-
Notifications
You must be signed in to change notification settings - Fork 680
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Fix kusama validators getting 0 backing rewards the first session they enter the active set #3722
Conversation
…y enter active set Signed-off-by: Alexandru Gheorghe <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nice find 👍
Not sure you need to test this on versi. However, the tests should be updated. Especially as there is already one authorities_returns_current_and_next_authority_set
;)
Co-authored-by: Bastian Köcher <[email protected]>
Signed-off-by: Alexandru Gheorghe <[email protected]>
Signed-off-by: Alexandru Gheorghe <[email protected]>
I agree the fix is pretty clear, I just want to make sure there aren't other places that I missed. We don't need to wait with merging this PR, until I get that setup working.
Done.
Did my best hopefully it is clearer now. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good! I think, although it would be good to test on Versi, it's not strictly required. By the nature of changes, the worst possible scenario is that we're updating something we shouldn't update with the data still correct, so it doesn't make any worse anyway.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
👍
If you want to test this, run a test network locally. Set |
@alexggh please add some small prdoc explaining the change. |
Does this also fix #1819? |
It does help, statement-v1 does not require direct connections to your peers in the validator group, so you might get lucky with the random gossiping when you are sparsely connected, but fully lucky, so I guess that's why we would see just some missed statements instead of missing all the statements as it happens with statements-v2. |
Yeah, that's what I ended-up doing, it is not that easy, because ValidatorManager changes the set every session, so it was masking the problem also you need some bigger waiting times and sessions because DHT is updated every 10 minutes. Managed to reproduce it with zombienet and a modified ValidatorManager and confirmed this patch is fixing the issue. |
Signed-off-by: Alexandru Gheorghe <[email protected]>
Done. |
…y enter the active set (#3722) There is a problem in the way we update `authorithy-discovery` next keys and because of that nodes that enter the active set would be noticed at the start of the session they become active, instead of the start of the previous session as it was intended. This is problematic because: 1. The node itself advertises its addresses on the DHT only when it notices it should become active on around ~10m loop, so in this case it would notice after it becomes active. 2. The other nodes won't be able to detect the new nodes addresses at the beginning of the session, so it won't added them to the reserved set. With 1 + 2, we end-up in a situation where the the new node won't be able to properly connect to its peers because it won't be in its peers reserved set. Now, the nodes accept by default`MIN_GOSSIP_PEERS: usize = 25` connections to nodes that are not in the reserved set, but given Kusama size(> 1000 nodes) you could easily have more than`25` new nodes entering the active set or simply the nodes don't have slots anymore because, they already have connections to peers not in the active set. In the end what the node would notice is 0 backing rewards because it wasn't directly connected to the peers in its backing group. ## Root-cause The flow is like this: 1. At BAD_SESSION - 1, in `rotate_session` new nodes are added to QueuedKeys https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L609 ``` <QueuedKeys<T>>::put(queued_amalgamated.clone()); <QueuedChanged<T>>::put(next_changed); ``` 2. AuthorityDiscovery::on_new_session is called with `changed` being the value of `<QueuedChanged<T>>:` at BAD_SESSION - **2** because it was saved before being updated https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L613 3. At BAD_SESSION - 1, `AuthorityDiscovery::on_new_session` doesn't updated its next_keys because `changed` was false. 4. For the entire durations of `BAD_SESSION - 1` everyone calling runtime api `authorities`(should return past, present and future authorities) won't discover the nodes that should become active . 5. At the beginning of BAD_SESSION, all nodes discover the new nodes are authorities, but it is already too late because reserved_nodes are updated only at the beginning of the session by the `gossip-support`. See above why this bad. ## Fix Update next keys with the queued_validators at every session, not matter the value of `changed` this is the same way babe pallet correctly does it. https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/babe/src/lib.rs#L655 ## Notes - The issue doesn't reproduce with proof-authorities changes like `versi` because `changed` would always be true and `AuthorityDiscovery` correctly updates its next_keys every time. - Confirmed at session `37651` on kusama that this is exactly what it happens by looking at blocks with polkadot.js. ## TODO - [ ] Move versi on proof of stake and properly test before and after fix to confirm there is no other issue. --------- Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]> (cherry picked from commit 8d0cd4f) Signed-off-by: Alexandru Gheorghe <[email protected]>
