## Issue Addressed
Partly resolves https://github.com/sigp/lighthouse/issues/3032
## Proposed Changes
Extracts some of the functionality of #3094 into a separate PR as the original PR requires a bit more work.
Do not unnecessarily penalize peers when we fail to validate received execution payloads because our execution layer is offline.
## Issue Addressed
currently we count a failed attempt for a syncing chain even if the peer is not at fault. This makes us do more work if the chain fails, and heavily penalize peers, when we can simply retry. Inspired by a proposal I made to #3094
## Proposed Changes
If a batch fails but the peer is not at fault, do not count the attempt
Also removes some annoying logs
## Additional Info
We still get a counter on ignored attempts.. just in case
## Issue Addressed
This fixes the low-hanging Clippy lints introduced in Rust 1.61 (due any hour now). It _ignores_ one lint, because fixing it requires a structural refactor of the validator client that needs to be done delicately. I've started on that refactor and will create another PR that can be reviewed in more depth in the coming days. I think we should merge this PR in the meantime to unblock CI.
## Issue Addressed
In very rare occasions we've seen most if not all our peers in a chain with which we don't agree. Purging these peers can take a very long time: number of retries of the chain. Meanwhile sync is caught in a loop trying the chain again and again. This makes it so that we fast track purging peers via registering the failed chain to prevent retrying for some time (30 seconds). Longer times could be dangerous since a chain can fail if a batch fails to download for example. In this case, I think it's still acceptable to fast track purging peers since they are nor providing the required info anyway
Co-authored-by: Divma <26765164+divagant-martian@users.noreply.github.com>
## Issue Addressed
Removes the await points in sync waiting for a processor response for rpc block processing. Built on top of #3029
This also handles a couple of bugs in the previous code and adds a relatively comprehensive test suite.
## Proposed Changes
Initially the idea was to remove hashing of blocks in backfill sync. After considering it more, we conclude that we need to do it in both (forward and backfill) anyway. But since we forgot why we were doing it in the first place, this PR documents this logic.
Future us should find it useful
Co-authored-by: Divma <26765164+divagant-martian@users.noreply.github.com>
## Proposed Changes
Allocate less memory in sync by hashing the `SignedBeaconBlock`s in a batch directly, rather than going via SSZ bytes.
Credit to @paulhauner for finding this source of temporary allocations.
## Issue Addressed
NA
## Proposed Changes
I've observed some Prater nodes (and potentially some mainnet nodes) banning peers due to validator pubkey cache lock timeouts. For the `BeaconChainError`-type of errors, they're caused by internal faults and we can't necessarily tell if the peer is bad or not. I think this is causing us to ban peers unnecessarily when running on under-resourced machines.
## Additional Info
NA
We were batch removing chains when purging, and then updating the status of the collection for each of those. This makes the range status be out of sync with the real status. This represented no harm to the global sync status, but I've changed it to comply with a correct debug assertion that I got triggered while doing some testing.
Also added tests and improved code quality as per @paulhauner 's suggestions.
## Issue Addressed
Running a beacon node I triggered a sync debug panic. And so finally the time to create tests for sync arrived. Fortunately, te bug was not in the sync algorithm itself but a wrong assertion
## Proposed Changes
- Split Range's impl from the BeaconChain via a trait. This is needed for testing. The TestingRig/Harness is way bigger than needed and does not provide the modification functionalities that are needed to test sync. I find this simpler, tho some could disagree.
- Add a regression test for sync that fails before the changes.
- Fix the wrong assertion.
## Description
The `eth2_libp2p` crate was originally named and designed to incorporate a simple libp2p integration into lighthouse. Since its origins the crates purpose has expanded dramatically. It now houses a lot more sophistication that is specific to lighthouse and no longer just a libp2p integration.
As of this writing it currently houses the following high-level lighthouse-specific logic:
- Lighthouse's implementation of the eth2 RPC protocol and specific encodings/decodings
- Integration and handling of ENRs with respect to libp2p and eth2
- Lighthouse's discovery logic, its integration with discv5 and logic about searching and handling peers.
