src/table/table.rs

@@ -143,7 +143,7 @@ impl<F: TableSchema, R: TableReplication> Table<F, R> {
 		self.data.queue_insert(tx, e)
 	}
 
-	pub async fn insert_many<I, IE>(&self, entries: I) -> Result<(), Error>
+	pub async fn insert_many<I, IE>(self: &Arc<Self>, entries: I) -> Result<(), Error>
 	where
 		I: IntoIterator<Item = IE> + Send + Sync,
 		IE: Borrow<F::E> + Send + Sync,
@@ -161,52 +161,149 @@ impl<F: TableSchema, R: TableReplication> Table<F, R> {
 		Ok(())
 	}
 
-	async fn insert_many_internal<I, IE>(&self, entries: I) -> Result<(), Error>
+	async fn insert_many_internal<I, IE>(self: &Arc<Self>, entries: I) -> Result<(), Error>
 	where
 		I: IntoIterator<Item = IE> + Send + Sync,
 		IE: Borrow<F::E> + Send + Sync,
 	{
-		let mut call_list: HashMap<_, Vec<_>> = HashMap::new();
+		// The different items will have to be stored on possibly different nodes.
+		// We will here batch all items into a single request for each concerned
+		// node, with all of the entries it must store within that request.
+		// Each entry has to be saved to a specific list of "write sets", i.e. a set
+		// of node within wich a quorum must be achieved. In normal operation, there
+		// is a single write set which corresponds to the quorum in the current
+		// cluster layout, but when the layout is updated, multiple write sets might
+		// have to be handled at once. Here, since we are sending many entries, we
+		// will have to handle many write sets in all cases. The algorihtm is thus
+		// to send one request to each node with all the items it must save,
+		// and keep track of the OK responses within each write set: if for all sets
+		// a quorum of nodes has answered OK, then the insert has succeeded and
+		// consistency properties (read-after-write) are preserved.
 
+		// Some code here might feel redundant with RpcHelper::try_write_many_sets,
+		// but I think deduplicating could lead to more spaghetti instead of
+		// improving the readability, so I'm leaving as is.
+
+		let quorum = self.data.replication.write_quorum();
+
+		// Serialize all entries and compute the write sets for each of them.
+		// In the case of sharded table replication, this also takes an "ack lock"
+		// to the layout manager to avoid ack'ing newer versions which are not
+		// taken into account by writes in progress (the ack can happen later, once
+		// all writes that didn't take the new layout into account are finished).
+		// These locks are released when entries_vec is dropped, i.e. when this
+		// function returns.
+		let mut entries_vec = Vec::new();
 		for entry in entries.into_iter() {
 			let entry = entry.borrow();
 			let hash = entry.partition_key().hash();
-			// TODO: use write sets
+			let write_sets = self.data.replication.write_sets(&hash);
-			let who = self.data.replication.storage_nodes(&hash);
 			let e_enc = Arc::new(ByteBuf::from(entry.encode()?));
-			for node in who {
+			entries_vec.push((write_sets, e_enc));
-				call_list.entry(node).or_default().push(e_enc.clone());
+		}
+
+		// Compute a deduplicated list of all of the write sets,
+		// and compute an index from each node to the position of the sets in which
+		// it takes part, to optimize the detection of a quorum.
+		let mut write_sets = entries_vec
+			.iter()
+			.map(|(wss, _)| wss.as_ref().iter().map(|ws| ws.as_slice()))
+			.flatten()
+			.collect::<Vec<&[Uuid]>>();
+		write_sets.sort();
+		write_sets.dedup();
+		let mut write_set_index = HashMap::<&Uuid, Vec<usize>>::new();
+		for (i, write_set) in write_sets.iter().enumerate() {
+			for node in write_set.iter() {
+				write_set_index.entry(node).or_default().push(i);
 			}
 		}
 
-		let call_futures = call_list.drain().map(|(node, entries)| async move {
+		// Build a map of all nodes to the entries that must be sent to that node.
-			let rpc = TableRpc::<F>::Update(entries);
+		let mut call_list: HashMap<Uuid, Vec<_>> = HashMap::new();
+		for (write_sets, entry_enc) in entries_vec.iter() {
+			for write_set in write_sets.as_ref().iter() {
+				for node in write_set.iter() {
+					call_list.entry(*node).or_default().push(entry_enc.clone())
+				}
+			}
+		}
 
-			let resp = self
+		// Build futures to actually perform each of the corresponding RPC calls
-				.system
+		let call_count = call_list.len();
-				.rpc_helper()
+		let call_futures = call_list.into_iter().map(|(node, entries)| {
-				.call(
+			let this = self.clone();
-					&self.endpoint,
+			let tracer = opentelemetry::global::tracer("garage");
-					node,
+			let span = tracer.start(format!("RPC to {:?}", node));
-					rpc,
+			let fut = async move {
-					RequestStrategy::with_priority(PRIO_NORMAL),
+				let rpc = TableRpc::<F>::Update(entries);
-				)
+				let resp = this
-				.await?;
+					.system
-			Ok::<_, Error>((node, resp))
+					.rpc_helper()
+					.call(
+						&this.endpoint,
+						node,
+						rpc,
+						RequestStrategy::with_priority(PRIO_NORMAL).with_quorum(quorum),
+					)
+					.await;
+				(node, resp)
+			};
+			fut.with_context(Context::current_with_span(span))
 		});
+
+		// Run all requests in parallel thanks to FuturesUnordered, and collect results.
 		let mut resps = call_futures.collect::<FuturesUnordered<_>>();
+		let mut set_counters = vec![(0, 0); write_sets.len()];
+		let mut successes = 0;
 		let mut errors = vec![];
 
-		while let Some(resp) = resps.next().await {
+		while let Some((node, resp)) = resps.next().await {
-			if let Err(e) = resp {
+			match resp {
-				errors.push(e);
+				Ok(_) => {
+					successes += 1;
+					for set in write_set_index.get(&node).unwrap().iter() {
+						set_counters[*set].0 += 1;
+					}
+				}
+				Err(e) => {
+					errors.push(e);
+					for set in write_set_index.get(&node).unwrap().iter() {
+						set_counters[*set].1 += 1;
+					}
+				}
+			}
+
+			if set_counters.iter().all(|(ok_cnt, _)| *ok_cnt >= quorum) {
+				// Success
+
+				// Continue all other requests in background
+				tokio::spawn(async move {
+					resps.collect::<Vec<(Uuid, Result<_, _>)>>().await;
+				});
+
+				return Ok(());
+			}
+
+			if set_counters
+				.iter()
+				.enumerate()
+				.any(|(i, (_, err_cnt))| err_cnt + quorum > write_sets[i].len())
+			{
+				// Too many errors in this set, we know we won't get a quorum
+				break;
 			}
 		}
-		if errors.len() > self.data.replication.max_write_errors() {
+
-			Err(Error::Message("Too many errors".into()))
+		// Failure, could not get quorum within at least one set
-		} else {
+		let errors = errors.iter().map(|e| format!("{}", e)).collect::<Vec<_>>();
-			Ok(())
+		Err(Error::Quorum(
-		}
+			quorum,
+			Some(write_sets.len()),
+			successes,
+			call_count,
+			errors,
+		))
 	}
 
 	pub async fn get(