## Issue Addressed
Closes#3963 (hopefully)
## Proposed Changes
Compute attestation selection proofs gradually each slot rather than in a single `join_all` at the start of each epoch. On a machine with 5k validators this replaces 5k tasks signing 5k proofs with 1 task that signs 5k/32 ~= 160 proofs each slot.
Based on testing with Goerli validators this seems to reduce the average time to produce a signature by preventing Tokio and the OS from falling over each other trying to run hundreds of threads. My testing so far has been with local keystores, which run on a dynamic pool of up to 512 OS threads because they use [`spawn_blocking`](https://docs.rs/tokio/1.11.0/tokio/task/fn.spawn_blocking.html) (and we haven't changed the default).
An earlier version of this PR hyper-optimised the time-per-signature metric to the detriment of the entire system's performance (see the reverted commits). The current PR is conservative in that it avoids touching the attestation service at all. I think there's more optimising to do here, but we can come back for that in a future PR rather than expanding the scope of this one.
The new algorithm for attestation selection proofs is:
- We sign a small batch of selection proofs each slot, for slots up to 8 slots in the future. On average we'll sign one slot's worth of proofs per slot, with an 8 slot lookahead.
- The batch is signed halfway through the slot when there is unlikely to be contention for signature production (blocks are <4s, attestations are ~4-6 seconds, aggregates are 8s+).
## Performance Data
_See first comment for updated graphs_.
Graph of median signing times before this PR:
Graph of update attesters metric (includes selection proof signing) before this PR:
Median signing time after this PR (prototype from 12:00, updated version from 13:30):
99th percentile on signing times (bounded attestation signing from 16:55, now removed):
Attester map update timing after this PR:
Selection proof signings per second change:
## Link to late blocks
I believe this is related to the slow block signings because logs from Stakely in #3963 show these two logs almost 5 seconds apart:
> Feb 23 18:56:23.978 INFO Received unsigned block, slot: 5862880, service: block, module: validator_client::block_service:393
> Feb 23 18:56:28.552 INFO Publishing signed block, slot: 5862880, service: block, module: validator_client::block_service:416
The only thing that happens between those two logs is the signing of the block:
0fb58a680d/validator_client/src/block_service.rs (L410-L414)
Helpfully, Stakely noticed this issue without any Lighthouse BNs in the mix, which pointed to a clear issue in the VC.
## TODO
- [x] Further testing on testnet infrastructure.
- [x] Make the attestation signing parallelism configurable.
## Issue Addressed
Closes#3896Closes#3998Closes#3700
## Proposed Changes
- Optimise the calculation of withdrawals for payload attributes by avoiding state clones, avoiding unnecessary state advances and reading from the snapshot cache if possible.
- Use the execution layer's payload attributes cache to avoid re-calculating payload attributes. I actually implemented a new LRU cache just for withdrawals but it had the exact same key and most of the same data as the existing payload attributes cache, so I deleted it.
- Add a new SSE event that fires when payloadAttributes are calculated. This is useful for block builders, a la https://github.com/ethereum/beacon-APIs/issues/244.
- Add a new CLI flag `--always-prepare-payload` which forces payload attributes to be sent with every fcU regardless of connected proposers. This is intended for use by builders/relays.
For maximum effect, the flags I've been using to run Lighthouse in "payload builder mode" are:
```
--always-prepare-payload \
--prepare-payload-lookahead 12000 \
--suggested-fee-recipient 0x0000000000000000000000000000000000000000
```
The fee recipient is required so Lighthouse has something to pack in the payload attributes (it can be ignored by the builder). The lookahead causes fcU to be sent at the start of every slot rather than at 8s. As usual, fcU will also be sent after each change of head block. I think this combination is sufficient for builders to build on all viable heads. Often there will be two fcU (and two payload attributes) sent for the same slot: one sent at the start of the slot with the head from `n - 1` as the parent, and one sent after the block arrives with `n` as the parent.
Example usage of the new event stream:
```bash
curl -N "http://localhost:5052/eth/v1/events?topics=payload_attributes"
```
## Additional Info
- [x] Tests added by updating the proposer re-org tests. This has the benefit of testing the proposer re-org code paths with withdrawals too, confirming that the new changes don't interact poorly.
