991e4094f8
17 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Michael Sproul
|
775d222299 |
Enable proposer boost re-orging (#2860)
## Proposed Changes With proposer boosting implemented (#2822) we have an opportunity to re-org out late blocks. This PR adds three flags to the BN to control this behaviour: * `--disable-proposer-reorgs`: turn aggressive re-orging off (it's on by default). * `--proposer-reorg-threshold N`: attempt to orphan blocks with less than N% of the committee vote. If this parameter isn't set then N defaults to 20% when the feature is enabled. * `--proposer-reorg-epochs-since-finalization N`: only attempt to re-org late blocks when the number of epochs since finalization is less than or equal to N. The default is 2 epochs, meaning re-orgs will only be attempted when the chain is finalizing optimally. For safety Lighthouse will only attempt a re-org under very specific conditions: 1. The block being proposed is 1 slot after the canonical head, and the canonical head is 1 slot after its parent. i.e. at slot `n + 1` rather than building on the block from slot `n` we build on the block from slot `n - 1`. 2. The current canonical head received less than N% of the committee vote. N should be set depending on the proposer boost fraction itself, the fraction of the network that is believed to be applying it, and the size of the largest entity that could be hoarding votes. 3. The current canonical head arrived after the attestation deadline from our perspective. This condition was only added to support suppression of forkchoiceUpdated messages, but makes intuitive sense. 4. The block is being proposed in the first 2 seconds of the slot. This gives it time to propagate and receive the proposer boost. ## Additional Info For the initial idea and background, see: https://github.com/ethereum/consensus-specs/pull/2353#issuecomment-950238004 There is also a specification for this feature here: https://github.com/ethereum/consensus-specs/pull/3034 Co-authored-by: Michael Sproul <micsproul@gmail.com> Co-authored-by: pawan <pawandhananjay@gmail.com> |
||
|
Michael Sproul
|
84392d63fa |
Delete DB schema migrations for v11 and earlier (#3761)
## Proposed Changes Now that the Gnosis merge is scheduled, all users should have upgraded beyond Lighthouse v3.0.0. Accordingly we can delete schema migrations for versions prior to v3.0.0. ## Additional Info I also deleted the state cache stuff I added in #3714 as it turned out to be useless for the light client proofs due to the one-slot offset. |
||
|
Michael Sproul
|
3be41006a6 |
Add --light-client-server flag and state cache utils (#3714)
## Issue Addressed Part of https://github.com/sigp/lighthouse/issues/3651. ## Proposed Changes Add a flag for enabling the light client server, which should be checked before gossip/RPC traffic is processed (e.g. https://github.com/sigp/lighthouse/pull/3693, https://github.com/sigp/lighthouse/pull/3711). The flag is available at runtime from `beacon_chain.config.enable_light_client_server`. Additionally, a new method `BeaconChain::with_mutable_state_for_block` is added which I envisage being used for computing light client updates. Unfortunately its performance will be quite poor on average because it will only run quickly with access to the tree hash cache. Each slot the tree hash cache is only available for a brief window of time between the head block being processed and the state advance at 9s in the slot. When the state advance happens the cache is moved and mutated to get ready for the next slot, which makes it no longer useful for merkle proofs related to the head block. Rather than spend more time trying to optimise this I think we should continue prototyping with this code, and I'll make sure `tree-states` is ready to ship before we enable the light client server in prod (cf. https://github.com/sigp/lighthouse/pull/3206). ## Additional Info I also fixed a bug in the implementation of `BeaconState::compute_merkle_proof` whereby the tree hash cache was moved with `.take()` but never put back with `.restore()`. |
||
Paul Hauner
|
be4e261e74 |
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 😅)
- 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](
|
||
|
Michael Sproul
|
efebf712dd |
Avoid cloning snapshots during sync (#3271)