…y enter the active set (#3722) There is a problem in the way we update `authorithy-discovery` next keys and because of that nodes that enter the active set would be noticed at the start of the session they become active, instead of the start of the previous session as it was intended. This is problematic because: 1. The node itself advertises its addresses on the DHT only when it notices it should become active on around ~10m loop, so in this case it would notice after it becomes active. 2. The other nodes won't be able to detect the new nodes addresses at the beginning of the session, so it won't added them to the reserved set. With 1 + 2, we end-up in a situation where the the new node won't be able to properly connect to its peers because it won't be in its peers reserved set. Now, the nodes accept by default`MIN_GOSSIP_PEERS: usize = 25` connections to nodes that are not in the reserved set, but given Kusama size(> 1000 nodes) you could easily have more than`25` new nodes entering the active set or simply the nodes don't have slots anymore because, they already have connections to peers not in the active set. In the end what the node would notice is 0 backing rewards because it wasn't directly connected to the peers in its backing group. ## Root-cause The flow is like this: 1. At BAD_SESSION - 1, in `rotate_session` new nodes are added to QueuedKeys https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L609 ``` <QueuedKeys<T>>::put(queued_amalgamated.clone()); <QueuedChanged<T>>::put(next_changed); ``` 2. AuthorityDiscovery::on_new_session is called with `changed` being the value of `<QueuedChanged<T>>:` at BAD_SESSION - **2** because it was saved before being updated https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L613 3. At BAD_SESSION - 1, `AuthorityDiscovery::on_new_session` doesn't updated its next_keys because `changed` was false. 4. For the entire durations of `BAD_SESSION - 1` everyone calling runtime api `authorities`(should return past, present and future authorities) won't discover the nodes that should become active . 5. At the beginning of BAD_SESSION, all nodes discover the new nodes are authorities, but it is already too late because reserved_nodes are updated only at the beginning of the session by the `gossip-support`. See above why this bad. ## Fix Update next keys with the queued_validators at every session, not matter the value of `changed` this is the same way babe pallet correctly does it. https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/babe/src/lib.rs#L655 ## Notes - The issue doesn't reproduce with proof-authorities changes like `versi` because `changed` would always be true and `AuthorityDiscovery` correctly updates its next_keys every time. - Confirmed at session `37651` on kusama that this is exactly what it happens by looking at blocks with polkadot.js. ## TODO - [ ] Move versi on proof of stake and properly test before and after fix to confirm there is no other issue. --------- Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]> (cherry picked from commit 8d0cd4f) Signed-off-by: Alexandru Gheorghe <[email protected]>
…y enter the active set (#3722) There is a problem in the way we update `authorithy-discovery` next keys and because of that nodes that enter the active set would be noticed at the start of the session they become active, instead of the start of the previous session as it was intended. This is problematic because: 1. The node itself advertises its addresses on the DHT only when it notices it should become active on around ~10m loop, so in this case it would notice after it becomes active. 