- [ ] Benchmarking with `blockdreamer` on devnet-7 showed promising results but I'm yet to do a comparison to `unstable`.
Co-authored-by: Michael Sproul <micsproul@gmail.com>
## Issue Addressed
NA
## Proposed Changes
Adds a service which periodically polls (11s into each mainnet slot) the `node/version` endpoint on each BN and roughly measures the round-trip latency. The latency is exposed as a `DEBG` log and a Prometheus metric.
The `--latency-measurement-service` has been added to the VC, with the following options:
- `--latency-measurement-service true`: enable the service (default).
- `--latency-measurement-service`: (without a value) has the same effect.
- `--latency-measurement-service false`: disable the service.
## Additional Info
Whilst looking at our staking setup, I think the BN+VC latency is contributing to late blocks. Now that we have to wait for the builders to respond it's nice to try and do everything we can to reduce that latency. Having visibility is the first step.
## Issue Addressed
NA
## Proposed Changes
As discovered in #4034, Lighthouse is not accepting `latest_valid_hash == None` in an `INVALID` response to `newPayload`. The `null`/`None` response *was* illegal at one point, however it was added in https://github.com/ethereum/execution-apis/pull/254.
This PR brings Lighthouse in line with the standard and should fix the root cause of what #4034 patched around.
## Additional Info
NA
## Issue Addressed
Cleaner resolution for #4006
## Proposed Changes
We are currently subscribing to core topics of new forks way before the actual fork since we had just a single `CORE_TOPICS` array. This PR separates the core topics for every fork and subscribes to only required topics based on the current fork.
Also adds logic for subscribing to the core topics of a new fork only 2 slots before the fork happens.
2 slots is to give enough time for the gossip meshes to form.
Currently doesn't add logic to remove topics from older forks in new forks. For e.g. in the coupled 4844 world, we had to remove the `BeaconBlock` topic in favour of `BeaconBlocksAndBlobsSidecar` at the 4844 fork. It should be easy enough to add though. Not adding it because I'm assuming that #4019 will get merged before this PR and we won't require any deletion logic. Happy to add it regardless though.
## Proposed Changes
Remove built-in support for Ropsten and Kiln via the `--network` flag. Both testnets are long dead and deprecated.
This shaves about 30MiB off the binary size, from 135MiB to 103MiB (maxperf), or 165MiB to 135MiB (release).
## Issue Addressed
Cleans up all the remnants of 4844 in capella. This makes sure when 4844 is reviewed there is nothing we are missing because it got included here
## Proposed Changes
drop a bomb on every 4844 thing
## Additional Info
Merge process I did (locally) is as follows:
- squash merge to produce one commit
- in new branch off unstable with the squashed commit create a `git revert HEAD` commit
- merge that new branch onto 4844 with `--strategy ours`
- compare local 4844 to remote 4844 and make sure the diff is empty
- enjoy
Co-authored-by: Paul Hauner <paul@paulhauner.com>
## Issue Addressed
NA
## Proposed Changes
Adds two new `DEBG` logs to the HTTP API:
1. As soon as we are requested to produce a block.
2. As soon as a signed block is received.
In #3858 we added some very helpful logs to the VC so we could see when things are happening with block proposals in the VC. After doing some more debugging, I found that I can tell when the VC is sending a block but I *can't* tell the time that the BN receives it (I can only get the time after the BN has started doing some work with the block). Knowing when the VC published and as soon as the BN receives is useful for determining the delays introduced by network latency (and some other things like JSON decoding, etc).
## Additional Info
NA
## Issue Addressed
NA
## Proposed Changes
Updates our `ef_tests` to use: https://github.com/ethereum/consensus-specs/releases/tag/v1.3.0-rc.3
This required:
- Skipping a `merkle_proof_validity` test (see #4022)
- Account for the `eip4844` tests changing name to `deneb`
- My IDE did some Python linting during this change. It seemed simple and nice so I left it there.
## Additional Info
NA
This fixes issues with certain metrics scrapers, which might error if the content-type is not correctly set.
