Request based availability distribution (#2423)

* WIP * availability distribution, still very wip. Work on the requesting side of things. * Some docs on what I intend to do. * Checkpoint of session cache implementation as I will likely replace it with something smarter. * More work, mostly on cache and getting things to type check. * Only derive MallocSizeOf and Debug for std. * availability-distribution: Cache feature complete. * Sketch out logic in `FetchTask` for actual fetching. - Compile fixes. - Cleanup. * Format cleanup. * More format fixes. * Almost feature complete `fetch_task`. Missing: - Check for cancel - Actual querying of peer ids. * Finish FetchTask so far. * Directly use AuthorityDiscoveryId in protocol and cache. * Resolve `AuthorityDiscoveryId` on sending requests. * Rework fetch_task - also make it impossible to check the wrong chunk index. - Export needed function in validator_discovery. * From<u32> implementation for `ValidatorIndex`. * Fixes and more integration work. * Make session cache proper lru cache. * Use proper lru cache. * Requester finished. * ProtocolState -> Requester Also make sure to not fetch our own chunk. * Cleanup + fixes. * Remove unused functions - FetchTask::is_finished - SessionCache::fetch_session_info * availability-distribution responding side. * Cleanup + Fixes. * More fixes. * More fixes. adder-collator is running! * Some docs. * Docs. * Fix reporting of bad guys. * Fix tests * Make all tests compile. * Fix test. * Cleanup + get rid of some warnings. * state -> requester * Mostly doc fixes. * Fix test suite. * Get rid of now redundant message types. * WIP * Rob's review remarks. * Fix test suite. * core.relay_parent -> leaf for session request. * Style fix. * Decrease request timeout. * Cleanup obsolete errors. * Metrics + don't fail on non fatal errors. * requester.rs -> requester/mod.rs * Panic on invalid BadValidator report. * Fix indentation. * Use typed default timeout constant. * Make channel size 0, as each sender gets one slot anyways. * Fix incorrect metrics initialization. * Fix build after merge. * More fixes. * Hopefully valid metrics names. * Better metrics names. * Some tests that already work. * Slightly better docs. * Some more tests. * Fix network bridge test.
2026-06-30 20:07:24 +00:00 · 2021-02-26 18:58:07 +01:00
parent 241b1f12a7
commit 48409e5548
45 changed files with 2037 additions and 1523 deletions
@@ -0,0 +1,275 @@
+// Copyright 2021 Parity Technologies (UK) Ltd.
+// This file is part of Polkadot.
+
+// Polkadot is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+
+// Polkadot is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License
+// along with Polkadot.  If not, see <http://www.gnu.org/licenses/>.
+
+use std::collections::HashSet;
+
+use lru::LruCache;
+use rand::{seq::SliceRandom, thread_rng};
+
+use sp_application_crypto::AppKey;
+use sp_core::crypto::Public;
+use sp_keystore::{CryptoStore, SyncCryptoStorePtr};
+
+use polkadot_node_subsystem_util::{
+	request_session_index_for_child_ctx, request_session_info_ctx,
+};
+use polkadot_primitives::v1::SessionInfo as GlobalSessionInfo;
+use polkadot_primitives::v1::{
+	AuthorityDiscoveryId, GroupIndex, Hash, SessionIndex, ValidatorId, ValidatorIndex,
+};
+use polkadot_subsystem::SubsystemContext;
+
+use super::{
+	error::{recv_runtime, Result},
+	Error,
+	LOG_TARGET,
+};
+
+/// Caching of session info as needed by availability distribution.
+///
+/// It should be ensured that a cached session stays live in the cache as long as we might need it.
+pub struct SessionCache {
+	/// Get the session index for a given relay parent.
+	///
+	/// We query this up to a 100 times per block, so caching it here without roundtrips over the
+	/// overseer seems sensible.