## Issue Addressed Closes https://github.com/sigp/lighthouse/issues/2944 ## Proposed Changes Remove snapshots from the cache during sync rather than cloning them. This reduces unnecessary cloning and memory fragmentation during sync. ## Additional Info This PR relies on the fact that the `block_delay` cache is not populated for blocks from sync. Relying on block delay may have the side effect that a change in `block_delay` calculation could lead to: a) more clones, if block delays are added for syncing blocks or b) less clones, if blocks near the head are erroneously provided without a `block_delay`. Case (a) would be a regression to the current status quo, and (b) is low-risk given we know that the snapshot cache is current susceptible to misses (hence `tree-states`). |
||
|
Michael Sproul
|
bcdd960ab1 |
Separate execution payloads in the DB (#3157)
## Proposed Changes Reduce post-merge disk usage by not storing finalized execution payloads in Lighthouse's database. ⚠️ **This is achieved in a backwards-incompatible way for networks that have already merged** ⚠️. Kiln users and shadow fork enjoyers will be unable to downgrade after running the code from this PR. The upgrade migration may take several minutes to run, and can't be aborted after it begins. The main changes are: - New column in the database called `ExecPayload`, keyed by beacon block root. - The `BeaconBlock` column now stores blinded blocks only. - Lots of places that previously used full blocks now use blinded blocks, e.g. analytics APIs, block replay in the DB, etc. - On finalization: - `prune_abanonded_forks` deletes non-canonical payloads whilst deleting non-canonical blocks. - `migrate_db` deletes finalized canonical payloads whilst deleting finalized states. - Conversions between blinded and full blocks are implemented in a compositional way, duplicating some work from Sean's PR #3134. - The execution layer has a new `get_payload_by_block_hash` method that reconstructs a payload using the EE's `eth_getBlockByHash` call. - I've tested manually that it works on Kiln, using Geth and Nethermind. - This isn't necessarily the most efficient method, and new engine APIs are being discussed to improve this: https://github.com/ethereum/execution-apis/pull/146. - We're depending on the `ethers` master branch, due to lots of recent changes. We're also using a workaround for https://github.com/gakonst/ethers-rs/issues/1134. - Payload reconstruction is used in the HTTP API via `BeaconChain::get_block`, which is now `async`. Due to the `async` fn, the `blocking_json` wrapper has been removed. - Payload reconstruction is used in network RPC to serve blocks-by-{root,range} responses. Here the `async` adjustment is messier, although I think I've managed to come up with a reasonable compromise: the handlers take the `SendOnDrop` by value so that they can drop it on _task completion_ (after the `fn` returns). Still, this is introducing disk reads onto core executor threads, which may have a negative performance impact (thoughts appreciated). ## Additional Info - [x] For performance it would be great to remove the cloning of full blocks when converting them to blinded blocks to write to disk. I'm going to experiment with a `put_block` API that takes the block by value, breaks it into a blinded block and a payload, stores the blinded block, and then re-assembles the full block for the caller. - [x] We should measure the latency of blocks-by-root and blocks-by-range responses. - [x] We should add integration tests that stress the payload reconstruction (basic tests done, issue for more extensive tests: https://github.com/sigp/lighthouse/issues/3159) - [x] We should (manually) test the schema v9 migration from several prior versions, particularly as blocks have changed on disk and some migrations rely on being able to load blocks. Co-authored-by: Paul Hauner <paul@paulhauner.com> |
||
|
Paul Hauner
|
c11253a82f |
Remove grandparents from snapshot cache (#2917)
## Issue Addressed NA ## Proposed Changes In https://github.com/sigp/lighthouse/pull/2832 we made some changes to the `SnapshotCache` to help deal with the one-block reorgs seen on mainnet (and testnets). I believe the change in #2832 is good and we should keep it, but I think that in its present form it is causing the `SnapshotCache` to hold onto states that it doesn't need anymore. For example, a skip slot will result in one more `BeaconSnapshot` being stored in the cache. This PR adds a new type of pruning that happens after a block is inserted to the cache. We will remove any snapshot from the cache that is a *grandparent* of the block being imported. Since we know the grandparent has two valid blocks built atop it, it is not at risk from a one-block re-org. ## Additional Info NA |
||