## Issue Addressed
Fixes https://github.com/sigp/lighthouse/issues/3437
## Proposed Changes
Simply set header: `Content-Type: text/plain` on metrics server response. Seems like the errored branch does this correctly already.
## Additional Info
This is needed also to enable influx-db metric scraping which work very nicely with Geth.
## Proposed Changes
Allowing compiling without MDBX by running:
```bash
CARGO_INSTALL_EXTRA_FLAGS="--no-default-features" make
```
The reasons to do this are several:
- Save compilation time if the slasher won't be used
- Work around compilation errors in slasher backend dependencies (our pinned version of MDBX is currently not compiling on FreeBSD with certain compiler versions).
## Additional Info
When I opened this PR we were using resolver v1 which [doesn't disable default features in dependencies](https://doc.rust-lang.org/cargo/reference/features.html#resolver-version-2-command-line-flags), and `mdbx` is default for the `slasher` crate. Even after the resolver got changed to v2 in #3697 compiling with `--no-default-features` _still_ wasn't turning off the slasher crate's default features, so I added `default-features = false` in all the places we depend on it.
Co-authored-by: Michael Sproul <micsproul@gmail.com>
## Issue Addressed
NA
## Proposed Changes
- Bump versions
## Sepolia Capella Upgrade
This release will enable the Capella fork on Sepolia. We are planning to publish this release on the 23rd of Feb 2023.
Users who can build from source and wish to do pre-release testing can use this branch.
## Additional Info
- [ ] Requires further testing
This is a correction to #3757.
The correction registers a peer that is being disconnected in the local peer manager db to ensure we are tracking the correct state.
## Issue Addressed
N/A
## Proposed Changes
The doppelganger tests were failing silently since the `PROPOSER_BOOST` config was not set. Sets the config and script returns an error if any subprocess fails.
## Issue Addressed
#3804
## Proposed Changes
- Add `total_balance` to the validator monitor and adjust the number of historical epochs which are cached.
- Allow certain values in the cache to be served out via the HTTP API without requiring a state read.
## Usage
```
curl -X POST "http://localhost:5052/lighthouse/ui/validator_info" -d '{"indices": [0]}' -H "Content-Type: application/json" | jq
```
```
{
"data": {
"validators": {
"0": {
"info": [
{
"epoch": 172981,
"total_balance": 36566388519
},
...
{
"epoch": 172990,
"total_balance": 36566496513
}
]
},
"1": {
"info": [
{
"epoch": 172981,
"total_balance": 36355797968
},
...
{
"epoch": 172990,
"total_balance": 36355905962
}
]
}
}
}
}
```
## Additional Info
This requires no historical states to operate which mean it will still function on the freshly checkpoint synced node, however because of this, the values will populate each epoch (up to a maximum of 10 entries).
Another benefit of this method, is that we can easily cache any other values which would normally require a state read and serve them via the same endpoint. However, we would need be cautious about not overly increasing block processing time by caching values from complex computations.
This also caches some of the validator metrics directly, rather than pulling them from the Prometheus metrics when the API is called. This means when the validator count exceeds the individual monitor threshold, the cached values will still be available.
Co-authored-by: Paul Hauner <paul@paulhauner.com>
* Modify comment to only include 4844
Capella only modifies per epoch processing by adding
`process_historical_summaries_update`, which does not change the realization of
justification or finality.
Whilst 4844 does not currently modify realization, the spec is not yet final
enough to say that it never will.
* Clarify address change verification comment
The verification of the address change doesn't really have anything to do with
the current epoch. I think this was just a copy-paste from a function like
`verify_exit`.
* Add extra encoding/decoding tests
* Remove TODO
The method LGTM
* Remove `FreeAttestation`
This is an ancient relic, I'm surprised it still existed!
* Add paranoid check for eip4844 code
This is not technically necessary, but I think it's nice to be explicit about
EIP4844 consensus code for the time being.
* Reduce big-O complexity of address change pruning
I'm not sure this is *actually* useful, but it might come in handy if we see a
ton of address changes at the fork boundary. I know the devops team have been
testing with ~100k changes, so maybe this will help in that case.
