lighthouse/beacon_node/http_api/tests/interactive_tests.rs

136 lines
4.6 KiB
Rust
Raw Normal View History

//! Generic tests that make use of the (newer) `InteractiveApiTester`
use crate::common::*;
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
use beacon_chain::test_utils::{AttestationStrategy, BlockStrategy};
use eth2::types::DepositContractData;
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
use tree_hash::TreeHash;
use types::{EthSpec, FullPayload, MainnetEthSpec, Slot};
type E = MainnetEthSpec;
// Test that the deposit_contract endpoint returns the correct chain_id and address.
// Regression test for https://github.com/sigp/lighthouse/issues/2657
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn deposit_contract_custom_network() {
let validator_count = 24;
let mut spec = E::default_spec();
// Rinkeby, which we don't use elsewhere.
spec.deposit_chain_id = 4;
spec.deposit_network_id = 4;
// Arbitrary contract address.
spec.deposit_contract_address = "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa".parse().unwrap();
let tester = InteractiveTester::<E>::new(Some(spec.clone()), validator_count).await;
let client = &tester.client;
let result = client.get_config_deposit_contract().await.unwrap().data;
let expected = DepositContractData {
address: spec.deposit_contract_address,
chain_id: spec.deposit_chain_id,
};
assert_eq!(result, expected);
}
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
// Test that running fork choice before proposing results in selection of the correct head.
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
pub async fn fork_choice_before_proposal() {
// Validator count needs to be at least 32 or proposer boost gets set to 0 when computing
// `validator_count // 32`.
let validator_count = 32;
let all_validators = (0..validator_count).collect::<Vec<_>>();
let num_initial: u64 = 31;
let tester = InteractiveTester::<E>::new(None, validator_count).await;
let harness = &tester.harness;
// Create some chain depth.
harness.advance_slot();
Use async code when interacting with EL (#3244) ## Overview This rather extensive PR achieves two primary goals: 1. Uses the finalized/justified checkpoints of fork choice (FC), rather than that of the head state. 2. Refactors fork choice, block production and block processing to `async` functions. Additionally, it achieves: - Concurrent forkchoice updates to the EL and cache pruning after a new head is selected. - Concurrent "block packing" (attestations, etc) and execution payload retrieval during block production. - Concurrent per-block-processing and execution payload verification during block processing. - The `Arc`-ification of `SignedBeaconBlock` during block processing (it's never mutated, so why not?): - I had to do this to deal with sending blocks into spawned tasks. - Previously we were cloning the beacon block at least 2 times during each block processing, these clones are either removed or turned into cheaper `Arc` clones. - We were also `Box`-ing and un-`Box`-ing beacon blocks as they moved throughout the networking crate. This is not a big deal, but it's nice to avoid shifting things between the stack and heap. - Avoids cloning *all the blocks* in *every chain segment* during sync. - It also has the potential to clean up our code where we need to pass an *owned* block around so we can send it back in the case of an error (I didn't do much of this, my PR is already big enough :sweat_smile:) - The `BeaconChain::HeadSafetyStatus` struct was removed. It was an old relic from prior merge specs. For motivation for this change, see https://github.com/sigp/lighthouse/pull/3244#issuecomment-1160963273 ## Changes to `canonical_head` and `fork_choice` Previously, the `BeaconChain` had two separate fields: ``` canonical_head: RwLock<Snapshot>, fork_choice: RwLock<BeaconForkChoice> ``` Now, we have grouped these values under a single struct: ``` canonical_head: CanonicalHead { cached_head: RwLock<Arc<Snapshot>>, fork_choice: RwLock<BeaconForkChoice> } ``` Apart from ergonomics, the only *actual* change here is wrapping the canonical head snapshot in an `Arc`. This means that we no longer need to hold the `cached_head` (`canonical_head`, in old terms) lock when we want to pull some values from it. This was done to avoid deadlock risks by preventing functions from acquiring (and holding) the `cached_head` and `fork_choice` locks simultaneously. ## Breaking Changes ### The `state` (root) field in the `finalized_checkpoint` SSE event Consider the scenario where epoch `n` is just finalized, but `start_slot(n)` is skipped. There are two state roots we might in the `finalized_checkpoint` SSE event: 1. The state root of the finalized block, which is `get_block(finalized_checkpoint.root).state_root`. 4. The state root at slot of `start_slot(n)`, which would be the state from (1), but "skipped forward" through any skip slots. Previously, Lighthouse would choose (2). However, we can see that when [Teku generates that event](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/beaconrestapi/src/main/java/tech/pegasys/teku/beaconrestapi/handlers/v1/events/EventSubscriptionManager.java#L171-L182) it uses [`getStateRootFromBlockRoot`](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/provider/src/main/java/tech/pegasys/teku/api/ChainDataProvider.java#L336-L341) which uses (1). I have switched Lighthouse from (2) to (1). I think it's a somewhat arbitrary choice between the two, where (1) is easier to compute and is consistent with Teku. ## Notes for Reviewers I've renamed `BeaconChain::fork_choice` to `BeaconChain::recompute_head`. Doing this helped ensure I broke all previous