- Lighthouse's peer manager - This is a large module handling various aspects of Lighthouse's network, such as peer scoring, handling pings and metadata, connection maintenance and recording, etc.
- Lighthouse's peer database - This is a collection of information stored for each individual peer which is specific to lighthouse. We store connection state, sync state, last seen ips and scores etc. The data stored for each peer is designed for various elements of the lighthouse code base such as syncing and the http api.
- Gossipsub scoring - This stores a collection of gossipsub 1.1 scoring mechanisms that are continuously analyssed and updated based on the ethereum 2 networks and how Lighthouse performs on these networks.
- Lighthouse specific types for managing gossipsub topics, sync status and ENR fields
- Lighthouse's network HTTP API metrics - A collection of metrics for lighthouse network monitoring
- Lighthouse's custom configuration of all networking protocols, RPC, gossipsub, discovery, identify and libp2p.
Therefore it makes sense to rename the crate to be more akin to its current purposes, simply that it manages the majority of Lighthouse's network stack. This PR renames this crate to `lighthouse_network`
Co-authored-by: Paul Hauner <paul@paulhauner.com>
## Issue Addressed
Closes#1891Closes#1784
## Proposed Changes
Implement checkpoint sync for Lighthouse, enabling it to start from a weak subjectivity checkpoint.
## Additional Info
- [x] Return unavailable status for out-of-range blocks requested by peers (#2561)
- [x] Implement sync daemon for fetching historical blocks (#2561)
- [x] Verify chain hashes (either in `historical_blocks.rs` or the calling module)
- [x] Consistency check for initial block + state
- [x] Fetch the initial state and block from a beacon node HTTP endpoint
- [x] Don't crash fetching beacon states by slot from the API
- [x] Background service for state reconstruction, triggered by CLI flag or API call.
Considered out of scope for this PR:
- Drop the requirement to provide the `--checkpoint-block` (this would require some pretty heavy refactoring of block verification)
Co-authored-by: Diva M <divma@protonmail.com>
## Issue Addressed
N/A
## Proposed Changes
- Removing a bunch of unnecessary references
- Updated `Error::VariantError` to `Error::Variant`
- There were additional enum variant lints that I ignored, because I thought our variant names were fine
- removed `MonitoredValidator`'s `pubkey` field, because I couldn't find it used anywhere. It looks like we just use the string version of the pubkey (the `id` field) if there is no index
## Additional Info
Co-authored-by: realbigsean <seananderson33@gmail.com>
## Issue Addressed
`make lint` failing on rust 1.53.0.
## Proposed Changes
1.53.0 updates
## Additional Info
I haven't figure out why yet, we were now hitting the recursion limit in a few crates. So I had to add `#![recursion_limit = "256"]` in a few places
Co-authored-by: realbigsean <seananderson33@gmail.com>
Co-authored-by: Michael Sproul <michael@sigmaprime.io>
## 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.
The current implementation assumes the range offset of slots downloaded on a batch to equal zero. This conflicts with the condition to consider this chain as sync. For finalized sync, it results in one extra batch being downloaded which can't be processed.
CC @wemeetagain
## Issue Addressed
`test_dht_persistence` failing
## Proposed Changes
Bind `NetworkService::start` to an underscore prefixed variable rather than `_`. `_` was causing it to be dropped immediately
This was failing 5/100 times before this update, but I haven't been able to get it to fail after updating it
Co-authored-by: realbigsean <seananderson33@gmail.com>
This is a little bit of a tip-of-the-iceberg PR. It houses a lot of code changes in the libp2p dependency.
This needs a bit of thorough testing before merging.
The primary code changes are:
- General libp2p dependency update
- Gossipsub refactor to shift compression into gossipsub providing performance improvements and improved API for handling compression
Co-authored-by: Paul Hauner <paul@paulhauner.com>
## Issue Addressed
Following slog's documentation, this should help a bit with string allocations. I left it run for two days and mem usage is lower. This is of course anecdotal, but shouldn't harm anyway
## Proposed Changes
remove `String` creation in logs when possible
## Issue Addressed
- Add metrics to keep track of peer counts by sync type
- Add metric to keep track of the number of syncing chains in range
## Proposed Changes
Plugin to the network metrics update interval and update too the counts for peers wrt to their sync status with us
## Additional Info
For the peer counts
- By the way it is implemented the numbers won't always match to the total peer count in the `libp2p` metric.