Mac L
|
a7a7edb6cf |
Optimise snapshot cache for late blocks (#2832)
## Proposed Changes In the event of a late block, keep the block in the snapshot cache by cloning it. This helps us process new blocks quickly in the event the late block was re-org'd. Co-authored-by: Michael Sproul <michael@sigmaprime.io> |
||
|
Paul Hauner
|
e2d09bb8ac |
Add BeaconChainHarness::builder (#2707)
## Issue Addressed NA ## Proposed Changes This PR is near-identical to https://github.com/sigp/lighthouse/pull/2652, however it is to be merged into `unstable` instead of `merge-f2f`. Please see that PR for reasoning. I'm making this duplicate PR to merge to `unstable` in an effort to shrink the diff between `unstable` and `merge-f2f` by doing smaller, lead-up PRs. ## Additional Info NA |
||
|
Michael Sproul
|
b4689e20c6 |
Altair consensus changes and refactors (#2279)
## Proposed Changes Implement the consensus changes necessary for the upcoming Altair hard fork. ## Additional Info This is quite a heavy refactor, with pivotal types like the `BeaconState` and `BeaconBlock` changing from structs to enums. This ripples through the whole codebase with field accesses changing to methods, e.g. `state.slot` => `state.slot()`. Co-authored-by: realbigsean <seananderson33@gmail.com> |
||
|
Paul Hauner
|
015ab7d0a7 |
Optimize validator duties (#2243)
## Issue Addressed Closes #2052 ## Proposed Changes - Refactor the attester/proposer duties endpoints in the BN - Performance improvements - Fixes some potential inconsistencies with the dependent root fields. - Removes `http_api::beacon_proposer_cache` and just uses the one on the `BeaconChain` instead. - Move the code for the proposer/attester duties endpoints into separate files, for readability. - Refactor the `DutiesService` in the VC - Required to reduce the delay on broadcasting new blocks. - Gets rid of the `ValidatorDuty` shim struct that came about when we adopted the standard API. - Separate block/attestation duty tasks so that they don't block each other when one is slow. - In the VC, use `PublicKeyBytes` to represent validators instead of `PublicKey`. `PublicKey` is a legit crypto object whilst `PublicKeyBytes` is just a byte-array, it's much faster to clone/hash `PublicKeyBytes` and this change has had a significant impact on runtimes. - Unfortunately this has created lots of dust changes. - In the BN, store `PublicKeyBytes` in the `beacon_proposer_cache` and allow access to them. The HTTP API always sends `PublicKeyBytes` over the wire and the conversion from `PublicKey` -> `PublickeyBytes` is non-trivial, especially when queries have 100s/1000s of validators (like Pyrmont). - Add the `state_processing::state_advance` mod which dedups a lot of the "apply `n` skip slots to the state" code. - This also fixes a bug with some functions which were failing to include a state root as per [this comment]( |
||
|
Paul Hauner
|
e4eb0eb168 |
Use advanced state for block production (#2241)
## Issue Addressed NA ## Proposed Changes - Use the pre-states from #2174 during block production. - Running this on Pyrmont shows block production times dropping from ~550ms to ~150ms. - Create `crit` and `warn` logs when a block is published to the API later than we expect. - On mainnet we are issuing a warn if the block is published more than 1s later than the slot start and a crit for more than 3s. - Rename some methods on the `SnapshotCache` for clarity. - Add the ability to pass the state root to `BeaconChain::produce_block_on_state` to avoid computing a state root. This is a very common LH optimization. - Add a metric that tracks how late we broadcast blocks received from the HTTP API. This is *technically* a duplicate of a `ValidatorMonitor` log, but I wanted to have it for the case where we aren't monitoring validators too. |
||
|
Paul Hauner
|
a764c3b247 |
Handle early blocks (#2155)
## Issue Addressed
NA
## Problem this PR addresses
There's an issue where Lighthouse is banning a lot of peers due to the following sequence of events:
1. Gossip block 0xabc arrives ~200ms early
- It is propagated across the network, with respect to [`MAXIMUM_GOSSIP_CLOCK_DISPARITY`](https://github.com/ethereum/eth2.0-specs/blob/v1.0.0/specs/phase0/p2p-interface.md#why-is-there-maximum_gossip_clock_disparity-when-validating-slot-ranges-of-messages-in-gossip-subnets).
