-
Notifications
You must be signed in to change notification settings - Fork 781
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
[Merged by Bors] - Run fork choice before block proposal #3168
[Merged by Bors] - Run fork choice before block proposal #3168
Conversation
8a8eb08
to
a247b35
Compare
I've decided we need to move to the fork choice schedule that Prysm use: always run fork choice in the start of the slot, and then wait for that to finish (perhaps with a timeout) when proposing a block. As an optimisation we can also run fork choice preemptively in the state advance timer at 9s through the previous slot. |
Current TODO:
|
To establish that the impact of this change on merge networks is minimal, let's consider a few scenarios. We'll assume that the BN needs to propose a block in slot Case 1: Healthy chain, blocks every slot
Aside from the 2nd payload ID issue, I think this case looks good: we've added just 1 extra fcU relative to what we have currently. Case 2: Skip slot at
|
In the last week there was one instance of a warning from this branch being tripped on Prater, and it was on the notoriously flaky Singapore node while it was overwhelmed doing a bunch of other stuff:
Even so, the timeout was effective, and the block managed to become part of the canonical chain: https://prater.beaconcha.in/block/3017089 |
Based on out-of-band discussions with @paulhauner I've just pushed two commit 06908ad and 9808754 which move the preemptive fork choice run to 500ms before the start of the next slot, rather than 3s before. This has the following advantages:
The disadvantages are:
In commit 06908ad I also fixed an issue that occurs when the state advance runs over the slot during which it's started, which would cause it to skip the following slot and wait for the next one. I'm not strongly attached to this change and would be happy to revert the offset to its previous value which assumes that the state advance can run to completion before the end of the slot. |
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.
LGTM! 🚀
bors r+ |
## Issue Addressed Upcoming spec change ethereum/consensus-specs#2878 ## Proposed Changes 1. Run fork choice at the start of every slot, and wait for this run to complete before proposing a block. 2. As an optimisation, also run fork choice 3/4 of the way through the slot (at 9s), _dequeueing attestations for the next slot_. 3. Remove the fork choice run from the state advance timer that occurred before advancing the state. ## Additional Info ### Block Proposal Accuracy This change makes us more likely to propose on top of the correct head in the presence of re-orgs with proposer boost in play. The main scenario that this change is designed to address is described in the linked spec issue. ### Attestation Accuracy This change _also_ makes us more likely to attest to the correct head. Currently in the case of a skipped slot at `slot` we only run fork choice 9s into `slot - 1`. This means the attestations from `slot - 1` aren't taken into consideration, and any boost applied to the block from `slot - 1` is not removed (it should be). In the language of the linked spec issue, this means we are liable to attest to C, even when the majority voting weight has already caused a re-org to B. ### Why remove the call before the state advance? If we've run fork choice at the start of the slot then it has already dequeued all the attestations from the previous slot, which are the only ones eligible to influence the head in the current slot. Running fork choice again is unnecessary (unless we run it for the next slot and try to pre-empt a re-org, but I don't currently think this is a great idea). ### Performance Based on Prater testing this adds about 5-25ms of runtime to block proposal times, which are 500-1000ms on average (and spike to 5s+ sometimes due to state handling issues 😢 ). I believe this is a small enough penalty to enable it by default, with the option to disable it via the new flag `--fork-choice-before-proposal-timeout 0`. Upcoming work on block packing and state representation will also reduce block production times in general, while removing the spikes. ### Implementation Fork choice gets invoked at the start of the slot via the `per_slot_task` function called from the slot timer. It then uses a condition variable to signal to block production that fork choice has been updated. This is a bit funky, but it seems to work. One downside of the timer-based approach is that it doesn't happen automatically in most of the tests. The test added by this PR has to trigger the run manually.
