Reduce network bandwidth, improve parablock times: optimize approval-distribution (#5164)

* gossip-support: be explicit about dimensions * some guide updates * update network-bridge to distinguish x and y dimensions * get everything to compile * beginnings * some TODOs * polkadot runtime: use relevant_authorities * make gossip topologies per-session * better formatting * gossip support: use current session validators * expand in comment * adjust tests and fix index bug * add past/present/future connection test and clean up code * fmt * network bridge: updated types * update protocols to new gossip topology message * guide updates * add session to BlockApprovalMeta * add session to block info * refactor knowledge and remove most unify logic * start replacing gossip_peers with new SessionTopologies * add routing information to message state * add some utilities to SessionTopology * implement new gossip topology logic * re-implement unify_with_peer * distribute assignments according to topology * finish grid topology implementation * refactor network bridge slightly * issue connection requests on all past/present/future * fmt * address grumbles * tighten invariants in unify_with_peer * implement random propagation * refactor: extract required routing adjustment logic * some block-age logic * aggressively propagate messages when finality is slow * overhaul aggression system to have 3 levels * add aggression metrics * remove aggression L3 * reduce random circulation * remove PeerData * get approval tests compiling * use btree_map in known_by to make deterministic * Revert "use btree_map in known_by to make deterministic" This reverts commit 330d65343a7bb6fe4dd0f24bd8dbc15c0cbdbd9d. * test XY grid propagation * remove stray println * test unshared dimension propagation * add random gossip check * test unify_with_peer better * test sending after getting gossip topology * test L1 aggression on originator * test L1 aggression for non-originators * test non-originator aggression L2 * fnt * ~spellcheck * fix statement-distribution tests * fix flaky test * fix metrics typo * re-send periodically * test resending * typo Co-authored-by: Bernhard Schuster <bernhard@ahoi.io> * add more metrics about apd messages * add back unify_with_peer logs * make Resend an enum * be more explicit when resending * fmt * fix error * add a TODO for refactoring * remove debug metrics * add some guide stuff * fmt * update runtime API in test-runtim Co-authored-by: Bernhard Schuster <bernhard@ahoi.io>
2026-04-30 11:57:56 +00:00 · 2022-04-19 13:26:55 -05:00
parent edfa24bbc5
commit 79ecc53801
25 changed files with 2563 additions and 499 deletions
@@ -22,6 +22,16 @@ For assignments, what we need to be checking is whether we are aware of the (blo

 However, awareness on its own of a (block, candidate) pair would imply that even ancient candidates all the way back to the genesis are relevant. We are actually not interested in anything before finality.

+We gossip assignments along a grid topology produced by the [Gossip Support Subsystem](../utility/gossip-support.md) and also to a few random peers. The first time we accept an assignment or approval, regardless of the source, which originates from a validator peer in a shared dimension of the grid, we propagate the message to validator peers in the unshared dimension as well as a few random peers.
+
+But, in case these mechanisms don't work on their own, we need to trade bandwidth for protocol liveness by introducing aggression.
+
+Aggression has 3 levels:
+    Aggression Level 0: The basic behaviors described above.
+    Aggression Level 1: The originator of a message sends to all peers. Other peers follow the rules above.
+    Aggression Level 2: All peers send all messages to all their row and column neighbors. This means that each validator will, on average, receive each message approximately 2*sqrt(n) times.
+
+These aggression levels are chosen based on how long a block has taken to finalize: assignments and approvals related to the unfinalized block will be propagated with more aggression. In particular, it's only the earliest unfinalized blocks that aggression should be applied to, because descendants may be unfinalized only by virtue of being descendants.

 ## Protocol