- However, it is not imported to our database since the block is early.
2. Attestations for 0xabc arrive, but the block was not imported.
- The peer that sent the attestation is down-voted.
- Each unknown-block attestation causes a score loss of 1, the peer is banned at -100.
- When the peer is on an attestation subnet there can be hundreds of attestations, so the peer is banned quickly (before the missed block can be obtained via rpc).
## Potential solutions
I can think of three solutions to this:
1. Wait for attestation-queuing (#635) to arrive and solve this.
- Easy
- Not immediate fix.
- Whilst this would work, I don't think it's a perfect solution for this particular issue, rather (3) is better.
1. Allow importing blocks with a tolerance of `MAXIMUM_GOSSIP_CLOCK_DISPARITY`.
- Easy
- ~~I have implemented this, for now.~~
1. If a block is verified for gossip propagation (i.e., signature verified) and it's within `MAXIMUM_GOSSIP_CLOCK_DISPARITY`, then queue it to be processed at the start of the appropriate slot.
- More difficult
- Feels like the best solution, I will try to implement this.
**This PR takes approach (3).**
## Changes included
- Implement the `block_delay_queue`, based upon a [`DelayQueue`](https://docs.rs/tokio-util/0.6.3/tokio_util/time/delay_queue/struct.DelayQueue.html) which can store blocks until it's time to import them.
- Add a new `DelayedImportBlock` variant to the `beacon_processor::WorkEvent` enum to handle this new event.
- In the `BeaconProcessor`, refactor a `tokio::select!` to a struct with an explicit `Stream` implementation. I experienced some issues with `tokio::select!` in the block delay queue and I also found it hard to debug. I think this explicit implementation is nicer and functionally equivalent (apart from the fact that `tokio::select!` randomly chooses futures to poll, whereas now we're deterministic).
- Add a testing framework to the `beacon_processor` module that tests this new block delay logic. I also tested a handful of other operations in the beacon processor (attns, slashings, exits) since it was super easy to copy-pasta the code from the `http_api` tester.
- To implement these tests I added the concept of an optional `work_journal_tx` to the `BeaconProcessor` which will spit out a log of events. I used this in the tests to ensure that things were happening as I expect.
- The tests are a little racey, but it's hard to avoid that when testing timing-based code. If we see CI failures I can revise. I haven't observed *any* failures due to races on my machine or on CI yet.
- To assist with testing I allowed for directly setting the time on the `ManualSlotClock`.
- I gave the `beacon_processor::Worker` a `Toolbox` for two reasons; (a) it avoids changing tons of function sigs when you want to pass a new object to the worker and (b) it seemed cute.
|
||
|
Paul Hauner
|
88cc222204 |
Advance state to next slot after importing block (#2174)
## Issue Addressed
NA
## Proposed Changes
Add an optimization to perform `per_slot_processing` from the *leading-edge* of block processing to the *trailing-edge*. Ultimately, this allows us to import the block at slot `n` faster because we used the tail-end of slot `n - 1` to perform `per_slot_processing`.
Additionally, add a "block proposer cache" which allows us to cache the block proposer for some epoch. Since we're now doing trailing-edge `per_slot_processing`, we can prime this cache with the values for the next epoch before those blocks arrive (assuming those blocks don't have some weird forking).
There were several ancillary changes required to achieve this:
- Remove the `state_root` field of `BeaconSnapshot`, since there's no need to know it on a `pre_state` and in all other cases we can just read it from `block.state_root()`.
- This caused some "dust" changes of `snapshot.beacon_state_root` to `snapshot.beacon_state_root()`, where the `BeaconSnapshot::beacon_state_root()` func just reads the state root from the block.
- Rename `types::ShuffingId` to `AttestationShufflingId`. I originally did this because I added a `ProposerShufflingId` struct which turned out to be not so useful. I thought this new name was more descriptive so I kept it.