uses of fork choice and I also find it more descriptive. It describes an action and can't be confused with trying to get a reference to the `ForkChoice` struct. I've changed the ordering of SSE events when a block is received. It used to be `[block, finalized, head]` and now it's `[block, head, finalized]`. It was easier this way and I don't think we were making any promises about SSE event ordering so it's not "breaking". I've made it so fork choice will run when it's first constructed. I did this because I wanted to have a cached version of the last call to `get_head`. Ensuring `get_head` has been run *at least once* means that the cached values doesn't need to wrapped in an `Option`. This was fairly simple, it just involved passing a `slot` to the constructor so it knows *when* it's being run. When loading a fork choice from the store and a slot clock isn't handy I've just used the `slot` that was saved in the `fork_choice_store`. That seems like it would be a faithful representation of the slot when we saved it. I added the `genesis_time: u64` to the `BeaconChain`. It's small, constant and nice to have around. Since we're using FC for the fin/just checkpoints, we no longer get the `0x00..00` roots at genesis. You can see I had to remove a work-around in `ef-tests` here: b56be3bc2. I can't find any reason why this would be an issue, if anything I think it'll be better since the genesis-alias has caught us out a few times (0x00..00 isn't actually a real root). Edit: I did find a case where the `network` expected the 0x00..00 alias and patched it here: 3f26ac3e2. You'll notice a lot of changes in tests. Generally, tests should be functionally equivalent. Here are the things creating the most diff-noise in tests: - Changing tests to be `tokio::async` tests. - Adding `.await` to fork choice, block processing and block production functions. - Refactor of the `canonical_head` "API" provided by the `BeaconChain`. E.g., `chain.canonical_head.cached_head()` instead of `chain.canonical_head.read()`. - Wrapping `SignedBeaconBlock` in an `Arc`. - In the `beacon_chain/tests/block_verification`, we can't use the `lazy_static` `CHAIN_SEGMENT` variable anymore since it's generated with an async function. We just generate it in each test, not so efficient but hopefully insignificant. I had to disable `rayon` concurrent tests in the `fork_choice` tests. This is because the use of `rayon` and `block_on` was causing a panic. Co-authored-by: Mac L <mjladson@pm.me>
2022-07-03 05:36:50 +00:00
harness
.extend_chain(
num_initial as usize,
BlockStrategy::OnCanonicalHead,
AttestationStrategy::AllValidators,
)
.await;
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
// We set up the following block graph, where B is a block that is temporarily orphaned by C,
// but is then reinstated and built upon by D.
//
// A | B | - | D |
// ^ | - | C |
let slot_a = Slot::new(num_initial);
let slot_b = slot_a + 1;
let slot_c = slot_a + 2;
let slot_d = slot_a + 3;
let state_a = harness.get_current_state();
Use async code when interacting with EL (#3244) ## Overview This rather extensive PR achieves two primary goals: 1. Uses the finalized/justified checkpoints of fork choice (FC), rather than that of the head state. 2. Refactors fork choice, block production and block processing to `async` functions. Additionally, it achieves: - Concurrent forkchoice updates to the EL and cache pruning after a new head is selected. - Concurrent "block packing" (attestations, etc) and execution payload retrieval during block production. - Concurrent per-block-processing and execution payload verification during block processing. - The `Arc`-ification of `SignedBeaconBlock` during block processing (it's never mutated, so why not?): - I had to do this to deal with sending blocks into spawned tasks. - Previously we were cloning the beacon block at least 2 times during each block processing, these clones are either removed or turned into cheaper `Arc` clones. - We were also `Box`-ing and un-`Box`-ing beacon blocks as they moved throughout the networking crate. This is not a big deal, but it's nice to avoid shifting things between the stack and heap. - Avoids cloning *all the blocks* in *every chain segment* during sync. - It also has the potential to clean up our code where we need to pass an *owned* block around so we can send it back in the case of an error (I didn't do much of this, my PR is already big enough :sweat_smile:) - The `BeaconChain::HeadSafetyStatus` struct was removed. It was an old relic from prior merge specs. For motivation for this change, see https://github.com/sigp/lighthouse/pull/3244#issuecomment-1160963273 ## Changes to `canonical_head` and `fork_choice` Previously, the `BeaconChain` had two separate fields: ``` canonical_head: RwLock<Snapshot>, fork_choice: RwLock<BeaconForkChoice> ``` Now, we have grouped these values under a single struct: ``` canonical_head: CanonicalHead { cached_head: RwLock<Arc<Snapshot>>, fork_choice: RwLock<BeaconForkChoice> } ``` Apart from ergonomics, the only *actual* change here is wrapping the canonical head snapshot in an `Arc`. This means that we no longer need to hold the `cached_head` (`canonical_head`, in old terms) lock when we want to pull some values from it. This was done to avoid deadlock risks by preventing functions from acquiring (and holding) the `cached_head` and `fork_choice` locks simultaneously. ## Breaking Changes ### The `state` (root) field in the `finalized_checkpoint` SSE event Consider the scenario where epoch `n` is just finalized, but `start_slot(n)` is skipped. There are two state roots we might in the `finalized_checkpoint` SSE event: 1. The state root of the finalized block, which is `get_block(finalized_checkpoint.root).state_root`. 