Build failed (retrying...): |
## Issue Addressed Upcoming spec change ethereum/consensus-specs#2878 ## Proposed Changes 1. Run fork choice at the start of every slot, and wait for this run to complete before proposing a block. 2. As an optimisation, also run fork choice 3/4 of the way through the slot (at 9s), _dequeueing attestations for the next slot_. 3. Remove the fork choice run from the state advance timer that occurred before advancing the state. ## Additional Info ### Block Proposal Accuracy This change makes us more likely to propose on top of the correct head in the presence of re-orgs with proposer boost in play. The main scenario that this change is designed to address is described in the linked spec issue. ### Attestation Accuracy This change _also_ makes us more likely to attest to the correct head. Currently in the case of a skipped slot at `slot` we only run fork choice 9s into `slot - 1`. This means the attestations from `slot - 1` aren't taken into consideration, and any boost applied to the block from `slot - 1` is not removed (it should be). In the language of the linked spec issue, this means we are liable to attest to C, even when the majority voting weight has already caused a re-org to B. ### Why remove the call before the state advance? If we've run fork choice at the start of the slot then it has already dequeued all the attestations from the previous slot, which are the only ones eligible to influence the head in the current slot. Running fork choice again is unnecessary (unless we run it for the next slot and try to pre-empt a re-org, but I don't currently think this is a great idea). ### Performance Based on Prater testing this adds about 5-25ms of runtime to block proposal times, which are 500-1000ms on average (and spike to 5s+ sometimes due to state handling issues 😢 ). I believe this is a small enough penalty to enable it by default, with the option to disable it via the new flag `--fork-choice-before-proposal-timeout 0`. Upcoming work on block packing and state representation will also reduce block production times in general, while removing the spikes. ### Implementation Fork choice gets invoked at the start of the slot via the `per_slot_task` function called from the slot timer. It then uses a condition variable to signal to block production that fork choice has been updated. This is a bit funky, but it seems to work. One downside of the timer-based approach is that it doesn't happen automatically in most of the tests. The test added by this PR has to trigger the run manually.
bors r- gotta fix udeps |
Canceled. |
bors r+ |
## Issue Addressed Upcoming spec change ethereum/consensus-specs#2878 ## Proposed Changes 1. Run fork choice at the start of every slot, and wait for this run to complete before proposing a block. 2. As an optimisation, also run fork choice 3/4 of the way through the slot (at 9s), _dequeueing attestations for the next slot_. 3. Remove the fork choice run from the state advance timer that occurred before advancing the state. ## Additional Info ### Block Proposal Accuracy This change makes us more likely to propose on top of the correct head in the presence of re-orgs with proposer boost in play. The main scenario that this change is designed to address is described in the linked spec issue. ### Attestation Accuracy This change _also_ makes us more likely to attest to the correct head. Currently in the case of a skipped slot at `slot` we only run fork choice 9s into `slot - 1`. This means the attestations from `slot - 1` aren't taken into consideration, and any boost applied to the block from `slot - 1` is not removed (it should be). In the language of the linked spec issue, this means we are liable to attest to C, even when the majority voting weight has already caused a re-org to B. ### Why remove the call before the state advance? If we've run fork choice at the start of the slot then it has already dequeued all the attestations from the previous slot, which are the only ones eligible to influence the head in the current slot. Running fork choice again is unnecessary (unless we run it for the next slot and try to pre-empt a re-org, but I don't currently think this is a great idea). ### Performance Based on Prater testing this adds about 5-25ms of runtime to block proposal times, which are 500-1000ms on average (and spike to 5s+ sometimes due to state handling issues 😢 ). I believe this is a small enough penalty to enable it by default, with the option to disable it via the new flag `--fork-choice-before-proposal-timeout 0`. Upcoming work on block packing and state representation will also reduce block production times in general, while removing the spikes. ### Implementation Fork choice gets invoked at the start of the slot via the `per_slot_task` function called from the slot timer. It then uses a condition variable to signal to block production that fork choice has been updated. This is a bit funky, but it seems to work. One downside of the timer-based approach is that it doesn't happen automatically in most of the tests. The test added by this PR has to trigger the run manually.