* Revert "Reduce big-O complexity of address change pruning"
This reverts commit e7d93e6cc7cf1b92dd5a9e1966ce47d4078121eb.
* Remove CapellaReadiness::NotSynced
Some EEs have a habit of flipping between synced/not-synced, which causes some
spurious "Not read for the merge" messages back before the merge. For the
merge, if the EE wasn't synced the CE simple wouldn't go through the transition
(due to optimistic sync stuff). However, we don't have that hard requirement
for Capella; the CE will go through the fork and just wait for the EE to catch
up. I think that removing `NotSynced` here will avoid false-positives on the
"Not ready logs..". We'll be creating other WARN/ERRO logs if the EE isn't
synced, anyway.
* Change some Capella readiness logging
There's two changes here:
1. Shorten the log messages, for readability.
2. Change the hints.
Connecting a Capella-ready LH to a non-Capella-ready EE gives this log:
```
WARN Not ready for Capella info: The execution endpoint does not appear to support the required engine api methods for Capella: Required Methods Unsupported: engine_getPayloadV2 engine_forkchoiceUpdatedV2 engine_newPayloadV2, service: slot_notifier
```
This variant of error doesn't get a "try updating" style hint, when it's the
one that needs it. This is because we detect the method-not-found reponse from
the EE and return default capabilities, rather than indicating that the request
fails. I think it's fair to say that an EE upgrade is required whenever it
doesn't provide the required methods.
I changed the `ExchangeCapabilitiesFailed` message since that can only happen
when the EE fails to respond with anything other than success or not-found.
## Proposed Changes
* Bump Go from 1.17 to 1.20. The latest Geth release v1.11.0 requires 1.18 minimum.
* Prevent a cache miss during payload building by using the right fee recipient. This prevents Geth v1.11.0 from building a block with 0 transactions. The payload building mechanism is overhauled in the new Geth to improve the payload every 2s, and the tests were failing because we were falling back on a `getPayload` call with no lookahead due to `get_payload_id` cache miss caused by the mismatched fee recipient. Alternatively we could hack the tests to send `proposer_preparation_data`, but I think the static fee recipient is simpler for now.
* Add support for optionally enabling Lighthouse logs in the integration tests. Enable using `cargo run --release --features logging/test_logger`. This was very useful for debugging.
## Issue Addressed
I discovered this issue while implementing [this test](https://github.com/jimmygchen/lighthouse/blob/test-example/beacon_node/network/src/beacon_processor/tests.rs#L895), where I tried to manipulate the slot clock with:
`rig.chain.slot_clock.set_current_time(duration);`
however the change doesn't get reflected in the `slot_clock` in `ReprocessQueue`, and I realised `slot_clock` was cloned a few times in the code, and therefore changing the time in `rig.chain.slot_clock` doesn't have any effect in `ReprocessQueue`.
I've incorporated the suggestion from the @paulhauner and @michaelsproul - wrapping the `ManualSlotClock.current_time` (`RwLock<Duration>)` in an `Arc`, and the above test now passes.
Let's see if this breaks any existing tests :)
## Issue Addressed
Windows tests for subscription and unsubscriptions fail in CI sporadically. We usually ignore this failures, so this PR aims to help reduce the failure noise. Associated issue is https://github.com/sigp/lighthouse/issues/3960
On heavily crowded networks, we are seeing many attempted connections to our node every second.
Often these connections come from peers that have just been disconnected. This can be for a number of reasons including:
- We have deemed them to be not as useful as other peers
- They have performed poorly
- They have dropped the connection with us
- The connection was spontaneously lost
- They were randomly removed because we have too many peers
In all of these cases, if we have reached or exceeded our target peer limit, there is no desire to accept new connections immediately after the disconnect from these peers. In fact, it often costs us resources to handle the established connections and defeats some of the logic of dropping them in the first place.
This PR adds a timeout, that prevents recently disconnected peers from reconnecting to us.
Technically we implement a ban at the swarm layer to prevent immediate re connections for at least 10 minutes. I decided to keep this light, and use a time-based LRUCache which only gets updated during the peer manager heartbeat to prevent added stress of polling a delay map for what could be a large number of peers.
This cache is bounded in time. An extra space bound could be added should people consider this a risk.