2. The other nodes won't be able to detect the new nodes addresses at the beginning of the session, so it won't added them to the reserved set. With 1 + 2, we end-up in a situation where the the new node won't be able to properly connect to its peers because it won't be in its peers reserved set. Now, the nodes accept by default`MIN_GOSSIP_PEERS: usize = 25` connections to nodes that are not in the reserved set, but given Kusama size(> 1000 nodes) you could easily have more than`25` new nodes entering the active set or simply the nodes don't have slots anymore because, they already have connections to peers not in the active set. In the end what the node would notice is 0 backing rewards because it wasn't directly connected to the peers in its backing group. ## Root-cause The flow is like this: 1. At BAD_SESSION - 1, in `rotate_session` new nodes are added to QueuedKeys https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L609 ``` <QueuedKeys<T>>::put(queued_amalgamated.clone()); <QueuedChanged<T>>::put(next_changed); ``` 2. AuthorityDiscovery::on_new_session is called with `changed` being the value of `<QueuedChanged<T>>:` at BAD_SESSION - **2** because it was saved before being updated https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L613 3. At BAD_SESSION - 1, `AuthorityDiscovery::on_new_session` doesn't updated its next_keys because `changed` was false. 4. For the entire durations of `BAD_SESSION - 1` everyone calling runtime api `authorities`(should return past, present and future authorities) won't discover the nodes that should become active . 5. At the beginning of BAD_SESSION, all nodes discover the new nodes are authorities, but it is already too late because reserved_nodes are updated only at the beginning of the session by the `gossip-support`. See above why this bad. ## Fix Update next keys with the queued_validators at every session, not matter the value of `changed` this is the same way babe pallet correctly does it. https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/babe/src/lib.rs#L655 ## Notes - The issue doesn't reproduce with proof-authorities changes like `versi` because `changed` would always be true and `AuthorityDiscovery` correctly updates its next_keys every time. - Confirmed at session `37651` on kusama that this is exactly what it happens by looking at blocks with polkadot.js. ## TODO - [ ] Move versi on proof of stake and properly test before and after fix to confirm there is no other issue. --------- Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]> (cherry picked from commit 8d0cd4f) Signed-off-by: Alexandru Gheorghe <[email protected]>
…y enter the active set (#3722) There is a problem in the way we update `authorithy-discovery` next keys and because of that nodes that enter the active set would be noticed at the start of the session they become active, instead of the start of the previous session as it was intended. This is problematic because: 1. The node itself advertises its addresses on the DHT only when it notices it should become active on around ~10m loop, so in this case it would notice after it becomes active. 2. The other nodes won't be able to detect the new nodes addresses at the beginning of the session, so it won't added them to the reserved set. With 1 + 2, we end-up in a situation where the the new node won't be able to properly connect to its peers because it won't be in its peers reserved set. Now, the nodes accept by default`MIN_GOSSIP_PEERS: usize = 25` connections to nodes that are not in the reserved set, but given Kusama size(> 1000 nodes) you could easily have more than`25` new nodes entering the active set or simply the nodes don't have slots anymore because, they already have connections to peers not in the active set. In the end what the node would notice is 0 backing rewards because it wasn't directly connected to the peers in its backing group. ## Root-cause The flow is like this: 1. At BAD_SESSION - 1, in `rotate_session` new nodes are added to QueuedKeys https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L609 ``` <QueuedKeys<T>>::put(queued_amalgamated.clone()); <QueuedChanged<T>>::put(next_changed); ``` 2. AuthorityDiscovery::on_new_session is called with `changed` being the value of `<QueuedChanged<T>>:` at BAD_SESSION - **2** because it was saved before being updated https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L613 3. At BAD_SESSION - 1, `AuthorityDiscovery::on_new_session` doesn't updated its next_keys because `changed` was false. 4. For the entire durations of `BAD_SESSION - 1` everyone calling runtime api `authorities`(should return past, present and future authorities) won't discover the nodes that should become active . 5. At the beginning of BAD_SESSION, all nodes discover the new nodes are authorities, but it is already too late because reserved_nodes are updated only at the beginning of the session by the `gossip-support`. See above why this bad. ## Fix Update next keys with the queued_validators at every session, not matter the value of `changed` this is the same way babe pallet correctly does it. https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/babe/src/lib.rs#L655 ## Notes - The issue doesn't reproduce with proof-authorities changes like `versi` because `changed` would always be true and `AuthorityDiscovery` correctly updates its next_keys every time. - Confirmed at session `37651` on kusama that this is exactly what it happens by looking at blocks with polkadot.js. ## TODO - [ ] Move versi on proof of stake and properly test before and after fix to confirm there is no other issue. --------- Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]> (cherry picked from commit 8d0cd4f) Signed-off-by: Alexandru Gheorghe <[email protected]>