*This PR was adapted from @pawanjay176's work in #3197.* ## Issue Addressed Fixes a regression in #3168 ## Proposed Changes #3168 added calls to `fork_choice` in `BeaconChain::per_slot_task` function. This leads to a panic as `per_slot_task` is called from an async context which calls fork choice, which then calls `block_on`. This PR changes the timer to call the `per_slot_task` function in a blocking thread. Co-authored-by: Pawan Dhananjay <[email protected]>
## Issue Addressed Fixes an issue that @paulhauner found with the v2.3.0 release candidate whereby the fork choice runs introduced by #3168 tripped over each other during sync: ``` May 24 23:06:40.542 WARN Error signalling fork choice waiter slot: 3884129, error: ForkChoiceSignalOutOfOrder { current: Slot(3884131), latest: Slot(3884129) }, service: beacon ``` This can occur because fork choice is called from the state advance _and_ the per-slot task. When one of these runs takes a long time it can end up finishing after a run from a later slot, tripping the error above. The problem is resolved by not running either of these fork choice calls during sync. Additionally, these parallel fork choice runs were causing issues in the database: ``` May 24 07:49:05.098 WARN Found a chain that should already have been pruned, head_slot: 92925, head_block_root: 0xa76c7bf1b98e54ed4b0d8686efcfdf853484e6c2a4c67e91cbf19e5ad1f96b17, service: beacon May 24 07:49:05.101 WARN Database migration failed error: HotColdDBError(FreezeSlotError { current_split_slot: Slot(92608), proposed_split_slot: Slot(92576) }), service: beacon ``` In this case, two fork choice calls triggering the finalization processing were being processed out of order due to differences in their processing time, causing the background migrator to try to advance finalization _backwards_ :flushed:. Removing the parallel fork choice runs from sync effectively addresses the issue, because these runs are most likely to have different finalized checkpoints (because of the speed at which fork choice advances during sync). In theory it's still possible to process updates out of order if any other fork choice runs end up completing out of order, but this should be much less common. Fixing out of order fork choice runs in general is difficult as it requires architectural changes like serialising fork choice updates through a single thread, or locking fork choice along with the head when it is mutated (#3175). ## Proposed Changes * Don't run per-slot fork choice during sync (if head is older than 4 slots) * Don't run state-advance fork choice during sync (if head is older than 4 slots) * Check for monotonic finalization updates in the background migrator. This is a good defensive check to have, and I'm not sure why we didn't have it before (we may have had it and wrongly removed it).
Issue Addressed
Upcoming spec change ethereum/consensus-specs#2878
Proposed Changes
Additional Info
Block Proposal Accuracy
This change makes us more likely to propose on top of the correct head in the presence of re-orgs with proposer boost in play. The main scenario that this change is designed to address is described in the linked spec issue.
Attestation Accuracy
This change also makes us more likely to attest to the correct head. Currently in the case of a skipped slot at
slot
we only run fork choice 9s intoslot - 1
. This means the attestations fromslot - 1
aren't taken into consideration, and any boost applied to the block fromslot - 1
is not removed (it should be). In the language of the linked spec issue, this means we are liable to attest to C, even when the majority voting weight has already caused a re-org to B.Why remove the call before the state advance?
If we've run fork choice at the start of the slot then it has already dequeued all the attestations from the previous slot, which are the only ones eligible to influence the head in the current slot. Running fork choice again is unnecessary (unless we run it for the next slot and try to pre-empt a re-org, but I don't currently think this is a great idea).
Performance
Based on Prater testing this adds about 5-25ms of runtime to block proposal times, which are 500-1000ms on average (and spike to 5s+ sometimes due to state handling issues 😢 ). I believe this is a small enough penalty to enable it by default, with the option to disable it via the new flag
--fork-choice-before-proposal-timeout 0
. Upcoming work on block packing and state representation will also reduce block production times in general, while removing the spikes.Implementation
Fork choice gets invoked at the start of the slot via the
per_slot_task
function called from the slot timer. It then uses a condition variable to signal to block production that fork choice has been updated. This is a bit funky, but it seems to work. One downside of the timer-based approach is that it doesn't happen automatically in most of the tests. The test added by this PR has to trigger the run manually.