- Updating the gauge with every change is messy because it requires to be updated on connection (in the `eth2_libp2p` crate, while metrics are defined in the `network` crate) on Goodbye sent (for an `IrrelevantPeer`) either in the `beacon_processor` or the `peer_manager`, and on disconnection. Since this is not a critical metric I think counting once every second is enough. If you think more accuracy is needed we can do it too, but it would be harder to maintain)
ATM those look like this
## Issue Addressed
Two issues related to empty batches
- Chain target's was not being advanced when the batch was successful, empty and the chain didn't have an optimistic batch
- Not switching finalized chains. We now switch finalized chains requiring a minimum work first
## Issue Addressed
we have a log saying we add a peer to a chain, and an another one in case the chain is not syncing. To avoid needing to peer there two (and reduce log entries) simply log the chain's syncing state in the chain's KV
## Issue Addressed
`BlocksByRange` requests were the main culprit of a series of timeouts to peer's requests in general because they produce build up in the router's processor. Those were moved to the blocking executor but a task is being spawned for each; also not ideal since the amount of resources we give to those is not controlled
## Proposed Changes
- Move `BlocksByRange` and `BlocksByRoots` to the `beacon_processor`. The processor crafts the responses and sends them.
- Move too the processing of `StatusMessage`s from other peers. This is a fast operation but it can also build up and won't scale if we keep it in the router (processing one at the time). These don't need to send an answer, so there is no harm in processing them "later" if that were to happen. Sending responses to status requests is still in the router, so we answer as soon as we see them.
- Some "extras" that are basically clean up:
- Split the `Worker` logic in sync methods (chain processing and rpc blocks), gossip methods (the majority of methods) and rpc methods (the new ones)
- Move the `status_message` function previously provided by the router's processor to a more central place since it is used by the router, sync, network_context and beacon_processor
- Some spelling
## Additional Info
What's left to decide/test more thoroughly is the length of the queues and the priority rules. @paulhauner suggested at some point to put status above attestations, and @AgeManning had described an importance of "protecting gossipsub" so my solution is leaving status requests in the router and RPC methods below attestations. Slashings and Exits are at the end.
## Issue Addressed
#1856
## Proposed Changes
- For clarity, the router's processor now only decides if a peer is compatible and it disconnects it or sends it to sync accordingly. No logic here regarding how useful is the peer.
- Update peer_sync_info's rules
- Add an `IrrelevantPeer` sync status to account for incompatible peers (maybe this should be "IncompatiblePeer" now that I think about it?) this state is update upon receiving an internal goodbye in the peer manager
- Misc code cleanups
- Reduce the need to create `StatusMessage`s (and thus, `Arc` accesses )
- Add missing calls to update the global sync state
The overall effect should be:
- More peers recognized as Behind, and less as Unknown
- Peers identified as incompatible
## Issue Addressed
Fixes head syncing
## Proposed Changes
- Get back to statusing peers after removing chain segments and making the peer manager deal with status according to the Sync status, preventing an old known deadlock
- Also a bug where a chain would get removed if the optimistic batch succeeds being empty
## Additional Info
Tested on Medalla and looking good
## Issue Addressed
Sync edge case when we get an empty optimistic batch that passes validation and is inside the download buffer. Eventually the chain would reach the batch and treat it as an ugly state.
## Proposed Changes
- Handle the edge case advancing the chain's target + code clarification
- Some largey changes for readability + ergonomics since rust has try ops
- Better handling of bad batch and chain states
## Description
This increases the logging of the underlying UPnP tasks to inform the user of UPnP error/success.