4. The state root at slot of `start_slot(n)`, which would be the state from (1), but "skipped forward" through any skip slots. Previously, Lighthouse would choose (2). However, we can see that when [Teku generates that event](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/beaconrestapi/src/main/java/tech/pegasys/teku/beaconrestapi/handlers/v1/events/EventSubscriptionManager.java#L171-L182) it uses [`getStateRootFromBlockRoot`](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/provider/src/main/java/tech/pegasys/teku/api/ChainDataProvider.java#L336-L341) which uses (1). I have switched Lighthouse from (2) to (1). I think it's a somewhat arbitrary choice between the two, where (1) is easier to compute and is consistent with Teku. ## Notes for Reviewers I've renamed `BeaconChain::fork_choice` to `BeaconChain::recompute_head`. Doing this helped ensure I broke all previous uses of fork choice and I also find it more descriptive. It describes an action and can't be confused with trying to get a reference to the `ForkChoice` struct. I've changed the ordering of SSE events when a block is received. It used to be `[block, finalized, head]` and now it's `[block, head, finalized]`. It was easier this way and I don't think we were making any promises about SSE event ordering so it's not "breaking". I've made it so fork choice will run when it's first constructed. I did this because I wanted to have a cached version of the last call to `get_head`. Ensuring `get_head` has been run *at least once* means that the cached values doesn't need to wrapped in an `Option`. This was fairly simple, it just involved passing a `slot` to the constructor so it knows *when* it's being run. When loading a fork choice from the store and a slot clock isn't handy I've just used the `slot` that was saved in the `fork_choice_store`. That seems like it would be a faithful representation of the slot when we saved it. I added the `genesis_time: u64` to the `BeaconChain`. It's small, constant and nice to have around. Since we're using FC for the fin/just checkpoints, we no longer get the `0x00..00` roots at genesis. You can see I had to remove a work-around in `ef-tests` here: b56be3bc2. I can't find any reason why this would be an issue, if anything I think it'll be better since the genesis-alias has caught us out a few times (0x00..00 isn't actually a real root). Edit: I did find a case where the `network` expected the 0x00..00 alias and patched it here: 3f26ac3e2. You'll notice a lot of changes in tests. Generally, tests should be functionally equivalent. Here are the things creating the most diff-noise in tests: - Changing tests to be `tokio::async` tests. - Adding `.await` to fork choice, block processing and block production functions. - Refactor of the `canonical_head` "API" provided by the `BeaconChain`. E.g., `chain.canonical_head.cached_head()` instead of `chain.canonical_head.read()`. - Wrapping `SignedBeaconBlock` in an `Arc`. - In the `beacon_chain/tests/block_verification`, we can't use the `lazy_static` `CHAIN_SEGMENT` variable anymore since it's generated with an async function. We just generate it in each test, not so efficient but hopefully insignificant. I had to disable `rayon` concurrent tests in the `fork_choice` tests. This is because the use of `rayon` and `block_on` was causing a panic. Co-authored-by: Mac L <mjladson@pm.me>
2022-07-03 05:36:50 +00:00
let (block_b, state_b) = harness.make_block(state_a.clone(), slot_b).await;
let block_root_b = harness
.process_block(slot_b, block_b.canonical_root(), block_b)
.await
.unwrap();
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
// Create attestations to B but keep them in reserve until after C has been processed.
let attestations_b = harness.make_attestations(
&all_validators,
&state_b,
state_b.tree_hash_root(),
block_root_b,
slot_b,
);
Use async code when interacting with EL (#3244) ## Overview This rather extensive PR achieves two primary goals: 1. Uses the finalized/justified checkpoints of fork choice (FC), rather than that of the head state. 2. Refactors fork choice, block production and block processing to `async` functions. Additionally, it achieves: - Concurrent forkchoice updates to the EL and cache pruning after a new head is selected. - Concurrent "block packing" (attestations, etc) and execution payload retrieval during block production. - Concurrent per-block-processing and execution payload verification during block processing. - The `Arc`-ification of `SignedBeaconBlock` during block processing (it's never mutated, so why not?): - I had to do this to deal with sending blocks into spawned tasks. - Previously we were cloning the beacon block at least 2 times during each block processing, these clones are either removed or turned into cheaper `Arc` clones. - We were also `Box`-ing and un-`Box`-ing beacon blocks as they moved throughout the networking crate. This is not a big deal, but it's nice to avoid shifting things between the stack and heap. - Avoids cloning *all the blocks* in *every chain segment* during sync. - It also has the potential to clean up our code where we need to pass an *owned* block around so we can send it back in the case of an error (I didn't do much of this, my PR is already big enough :sweat_smile:) - The `BeaconChain::HeadSafetyStatus` struct was removed. It was an old relic from prior merge specs. For motivation for this change, see https://github.com/sigp/lighthouse/pull/3244#issuecomment-1160963273 ## Changes to `canonical_head` and `fork_choice` Previously, the `BeaconChain` had two separate fields: ``` canonical_head: RwLock<Snapshot>, fork_choice: RwLock<BeaconForkChoice> ``` Now, we have grouped these values under a single struct: ``` canonical_head: CanonicalHead { cached_head: RwLock<Arc<Snapshot>>, fork_choice: RwLock<BeaconForkChoice> } ``` Apart from ergonomics, the only *actual* change here is wrapping the canonical head snapshot in an `Arc`. This means that we no longer need to hold the `cached_head` (`canonical_head`, in old terms) lock when we want to pull some values from it. This was done to avoid deadlock risks by preventing functions from acquiring (and holding) the `cached_head` and `fork_choice` locks simultaneously. ## Breaking Changes ### The `state` (root) field in the `finalized_checkpoint` SSE event Consider the scenario where epoch `n` is just finalized, but `start_slot(n)` is skipped. There are two state roots we might in the `finalized_checkpoint` SSE event: 1. The state root of the finalized block, which is `get_block(finalized_checkpoint.root).state_root`. 4. The state root at slot of `start_slot(n)`, which would be the state from (1), but "skipped forward" through any skip slots. Previously, Lighthouse would choose (2). However, we can see that when [Teku generates that event](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/beaconrestapi/src/main/java/tech/pegasys/teku/beaconrestapi/handlers/v1/events/EventSubscriptionManager.java#L171-L182) it uses [`getStateRootFromBlockRoot`](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/provider/src/main/java/tech/pegasys/teku/api/ChainDataProvider.java#L336-L341) which uses (1). I have switched Lighthouse from (2) to (1). I think it's a somewhat arbitrary choice between the two, where (1) is easier to compute and is consistent with Teku. ## Notes for Reviewers I've renamed `BeaconChain::fork_choice` to `BeaconChain::recompute_head`. Doing this helped ensure I broke all previous uses of fork choice and I also find it more descriptive. It describes an action and can't be confused with trying to get a reference to the `ForkChoice` struct. I've changed the ordering of SSE events when a block is received. It used to be `[block, finalized, head]` and now it's `[block, head, finalized]`. It was easier this way and I don't think we were making any promises about SSE event ordering so it's not "breaking". I've made it so fork choice will run when it's first constructed. I did this because I wanted to have a cached version of the last call to `get_head`. Ensuring `get_head` has been run *at least once* means that the cached values doesn't need to wrapped in an `Option`. This was fairly simple, it just involved passing a `slot` to the constructor so it knows *when* it's being run. When loading a fork choice from the store and a slot clock isn't handy I've just used the `slot` that was saved in the `fork_choice_store`. That seems like it would be a faithful representation of the slot when we saved it. I added the `genesis_time: u64` to the `BeaconChain`. It's small, constant and nice to have around. Since we're using FC for the fin/just checkpoints, we no longer get the `0x00..00` roots at genesis. You can see I had to remove a work-around in `ef-tests` here: b56be3bc2. I can't find any reason why this would be an issue, if anything I think it'll be better since the genesis-alias has caught us out a few times (0x00..00 isn't actually a real root). Edit: I did find a case where the `network` expected the 0x00..00 alias and patched it here: 3f26ac3e2. You'll notice a lot of changes in tests. Generally, tests should be functionally equivalent. Here are the things creating the most diff-noise in tests: - Changing tests to be `tokio::async` tests. - Adding `.await` to fork choice, block processing and block production functions. - Refactor of the `canonical_head` "API" provided by the `BeaconChain`. E.g., `chain.canonical_head.cached_head()` instead of `chain.canonical_head.read()`. - Wrapping `SignedBeaconBlock` in an `Arc`. - In the `beacon_chain/tests/block_verification`, we can't use the `lazy_static` `CHAIN_SEGMENT` variable anymore since it's generated with an async function. We just generate it in each test, not so efficient but hopefully insignificant. I had to disable `rayon` concurrent tests in the `fork_choice` tests. This is because the use of `rayon` and `block_on` was causing a panic. Co-authored-by: Mac L <mjladson@pm.me>
2022-07-03 05:36:50 +00:00
let (block_c, state_c) = harness.make_block(state_a, slot_c).await;
let block_root_c = harness
.process_block(slot_c, block_c.canonical_root(), block_c.clone())
Use async code when interacting with EL (#3244) ## Overview This rather extensive PR achieves two primary goals: 1. Uses the finalized/justified checkpoints of fork choice (FC), rather than that of the head state. 