…y enter the active set (#3722) There is a problem in the way we update `authorithy-discovery` next keys and because of that nodes that enter the active set would be noticed at the start of the session they become active, instead of the start of the previous session as it was intended. This is problematic because: 1. The node itself advertises its addresses on the DHT only when it notices it should become active on around ~10m loop, so in this case it would notice after it becomes active. 2. The other nodes won't be able to detect the new nodes addresses at the beginning of the session, so it won't added them to the reserved set. With 1 + 2, we end-up in a situation where the the new node won't be able to properly connect to its peers because it won't be in its peers reserved set. Now, the nodes accept by default`MIN_GOSSIP_PEERS: usize = 25` connections to nodes that are not in the reserved set, but given Kusama size(> 1000 nodes) you could easily have more than`25` new nodes entering the active set or simply the nodes don't have slots anymore because, they already have connections to peers not in the active set. In the end what the node would notice is 0 backing rewards because it wasn't directly connected to the peers in its backing group. ## Root-cause The flow is like this: 1. At BAD_SESSION - 1, in `rotate_session` new nodes are added to QueuedKeys https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L609 ``` <QueuedKeys<T>>::put(queued_amalgamated.clone()); <QueuedChanged<T>>::put(next_changed); ``` 2. AuthorityDiscovery::on_new_session is called with `changed` being the value of `<QueuedChanged<T>>:` at BAD_SESSION - **2** because it was saved before being updated https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L613 3. At BAD_SESSION - 1, `AuthorityDiscovery::on_new_session` doesn't updated its next_keys because `changed` was false. 4. For the entire durations of `BAD_SESSION - 1` everyone calling runtime api `authorities`(should return past, present and future authorities) won't discover the nodes that should become active . 5. At the beginning of BAD_SESSION, all nodes discover the new nodes are authorities, but it is already too late because reserved_nodes are updated only at the beginning of the session by the `gossip-support`. See above why this bad. ## Fix Update next keys with the queued_validators at every session, not matter the value of `changed` this is the same way babe pallet correctly does it. https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/babe/src/lib.rs#L655 ## Notes - The issue doesn't reproduce with proof-authorities changes like `versi` because `changed` would always be true and `AuthorityDiscovery` correctly updates its next_keys every time. - Confirmed at session `37651` on kusama that this is exactly what it happens by looking at blocks with polkadot.js. ## TODO - [ ] Move versi on proof of stake and properly test before and after fix to confirm there is no other issue. --------- Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]> (cherry picked from commit 8d0cd4f) Signed-off-by: Alexandru Gheorghe <[email protected]>
… enter the active set (#3735) This is a cherry-pick from master of #3722 Expected patches for (1.7.0): `pallet-authorithy-discovery` Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]>
… enter the active set (#3733) This is a cherry-pick from master of #3722 Expected patches for (1.7.0): `pallet-authorithy-discovery` Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]>
… enter the active set (#3735) This is a cherry-pick from master of #3722 Expected patches for (1.7.0): `pallet-authorithy-discovery` Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]>