+	session_index_cache: LruCache<Hash, SessionIndex>,
+
+	/// Look up cached sessions by SessionIndex.
+	///
+	/// Note: Performance of fetching is really secondary here, but we need to ensure we are going
+	/// to get any existing cache entry, before fetching new information, as we should not mess up
+	/// the order of validators in `SessionInfo::validator_groups`. (We want live TCP connections
+	/// wherever possible.)
+	session_info_cache: LruCache<SessionIndex, SessionInfo>,
+
+	/// Key store for determining whether we are a validator and what `ValidatorIndex` we have.
+	keystore: SyncCryptoStorePtr,
+}
+
+/// Localized session information, tailored for the needs of availability distribution.
+#[derive(Clone)]
+pub struct SessionInfo {
+	/// The index of this session.
+	pub session_index: SessionIndex,
+
+	/// Validator groups of the current session.
+	///
+	/// Each group's order is randomized. This way we achieve load balancing when requesting
+	/// chunks, as the validators in a group will be tried in that randomized order. Each node
+	/// should arrive at a different order, therefore we distribute the load on individual
+	/// validators.
+	pub validator_groups: Vec<Vec<AuthorityDiscoveryId>>,
+
+	/// Information about ourself:
+	pub our_index: ValidatorIndex,
+
+	/// Remember to which group we belong, so we won't start fetching chunks for candidates with
+	/// our group being responsible. (We should have that chunk already.)
+	pub our_group: GroupIndex,
+}
+
+/// Report of bad validators.
+///
+/// Fetching tasks will report back validators that did not respond as expected, so we can re-order
+/// them.
+pub struct BadValidators {
+	/// The session index that was used.
+	pub session_index: SessionIndex,
+	/// The group, the not properly responding validators belong to.
+	pub group_index: GroupIndex,
+	/// The list of bad validators.
+	pub bad_validators: Vec<AuthorityDiscoveryId>,
+}
+
+impl SessionCache {
+	/// Create a new `SessionCache`.
+	pub fn new(keystore: SyncCryptoStorePtr) -> Self {
+		SessionCache {
+			// 5 relatively conservative, 1 to 2 should suffice:
+			session_index_cache: LruCache::new(5),
+			// We need to cache the current and the last session the most:
+			session_info_cache: LruCache::new(2),
+			keystore,
+		}
+	}
+
+	/// Tries to retrieve `SessionInfo` and calls `with_info` if successful.
+	///
+	/// If this node is not a validator, the function will return `None`.
+	///
+	/// Use this function over any `fetch_session_info` if all you need is a reference to
+	/// `SessionInfo`, as it avoids an expensive clone.
+	pub async fn with_session_info<Context, F, R>(
+		&mut self,
+		ctx: &mut Context,
+		parent: Hash,
+		with_info: F,
+	) -> Result<Option<R>>
+	where
+		Context: SubsystemContext,
+		F: FnOnce(&SessionInfo) -> R,
+	{
+		let session_index = match self.session_index_cache.get(&parent) {
+			Some(index) => *index,
+			None => {
+				let index =
+					recv_runtime(request_session_index_for_child_ctx(parent, ctx).await)
+						.await?;
+				self.session_index_cache.put(parent, index);
+				index
+			}
+		};
+
+		if let Some(info) = self.session_info_cache.get(&session_index) {
+			return Ok(Some(with_info(info)));
+		}
+
+		if let Some(info) = self
+			.query_info_from_runtime(ctx, parent, session_index)
+			.await?
+		{
+			let r = with_info(&info);
+			self.session_info_cache.put(session_index, info);
+			return Ok(Some(r));
+		}
+		Ok(None)
+	}
+
+	/// Variant of `report_bad` that never fails, but just logs errors.
+	///
+	/// Not being able to report bad validators is not fatal, so we should not shutdown the
+	/// subsystem on this.