2. Refactors fork choice, block production and block processing to `async` functions. Additionally, it achieves: - Concurrent forkchoice updates to the EL and cache pruning after a new head is selected. - Concurrent "block packing" (attestations, etc) and execution payload retrieval during block production. - Concurrent per-block-processing and execution payload verification during block processing. - The `Arc`-ification of `SignedBeaconBlock` during block processing (it's never mutated, so why not?): - I had to do this to deal with sending blocks into spawned tasks. - Previously we were cloning the beacon block at least 2 times during each block processing, these clones are either removed or turned into cheaper `Arc` clones. - We were also `Box`-ing and un-`Box`-ing beacon blocks as they moved throughout the networking crate. This is not a big deal, but it's nice to avoid shifting things between the stack and heap. - Avoids cloning *all the blocks* in *every chain segment* during sync. - It also has the potential to clean up our code where we need to pass an *owned* block around so we can send it back in the case of an error (I didn't do much of this, my PR is already big enough :sweat_smile:) - The `BeaconChain::HeadSafetyStatus` struct was removed. It was an old relic from prior merge specs. For motivation for this change, see https://github.com/sigp/lighthouse/pull/3244#issuecomment-1160963273 ## Changes to `canonical_head` and `fork_choice` Previously, the `BeaconChain` had two separate fields: ``` canonical_head: RwLock<Snapshot>, fork_choice: RwLock<BeaconForkChoice> ``` Now, we have grouped these values under a single struct: ``` canonical_head: CanonicalHead { cached_head: RwLock<Arc<Snapshot>>, fork_choice: RwLock<BeaconForkChoice> } ``` Apart from ergonomics, the only *actual* change here is wrapping the canonical head snapshot in an `Arc`. This means that we no longer need to hold the `cached_head` (`canonical_head`, in old terms) lock when we want to pull some values from it. This was done to avoid deadlock risks by preventing functions from acquiring (and holding) the `cached_head` and `fork_choice` locks simultaneously. ## Breaking Changes ### The `state` (root) field in the `finalized_checkpoint` SSE event Consider the scenario where epoch `n` is just finalized, but `start_slot(n)` is skipped. There are two state roots we might in the `finalized_checkpoint` SSE event: 1. The state root of the finalized block, which is `get_block(finalized_checkpoint.root).state_root`. 4. The state root at slot of `start_slot(n)`, which would be the state from (1), but "skipped forward" through any skip slots. Previously, Lighthouse would choose (2). However, we can see that when [Teku generates that event](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/beaconrestapi/src/main/java/tech/pegasys/teku/beaconrestapi/handlers/v1/events/EventSubscriptionManager.java#L171-L182) it uses [`getStateRootFromBlockRoot`](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/provider/src/main/java/tech/pegasys/teku/api/ChainDataProvider.java#L336-L341) which uses (1). I have switched Lighthouse from (2) to (1). I think it's a somewhat arbitrary choice between the two, where (1) is easier to compute and is consistent with Teku. ## Notes for Reviewers I've renamed `BeaconChain::fork_choice` to `BeaconChain::recompute_head`. Doing this helped ensure I broke all previous uses of fork choice and I also find it more descriptive. It describes an action and can't be confused with trying to get a reference to the `ForkChoice` struct. I've changed the ordering of SSE events when a block is received. It used to be `[block, finalized, head]` and now it's `[block, head, finalized]`. It was easier this way and I don't think we were making any promises about SSE event ordering so it's not "breaking". I've made it so fork choice will run when it's first constructed. I did this because I wanted to have a cached version of the last call to `get_head`. Ensuring `get_head` has been run *at least once* means that the cached values doesn't need to wrapped in an `Option`. This was fairly simple, it just involved passing a `slot` to the constructor so it knows *when* it's being run. When loading a fork choice from the store and a slot clock isn't handy I've just used the `slot` that was saved in the `fork_choice_store`. That seems like it would be a faithful representation of the slot when we saved it. I added the `genesis_time: u64` to the `BeaconChain`. It's small, constant and nice to have around. Since we're using FC for the fin/just checkpoints, we no longer get the `0x00..00` roots at genesis. You can see I had to remove a work-around in `ef-tests` here: b56be3bc2. I can't find any reason why this would be an issue, if anything I think it'll be better since the genesis-alias has caught us out a few times (0x00..00 isn't actually a real root). Edit: I did find a case where the `network` expected the 0x00..00 alias and patched it here: 3f26ac3e2. You'll notice a lot of changes in tests. Generally, tests should be functionally equivalent. Here are the things creating the most diff-noise in tests: - Changing tests to be `tokio::async` tests. - Adding `.await` to fork choice, block processing and block production functions. - Refactor of the `canonical_head` "API" provided by the `BeaconChain`. E.g., `chain.canonical_head.cached_head()` instead of `chain.canonical_head.read()`. - Wrapping `SignedBeaconBlock` in an `Arc`. - In the `beacon_chain/tests/block_verification`, we can't use the `lazy_static` `CHAIN_SEGMENT` variable anymore since it's generated with an async function. We just generate it in each test, not so efficient but hopefully insignificant. I had to disable `rayon` concurrent tests in the `fork_choice` tests. This is because the use of `rayon` and `block_on` was causing a panic. Co-authored-by: Mac L <mjladson@pm.me>
2022-07-03 05:36:50 +00:00
.await
.unwrap();
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
// Create attestations to C from a small number of validators and process them immediately.
let attestations_c = harness.make_attestations(
&all_validators[..validator_count / 2],
&state_c,
state_c.tree_hash_root(),
block_root_c,
slot_c,
);
harness.process_attestations(attestations_c);
// Apply the attestations to B, but don't re-run fork choice.
harness.process_attestations(attestations_b);
// Due to proposer boost, the head should be C during slot C.