- Address https://github.com/ethereum/eth2.0-specs/pull/2196
- Add a debug log when we get a block with an unknown parent. There was previously no logging around this case.
- Add a function to `BeaconState` to compute all proposers for an epoch without re-computing the active indices for each slot.
## Additional Info
- ~~Blocked on #2173~~
- ~~Blocked on #2179~~ That PR was wrapped into this PR.
- There's potentially some places where we could avoid computing the proposer indices in `per_block_processing` but I haven't done this here. These would be an optimization beyond the issue at hand (improving block propagation times) and I think this PR is already doing enough. We can come back for that later.
## TODO
- [x] Tidy, improve comments.
- [x] ~~Try avoid computing proposer index in `per_block_processing`?~~
|
||
|
Michael Sproul
|
e5bf2576f1 |
Optimise tree hash caching for block production (#2106)
## Proposed Changes
`@potuz` on the Eth R&D Discord observed that Lighthouse blocks on Pyrmont were always arriving at other nodes after at least 1 second. Part of this could be due to processing and slow propagation, but metrics also revealed that the Lighthouse nodes were usually taking 400-600ms to even just produce a block before broadcasting it.
I tracked the slowness down to the lack of a pre-built tree hash cache (THC) on the states being used for block production. This was due to using the head state for block production, which lacks a THC in order to keep fork choice fast (cloning a THC takes at least 30ms for 100k validators). This PR modifies block production to clone a state from the snapshot cache rather than the head, which speeds things up by 200-400ms by avoiding the tree hash cache rebuild. In practice this seems to have cut block production time down to 300ms or less. Ideally we could _remove_ the snapshot from the cache (and save the 30ms), but it is required for when we re-process the block after signing it with the validator client.
## Alternatives
I experimented with 2 alternatives to this approach, before deciding on it:
* Alternative 1: ensure the `head` has a tree hash cache. This is too slow, as it imposes a +30ms hit on fork choice, which currently takes ~5ms (with occasional spikes).
* Alternative 2: use `Arc<BeaconSnapshot>` in the snapshot cache and share snapshots between the cache and the `head`. This made fork choice blazing fast (1ms), and block production the same as in this PR, but had a negative impact on block processing which I don't think is worth it. It ended up being necessary to clone the full state from the snapshot cache during block production, imposing the +30ms penalty there _as well_ as in block production.
In contract, the approach in this PR should only impact block production, and it improves it! Yay for pareto improvements 🎉
## Additional Info
This commit (ac59dfa) is currently running on all the Lighthouse Pyrmont nodes, and I've added a dashboard to the Pyrmont grafana instance with the metrics.
In future work we should optimise the attestation packing, which consumes around 30-60ms and is now a substantial contributor to the total.
|
||
|
Paul Hauner
|
b73c497be2 |
Support multiple BLS implementations (#1335)
## Issue Addressed NA ## Proposed Changes - Refactor the `bls` crate to support multiple BLS "backends" (e.g., milagro, blst, etc). - Removes some duplicate, unused code in `common/rest_types/src/validator.rs`. - Removes the old "upgrade legacy keypairs" functionality (these were unencrypted keys that haven't been supported for a few testnets, no one should be using them anymore). ## Additional Info Most of the files changed are just inconsequential changes to function names. ## TODO - [x] Optimization levels - [x] Infinity point: https://github.com/supranational/blst/issues/11 - [x] Ensure milagro *and* blst are tested via CI - [x] What to do with unsafe code? - [x] Test infinity point in signature sets |
||
|
Paul Hauner
|
2fb6b7c793
|
Add no-copy block processing cache (#863)
* Add state cache, remove store cache * Only build the head committee cache * Fix compile error * Fix compile error from merge * Rename state_cache -> checkpoint_cache * Rename Checkpoint -> Snapshot * Tidy, add comments * Tidy up find_head function * Change some checkpoint -> snapshot * Add tests * Expose max_len * Remove dead code * Tidy * Fix bug |