
 			// (TODO FIXME except if garage happens to shut down at that point)
-			let write_fut = async move { garage.block_manager.write_block(&m.hash, &m.data).await };
+			let write_fut = async move {
+				garage.block_manager.write_block(&m.hash, &m.data).await
+			};
 			tokio::spawn(write_fut).await?
 		}
 		Message::GetBlock(h) => garage.block_manager.read_block(&h).await,
 
 		Message::TableRPC(table, msg) => {
-			// For now, table RPCs use transactions that are not async so even if the future
+			// Same trick for table RPCs than for PutBlock
-			// is canceled, the db should be in a consistent state.
+			let op_fut = async move {
-			if let Some(rpc_handler) = garage.table_rpc_handlers.get(&table) {
+				if let Some(rpc_handler) = garage.table_rpc_handlers.get(&table) {
-				rpc_handler
+					rpc_handler
-					.handle(&msg[..])
+						.handle(&msg[..])
-					.await
+						.await
-					.map(|rep| Message::TableRPC(table.to_string(), rep))
+						.map(|rep| Message::TableRPC(table.to_string(), rep))
-			} else {
+				} else {
-				Ok(Message::Error(format!("Unknown table: {}", table)))
+					Ok(Message::Error(format!("Unknown table: {}", table)))
-			}
+				}
+			};
+			tokio::spawn(op_fut).await?
 		}
 
 		_ => Ok(Message::Error(format!("Unexpected message: {:?}", msg))),