assert_eq!(
Use async code when interacting with EL (#3244) ## Overview This rather extensive PR achieves two primary goals: 1. Uses the finalized/justified checkpoints of fork choice (FC), rather than that of the head state. 2. Refactors fork choice, block production and block processing to `async` functions. Additionally, it achieves: - Concurrent forkchoice updates to the EL and cache pruning after a new head is selected. - Concurrent "block packing" (attestations, etc) and execution payload retrieval during block production. - Concurrent per-block-processing and execution payload verification during block processing. - The `Arc`-ification of `SignedBeaconBlock` during block processing (it's never mutated, so why not?): - I had to do this to deal with sending blocks into spawned tasks. - Previously we were cloning the beacon block at least 2 times during each block processing, these clones are either removed or turned into cheaper `Arc` clones. - We were also `Box`-ing and un-`Box`-ing beacon blocks as they moved throughout the networking crate. This is not a big deal, but it's nice to avoid shifting things between the stack and heap. - Avoids cloning *all the blocks* in *every chain segment* during sync. - It also has the potential to clean up our code where we need to pass an *owned* block around so we can send it back in the case of an error (I didn't do much of this, my PR is already big enough :sweat_smile:) - The `BeaconChain::HeadSafetyStatus` struct was removed. It was an old relic from prior merge specs. For motivation for this change, see https://github.com/sigp/lighthouse/pull/3244#issuecomment-1160963273 ## Changes to `canonical_head` and `fork_choice` Previously, the `BeaconChain` had two separate fields: ``` canonical_head: RwLock<Snapshot>, fork_choice: RwLock<BeaconForkChoice> ``` Now, we have grouped these values under a single struct: ``` canonical_head: CanonicalHead { cached_head: RwLock<Arc<Snapshot>>, fork_choice: RwLock<BeaconForkChoice> } ``` Apart from ergonomics, the only *actual* change here is wrapping the canonical head snapshot in an `Arc`. This means that we no longer need to hold the `cached_head` (`canonical_head`, in old terms) lock when we want to pull some values from it. This was done to avoid deadlock risks by preventing functions from acquiring (and holding) the `cached_head` and `fork_choice` locks simultaneously. ## Breaking Changes ### The `state` (root) field in the `finalized_checkpoint` SSE event Consider the scenario where epoch `n` is just finalized, but `start_slot(n)` is skipped. There are two state roots we might in the `finalized_checkpoint` SSE event: 1. The state root of the finalized block, which is `get_block(finalized_checkpoint.root).state_root`. 4. The state root at slot of `start_slot(n)`, which would be the state from (1), but "skipped forward" through any skip slots. Previously, Lighthouse would choose (2). However, we can see that when [Teku generates that event](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/beaconrestapi/src/main/java/tech/pegasys/teku/beaconrestapi/handlers/v1/events/EventSubscriptionManager.java#L171-L182) it uses [`getStateRootFromBlockRoot`](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/provider/src/main/java/tech/pegasys/teku/api/ChainDataProvider.java#L336-L341) which uses (1). I have switched Lighthouse from (2) to (1). I think it's a somewhat arbitrary choice between the two, where (1) is easier to compute and is consistent with Teku. ## Notes for Reviewers I've renamed `BeaconChain::fork_choice` to `BeaconChain::recompute_head`. Doing this helped ensure I broke all previous uses of fork choice and I also find it more descriptive. It describes an action and can't be confused with trying to get a reference to the `ForkChoice` struct. I've changed the ordering of SSE events when a block is received. It used to be `[block, finalized, head]` and now it's `[block, head, finalized]`. It was easier this way and I don't think we were making any promises about SSE event ordering so it's not "breaking". I've made it so fork choice will run when it's first constructed. I did this because I wanted to have a cached version of the last call to `get_head`. Ensuring `get_head` has been run *at least once* means that the cached values doesn't need to wrapped in an `Option`. This was fairly simple, it just involved passing a `slot` to the constructor so it knows *when* it's being run. When loading a fork choice from the store and a slot clock isn't handy I've just used the `slot` that was saved in the `fork_choice_store`. That seems like it would be a faithful representation of the slot when we saved it. I added the `genesis_time: u64` to the `BeaconChain`. It's small, constant and nice to have around. Since we're using FC for the fin/just checkpoints, we no longer get the `0x00..00` roots at genesis. You can see I had to remove a work-around in `ef-tests` here: b56be3bc2. I can't find any reason why this would be an issue, if anything I think it'll be better since the genesis-alias has caught us out a few times (0x00..00 isn't actually a real root). Edit: I did find a case where the `network` expected the 0x00..00 alias and patched it here: 3f26ac3e2. You'll notice a lot of changes in tests. Generally, tests should be functionally equivalent. Here are the things creating the most diff-noise in tests: - Changing tests to be `tokio::async` tests. - Adding `.await` to fork choice, block processing and block production functions. - Refactor of the `canonical_head` "API" provided by the `BeaconChain`. E.g., `chain.canonical_head.cached_head()` instead of `chain.canonical_head.read()`. - Wrapping `SignedBeaconBlock` in an `Arc`. - In the `beacon_chain/tests/block_verification`, we can't use the `lazy_static` `CHAIN_SEGMENT` variable anymore since it's generated with an async function. We just generate it in each test, not so efficient but hopefully insignificant. I had to disable `rayon` concurrent tests in the `fork_choice` tests. This is because the use of `rayon` and `block_on` was causing a panic. Co-authored-by: Mac L <mjladson@pm.me>
2022-07-03 05:36:50 +00:00
harness.chain.canonical_head.cached_head().head_block_root(),
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
block_root_c.into()
);
// Ensure that building a block via the HTTP API re-runs fork choice and builds block D upon B.
// Manually prod the per-slot task, because the slot timer doesn't run in the background in
// these tests.