…y enter the active set (paritytech#3722) There is a problem in the way we update `authorithy-discovery` next keys and because of that nodes that enter the active set would be noticed at the start of the session they become active, instead of the start of the previous session as it was intended. This is problematic because: 1. The node itself advertises its addresses on the DHT only when it notices it should become active on around ~10m loop, so in this case it would notice after it becomes active. 2. The other nodes won't be able to detect the new nodes addresses at the beginning of the session, so it won't added them to the reserved set. With 1 + 2, we end-up in a situation where the the new node won't be able to properly connect to its peers because it won't be in its peers reserved set. Now, the nodes accept by default`MIN_GOSSIP_PEERS: usize = 25` connections to nodes that are not in the reserved set, but given Kusama size(> 1000 nodes) you could easily have more than`25` new nodes entering the active set or simply the nodes don't have slots anymore because, they already have connections to peers not in the active set. In the end what the node would notice is 0 backing rewards because it wasn't directly connected to the peers in its backing group. ## Root-cause The flow is like this: 1. At BAD_SESSION - 1, in `rotate_session` new nodes are added to QueuedKeys https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L609 ``` <QueuedKeys<T>>::put(queued_amalgamated.clone()); <QueuedChanged<T>>::put(next_changed); ``` 2. AuthorityDiscovery::on_new_session is called with `changed` being the value of `<QueuedChanged<T>>:` at BAD_SESSION - **2** because it was saved before being updated https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/session/src/lib.rs#L613 3. At BAD_SESSION - 1, `AuthorityDiscovery::on_new_session` doesn't updated its next_keys because `changed` was false. 4. For the entire durations of `BAD_SESSION - 1` everyone calling runtime api `authorities`(should return past, present and future authorities) won't discover the nodes that should become active . 5. At the beginning of BAD_SESSION, all nodes discover the new nodes are authorities, but it is already too late because reserved_nodes are updated only at the beginning of the session by the `gossip-support`. See above why this bad. ## Fix Update next keys with the queued_validators at every session, not matter the value of `changed` this is the same way babe pallet correctly does it. https://github.com/paritytech/polkadot-sdk/blob/02e1a7f476d7d7c67153e975ab9a1bdc02ffea12/substrate/frame/babe/src/lib.rs#L655 ## Notes - The issue doesn't reproduce with proof-authorities changes like `versi` because `changed` would always be true and `AuthorityDiscovery` correctly updates its next_keys every time. - Confirmed at session `37651` on kusama that this is exactly what it happens by looking at blocks with polkadot.js. ## TODO - [ ] Move versi on proof of stake and properly test before and after fix to confirm there is no other issue. --------- Signed-off-by: Alexandru Gheorghe <[email protected]> Co-authored-by: Bastian Köcher <[email protected]>
There is a problem in the way we update
authorithy-discovery
next keys and because of that nodes that enter the active set would be noticed at the start of the session they become active, instead of the start of the previous session as it was intended. This is problematic because:With 1 + 2, we end-up in a situation where the the new node won't be able to properly connect to its peers because it won't be in its peers reserved set. Now, the nodes accept by default
MIN_GOSSIP_PEERS: usize = 25
connections to nodes that are not in the reserved set, but given Kusama size(> 1000 nodes) you could easily have more than25
new nodes entering the active set or simply the nodes don't have slots anymore because, they already have connections to peers not in the active set.In the end what the node would notice is 0 backing rewards because it wasn't directly connected to the peers in its backing group.
Root-cause
The flow is like this:
rotate_session
new nodes are added to QueuedKeyspolkadot-sdk/substrate/frame/session/src/lib.rs
Line 609 in 02e1a7f
changed
being the value of<QueuedChanged<T>>:
at BAD_SESSION - 2 because it was saved before being updatedpolkadot-sdk/substrate/frame/session/src/lib.rs
Line 613 in 02e1a7f
AuthorityDiscovery::on_new_session
doesn't updated its next_keys becausechanged
was false.BAD_SESSION - 1
everyone calling runtime apiauthorities
(should return past, present and future authorities) won't discover the nodes that should become active .gossip-support
. See above why this bad.Fix
Update next keys with the queued_validators at every session, not matter the value of
changed
this is the same way babe pallet correctly does it.polkadot-sdk/substrate/frame/babe/src/lib.rs
Line 655 in 02e1a7f
Notes
versi
becausechanged
would always be true andAuthorityDiscovery
correctly updates its next_keys every time.37651
on kusama that this is exactly what it happens by looking at blocks with polkadot.js.TODO