+	pub fn report_bad_log(&mut self, report: BadValidators) {
+		if let Err(err) =  self.report_bad(report) {
+			tracing::warn!(
+				target: LOG_TARGET,
+				err= ?err,
+				"Reporting bad validators failed with error"
+			);
+		}
+	}
+
+	/// Make sure we try unresponsive or misbehaving validators last.
+	///
+	/// We assume validators in a group are tried in reverse order, so the reported bad validators
+	/// will be put at the beginning of the group.
+	#[tracing::instrument(level = "trace", skip(self, report), fields(subsystem = LOG_TARGET))]
+	pub fn report_bad(&mut self, report: BadValidators) -> Result<()> {
+		let session = self
+			.session_info_cache
+			.get_mut(&report.session_index)
+			.ok_or(Error::NoSuchCachedSession)?;
+		let group = session
+			.validator_groups
+			.get_mut(report.group_index.0 as usize)
+			.expect("A bad validator report must contain a valid group for the reported session. qed.");
+		let bad_set = report.bad_validators.iter().collect::<HashSet<_>>();
+
+		// Get rid of bad boys:
+		group.retain(|v| !bad_set.contains(v));
+
+		// We are trying validators in reverse order, so bad ones should be first:
+		let mut new_group = report.bad_validators;
+		new_group.append(group);
+		*group = new_group;
+		Ok(())
+	}
+
+	/// Query needed information from runtime.
+	///
+	/// We need to pass in the relay parent for our call to `request_session_info_ctx`. We should
+	/// actually don't need that: I suppose it is used for internal caching based on relay parents,
+	/// which we don't use here. It should not do any harm though.
+	async fn query_info_from_runtime<Context>(
+		&self,
+		ctx: &mut Context,
+		parent: Hash,
+		session_index: SessionIndex,
+	) -> Result<Option<SessionInfo>>
+	where
+		Context: SubsystemContext,
+	{
+		let GlobalSessionInfo {
+			validators,
+			discovery_keys,
+			mut validator_groups,
+			..
+		} = recv_runtime(request_session_info_ctx(parent, session_index, ctx).await)
+			.await?
+			.ok_or(Error::NoSuchSession(session_index))?;
+
+		if let Some(our_index) = self.get_our_index(validators).await {
+			// Get our group index:
+			let our_group = validator_groups
+				.iter()
+				.enumerate()
+				.find_map(|(i, g)| {
+					g.iter().find_map(|v| {
+						if *v == our_index {
+							Some(GroupIndex(i as u32))
+						} else {
+							None
+						}
+					})
+				})
+				.expect("Every validator should be in a validator group. qed.");
+
+			// Shuffle validators in groups:
+			let mut rng = thread_rng();
+			for g in validator_groups.iter_mut() {
+				g.shuffle(&mut rng)
+			}
+			// Look up `AuthorityDiscoveryId`s right away:
+			let validator_groups: Vec<Vec<_>> = validator_groups
+				.into_iter()
+				.map(|group| {
+					group
+						.into_iter()
+						.map(|index| {
+							discovery_keys.get(index.0 as usize)
+								.expect("There should be a discovery key for each validator of each validator group. qed.")
+								.clone()
+						})
+						.collect()
+				})
+				.collect();
+
+			let info = SessionInfo {
+				validator_groups,
+				our_index,
+				session_index,
+				our_group,
+			};
+			return Ok(Some(info));
+		}
+		return Ok(None);
+	}
+
+	/// Get our `ValidatorIndex`.
+	///
+	/// Returns: None if we are not a validator.
+	async fn get_our_index(&self, validators: Vec<ValidatorId>) -> Option<ValidatorIndex> {
+		for (i, v) in validators.iter().enumerate() {
+			if CryptoStore::has_keys(&*self.keystore, &[(v.to_raw_vec(), ValidatorId::ID)])
+				.await
+			{
+				return Some(ValidatorIndex(i as u32));
+			}
+		}
+		None
+	}
+}