harness.advance_slot();
Use async code when interacting with EL (#3244) ## Overview This rather extensive PR achieves two primary goals: 1. Uses the finalized/justified checkpoints of fork choice (FC), rather than that of the head state. 2. Refactors fork choice, block production and block processing to `async` functions. Additionally, it achieves: - Concurrent forkchoice updates to the EL and cache pruning after a new head is selected. - Concurrent "block packing" (attestations, etc) and execution payload retrieval during block production. - Concurrent per-block-processing and execution payload verification during block processing. - The `Arc`-ification of `SignedBeaconBlock` during block processing (it's never mutated, so why not?): - I had to do this to deal with sending blocks into spawned tasks. - Previously we were cloning the beacon block at least 2 times during each block processing, these clones are either removed or turned into cheaper `Arc` clones. - We were also `Box`-ing and un-`Box`-ing beacon blocks as they moved throughout the networking crate. This is not a big deal, but it's nice to avoid shifting things between the stack and heap. - Avoids cloning *all the blocks* in *every chain segment* during sync. - It also has the potential to clean up our code where we need to pass an *owned* block around so we can send it back in the case of an error (I didn't do much of this, my PR is already big enough :sweat_smile:) - The `BeaconChain::HeadSafetyStatus` struct was removed. It was an old relic from prior merge specs. For motivation for this change, see https://github.com/sigp/lighthouse/pull/3244#issuecomment-1160963273 ## Changes to `canonical_head` and `fork_choice` Previously, the `BeaconChain` had two separate fields: ``` canonical_head: RwLock<Snapshot>, fork_choice: RwLock<BeaconForkChoice> ``` Now, we have grouped these values under a single struct: ``` canonical_head: CanonicalHead { cached_head: RwLock<Arc<Snapshot>>, fork_choice: RwLock<BeaconForkChoice> } ``` Apart from ergonomics, the only *actual* change here is wrapping the canonical head snapshot in an `Arc`. This means that we no longer need to hold the `cached_head` (`canonical_head`, in old terms) lock when we want to pull some values from it. This was done to avoid deadlock risks by preventing functions from acquiring (and holding) the `cached_head` and `fork_choice` locks simultaneously. ## Breaking Changes ### The `state` (root) field in the `finalized_checkpoint` SSE event Consider the scenario where epoch `n` is just finalized, but `start_slot(n)` is skipped. There are two state roots we might in the `finalized_checkpoint` SSE event: 1. The state root of the finalized block, which is `get_block(finalized_checkpoint.root).state_root`. 4. The state root at slot of `start_slot(n)`, which would be the state from (1), but "skipped forward" through any skip slots. Previously, Lighthouse would choose (2). However, we can see that when [Teku generates that event](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/beaconrestapi/src/main/java/tech/pegasys/teku/beaconrestapi/handlers/v1/events/EventSubscriptionManager.java#L171-L182) it uses [`getStateRootFromBlockRoot`](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/provider/src/main/java/tech/pegasys/teku/api/ChainDataProvider.java#L336-L341) which uses (1). I have switched Lighthouse from (2) to (1). I think it's a somewhat arbitrary choice between the two, where (1) is easier to compute and is consistent with Teku. ## Notes for Reviewers I've renamed `BeaconChain::fork_choice` to `BeaconChain::recompute_head`. Doing this helped ensure I broke all previous uses of fork choice and I also find it more descriptive. It describes an action and can't be confused with trying to get a reference to the `ForkChoice` struct. I've changed the ordering of SSE events when a block is received. It used to be `[block, finalized, head]` and now it's `[block, head, finalized]`. It was easier this way and I don't think we were making any promises about SSE event ordering so it's not "breaking". I've made it so fork choice will run when it's first constructed. I did this because I wanted to have a cached version of the last call to `get_head`. Ensuring `get_head` has been run *at least once* means that the cached values doesn't need to wrapped in an `Option`. This was fairly simple, it just involved passing a `slot` to the constructor so it knows *when* it's being run. When loading a fork choice from the store and a slot clock isn't handy I've just used the `slot` that was saved in the `fork_choice_store`. That seems like it would be a faithful representation of the slot when we saved it. I added the `genesis_time: u64` to the `BeaconChain`. It's small, constant and nice to have around. Since we're using FC for the fin/just checkpoints, we no longer get the `0x00..00` roots at genesis. You can see I had to remove a work-around in `ef-tests` here: b56be3bc2. I can't find any reason why this would be an issue, if anything I think it'll be better since the genesis-alias has caught us out a few times (0x00..00 isn't actually a real root). Edit: I did find a case where the `network` expected the 0x00..00 alias and patched it here: 3f26ac3e2. You'll notice a lot of changes in tests. Generally, tests should be functionally equivalent. Here are the things creating the most diff-noise in tests: - Changing tests to be `tokio::async` tests. - Adding `.await` to fork choice, block processing and block production functions. - Refactor of the `canonical_head` "API" provided by the `BeaconChain`. E.g., `chain.canonical_head.cached_head()` instead of `chain.canonical_head.read()`. - Wrapping `SignedBeaconBlock` in an `Arc`. - In the `beacon_chain/tests/block_verification`, we can't use the `lazy_static` `CHAIN_SEGMENT` variable anymore since it's generated with an async function. We just generate it in each test, not so efficient but hopefully insignificant. I had to disable `rayon` concurrent tests in the `fork_choice` tests. This is because the use of `rayon` and `block_on` was causing a panic. Co-authored-by: Mac L <mjladson@pm.me>
2022-07-03 05:36:50 +00:00
harness.chain.per_slot_task().await;
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
let proposer_index = state_b
.get_beacon_proposer_index(slot_d, &harness.chain.spec)
.unwrap();
let randao_reveal = harness
.sign_randao_reveal(&state_b, proposer_index, slot_d)
.into();
let block_d = tester
.client
.get_validator_blocks::<E, FullPayload<E>>(slot_d, &randao_reveal, None)
.await
.unwrap()
.data;
// Head is now B.