Co-authored-by: Diva M <divma@protonmail.com>
## Issue Addressed
The documentation referring to build from source mismatches with the what gitworkflow uses.
aa5b7ef783/book/src/installation-source.md (L118-L120)
## Proposed Changes
Because the github workflow uses `cross` to build from source and for that build there is different env variable `CROSS_FEATURES` so need pass at the compile time.
## Additional Info
Verified that existing `-dev` builds does not contains the `minimal` spec enabled.
```bash
> docker run --rm --name node-5-cl-lighthouse sigp/lighthouse:latest-amd64-unstable-dev lighthouse --version
Lighthouse v3.4.0-aa5b7ef
BLS library: blst-portable
SHA256 hardware acceleration: true
Allocator: jemalloc
Specs: mainnet (true), minimal (false), gnosis (true)
```
## Issue Addressed
Fix a bug introduced by #3696. The bug is not expected to occur frequently, so releasing this PR is non-urgent.
## Proposed Changes
* Add a variant to `StoreOp` that allows a raw KV operation to be passed around.
* Return to using `self.store.do_atomically` rather than `self.store.hot_db.do_atomically`. This streamlines the write back into a single call and makes our auto-revert work again.
* Prevent `import_block_update_shuffling_cache` from failing block import. This is an outstanding bug from before v3.4.0 which may have contributed to some random unexplained database corruption.
## Additional Info
In #3696 I split the database write into two calls, one to convert the `StoreOp`s to `KeyValueStoreOp`s and one to write them. This had the unfortunate side-effect of damaging our atomicity guarantees in case of a write error. If the first call failed, we would be left with the block in fork choice but not on-disk (or the snapshot cache), which would prevent us from processing any descendant blocks. On `unstable` the first call is very unlikely to fail unless the disk is full, but on `tree-states` the conversion is more involved and a user reported database corruption after it failed in a way that should have been recoverable.
Additionally, as @emhane observed, #3696 also inadvertently removed the import of the new block into the block cache. Although this seems like it could have negatively impacted performance, there are several mitigating factors:
- For regular block processing we should almost always load the parent block (and state) from the snapshot cache.
- We often load blinded blocks, which bypass the block cache anyway.
- Metrics show no noticeable increase in the block cache miss rate with v3.4.0.
However, I expect the block cache _will_ be useful again in `tree-states`, so it is restored to use by this PR.
## Issue Addressed
NA
## Proposed Changes
Our `ERRO` stream has been rather noisy since the merge due to some unexpected behaviours of builders and EEs. Now that we've been running post-merge for a while, I think we can drop some of these `ERRO` to `WARN` so we're not "crying wolf".
The modified logs are:
#### `ERRO Execution engine call failed`
I'm seeing this quite frequently on Geth nodes. They seem to timeout when they're busy and it rarely indicates a serious issue. We also have logging across block import, fork choice updating and payload production that raise `ERRO` or `CRIT` when the EE times out, so I think we're not at risk of silencing actual issues.
#### `ERRO "Builder failed to reveal payload"`
In #3775 we reduced this log from `CRIT` to `ERRO` since it's common for builders to fail to reveal the block to the producer directly whilst still broadcasting it to the networ. I think it's worth dropping this to `WARN` since it's rarely interesting.
I elected to stay with `WARN` since I really do wish builders would fulfill their API promises by returning the block to us. Perhaps I'm just being pedantic here, I could be convinced otherwise.
#### `ERRO "Relay error when registering validator(s)"`
It seems like builders and/or mev-boost struggle to handle heavy loads of validator registrations. I haven't observed issues with validators not actually being registered, but I see timeouts on these endpoints many times a day. It doesn't seem like this `ERRO` is worth it.
#### `ERRO Error fetching block for peer ExecutionLayerErrorPayloadReconstruction`
This means we failed to respond to a peer on the P2P network with a block they requested because of an error in the `execution_layer`. It's very common to see timeouts or incomplete responses on this endpoint whilst the EE is busy and I don't think it's important enough for an `ERRO`. As long as the peer count stays high, I don't think the user needs to be actively concerned about how we're responding to peers.
## Additional Info
NA