assert_eq!(
Use async code when interacting with EL (#3244) ## Overview This rather extensive PR achieves two primary goals: 1. Uses the finalized/justified checkpoints of fork choice (FC), rather than that of the head state. 2. Refactors fork choice, block production and block processing to `async` functions. Additionally, it achieves: - Concurrent forkchoice updates to the EL and cache pruning after a new head is selected. - Concurrent "block packing" (attestations, etc) and execution payload retrieval during block production. - Concurrent per-block-processing and execution payload verification during block processing. - The `Arc`-ification of `SignedBeaconBlock` during block processing (it's never mutated, so why not?): - I had to do this to deal with sending blocks into spawned tasks. - Previously we were cloning the beacon block at least 2 times during each block processing, these clones are either removed or turned into cheaper `Arc` clones. - We were also `Box`-ing and un-`Box`-ing beacon blocks as they moved throughout the networking crate. This is not a big deal, but it's nice to avoid shifting things between the stack and heap. - Avoids cloning *all the blocks* in *every chain segment* during sync. - It also has the potential to clean up our code where we need to pass an *owned* block around so we can send it back in the case of an error (I didn't do much of this, my PR is already big enough :sweat_smile:) - The `BeaconChain::HeadSafetyStatus` struct was removed. It was an old relic from prior merge specs. For motivation for this change, see https://github.com/sigp/lighthouse/pull/3244#issuecomment-1160963273 ## Changes to `canonical_head` and `fork_choice` Previously, the `BeaconChain` had two separate fields: ``` canonical_head: RwLock<Snapshot>, fork_choice: RwLock<BeaconForkChoice> ``` Now, we have grouped these values under a single struct: ``` canonical_head: CanonicalHead { cached_head: RwLock<Arc<Snapshot>>, fork_choice: RwLock<BeaconForkChoice> } ``` Apart from ergonomics, the only *actual* change here is wrapping the canonical head snapshot in an `Arc`. This means that we no longer need to hold the `cached_head` (`canonical_head`, in old terms) lock when we want to pull some values from it. This was done to avoid deadlock risks by preventing functions from acquiring (and holding) the `cached_head` and `fork_choice` locks simultaneously. ## Breaking Changes ### The `state` (root) field in the `finalized_checkpoint` SSE event Consider the scenario where epoch `n` is just finalized, but `start_slot(n)` is skipped. There are two state roots we might in the `finalized_checkpoint` SSE event: 1. The state root of the finalized block, which is `get_block(finalized_checkpoint.root).state_root`. 4. The state root at slot of `start_slot(n)`, which would be the state from (1), but "skipped forward" through any skip slots. Previously, Lighthouse would choose (2). However, we can see that when [Teku generates that event](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/beaconrestapi/src/main/java/tech/pegasys/teku/beaconrestapi/handlers/v1/events/EventSubscriptionManager.java#L171-L182) it uses [`getStateRootFromBlockRoot`](https://github.com/ConsenSys/teku/blob/de2b2801c89ef5abf983d6bf37867c37fc47121f/data/provider/src/main/java/tech/pegasys/teku/api/ChainDataProvider.java#L336-L341) which uses (1). I have switched Lighthouse from (2) to (1). I think it's a somewhat arbitrary choice between the two, where (1) is easier to compute and is consistent with Teku. ## Notes for Reviewers I've renamed `BeaconChain::fork_choice` to `BeaconChain::recompute_head`. Doing this helped ensure I broke all previous uses of fork choice and I also find it more descriptive. It describes an action and can't be confused with trying to get a reference to the `ForkChoice` struct. I've changed the ordering of SSE events when a block is received. It used to be `[block, finalized, head]` and now it's `[block, head, finalized]`. It was easier this way and I don't think we were making any promises about SSE event ordering so it's not "breaking". I've made it so fork choice will run when it's first constructed. I did this because I wanted to have a cached version of the last call to `get_head`. Ensuring `get_head` has been run *at least once* means that the cached values doesn't need to wrapped in an `Option`. This was fairly simple, it just involved passing a `slot` to the constructor so it knows *when* it's being run. When loading a fork choice from the store and a slot clock isn't handy I've just used the `slot` that was saved in the `fork_choice_store`. That seems like it would be a faithful representation of the slot when we saved it. I added the `genesis_time: u64` to the `BeaconChain`. It's small, constant and nice to have around. Since we're using FC for the fin/just checkpoints, we no longer get the `0x00..00` roots at genesis. You can see I had to remove a work-around in `ef-tests` here: b56be3bc2. I can't find any reason why this would be an issue, if anything I think it'll be better since the genesis-alias has caught us out a few times (0x00..00 isn't actually a real root). Edit: I did find a case where the `network` expected the 0x00..00 alias and patched it here: 3f26ac3e2. You'll notice a lot of changes in tests. Generally, tests should be functionally equivalent. Here are the things creating the most diff-noise in tests: - Changing tests to be `tokio::async` tests. - Adding `.await` to fork choice, block processing and block production functions. - Refactor of the `canonical_head` "API" provided by the `BeaconChain`. E.g., `chain.canonical_head.cached_head()` instead of `chain.canonical_head.read()`. - Wrapping `SignedBeaconBlock` in an `Arc`. - In the `beacon_chain/tests/block_verification`, we can't use the `lazy_static` `CHAIN_SEGMENT` variable anymore since it's generated with an async function. We just generate it in each test, not so efficient but hopefully insignificant. I had to disable `rayon` concurrent tests in the `fork_choice` tests. This is because the use of `rayon` and `block_on` was causing a panic. Co-authored-by: Mac L <mjladson@pm.me>
2022-07-03 05:36:50 +00:00
harness.chain.canonical_head.cached_head().head_block_root(),
Run fork choice before block proposal (#3168) ## Issue Addressed Upcoming spec change https://github.com/ethereum/consensus-specs/pull/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 :cry: ). 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.
2022-05-20 05:02:11 +00:00
block_root_b.into()
);
// D's parent is B.
assert_eq!(block_d.parent_root(), block_root_b.into());
}