This also decreases the batch syncing size to two epochs per batch.
check for advanced peers and the state of the chain wrt the clock slot to decide if a chain is or not synced /transitioning to a head sync. Also a fix that prevented getting the right state while syncing heads
## Issue Addressed
#1614 and a couple of sync-stalling problems, the most important is a cyclic dependency between the sync manager and the peer manager
Squashed commit of the following:
commit f99373cbaec9adb2bdbae3f7e903284327962083
Author: Age Manning <Age@AgeManning.com>
Date: Mon Oct 5 18:44:09 2020 +1100
Clean up obsolute TODOs
## Issue Addressed
Downgrade inconsistent chain segment states from `panic` to `crit`. I don't love this solution but since range can always bounce back from any of those, we don't panic.
Co-authored-by: Age Manning <Age@AgeManning.com>
## Issue Addressed
chain state inconsistencies
## Proposed Changes
- a batch can be fake-failed by Range if it needs to move a peer to another chain. The peer will still send blocks/ errors / produce timeouts for those requests, so check when we get a response from the RPC that the request id matches, instead of only the peer, since a re-request can be directed to the same peer.
- if an optimistic batch succeeds, store the attempt to avoid trying it again when quickly switching chains. Also, use it only if ahead of our current target, instead of the segment's start epoch
## Issue Addressed
In principle.. closes#1551 but in general are improvements for performance, maintainability and readability. The logic for the optimistic sync in actually simple
## Proposed Changes
There are miscellaneous things here:
- Remove unnecessary `BatchProcessResult::Partial` to simplify the batch validation logic
- Make batches a state machine. This is done to ensure batch state transitions respect our logic (this was previously done by moving batches between `Vec`s) and to ease the cognitive load of the `SyncingChain` struct
- Move most batch-related logic to the batch
- Remove `PendingBatches` in favor of a map of peers to their batches. This is to avoid duplicating peers inside the chain (peer_pool and pending_batches)
- Add `must_use` decoration to the `ProcessingResult` so that chains that request to be removed are handled accordingly. This also means that chains are now removed in more places than before to account for unhandled cases
- Store batches in a sorted map (`BTreeMap`) access is not O(1) but since the number of _active_ batches is bounded this should be fast, and saves performing hashing ops. Batches are indexed by the epoch they start. Sorted, to easily handle chain advancements (range logic)
- Produce the chain Id from the identifying fields: target root and target slot. This, to guarantee there can't be duplicated chains and be able to consistently search chains by either Id or checkpoint
- Fix chain_id not being present in all chain loggers
- Handle mega-edge case where the processor's work queue is full and the batch can't be sent. In this case the chain would lose the blocks, remain in a "syncing" state and waiting for a result that won't arrive, effectively stalling sync.
- When a batch imports blocks or the chain starts syncing with a local finalized epoch greater that the chain's start epoch, the chain is advanced instead of reset. This is to avoid losing download progress and validate batches faster. This also means that the old `start_epoch` now means "current first unvalidated batch", so it represents more accurately the progress of the chain.
- Batch status peers from the same chain to reduce Arc access.
- Handle a couple of cases where the retry counters for a batch were not updated/checked are now handled via the batch state machine. Basically now if we forget to do it, we will know.
- Do not send back the blocks from the processor to the batch. Instead register the attempt before sending the blocks (does not count as failed)
- When re-requesting a batch, try to avoid not only the last failed peer, but all previous failed peers.
- Optimize requesting batches ahead in the buffer by shuffling idle peers just once (this is just addressing a couple of old TODOs in the code)
- In chain_collection, store chains by their id in a map
- Include a mapping from request_ids to (chain, batch) that requested the batch to avoid the double O(n) search on block responses
- Other stuff:
- impl `slog::KV` for batches
- impl `slog::KV` for syncing chains
- PSA: when logging, we can use `%thing` if `thing` implements `Display`. Same for `?` and `Debug`
### Optimistic syncing:
Try first the batch that contains the current head, if the batch imports any block, advance the chain. If not, if this optimistic batch is inside the current processing window leave it there for future use, if not drop it. The tolerance for this block is the same for downloading, but just once for processing
Co-authored-by: Age Manning <Age@AgeManning.com>
## Issue Addressed
N/A
## Proposed Changes
Prevent subscribing to core gossipsub topics until after we have achieved a full sync. This prevents us censoring gossipsub channels, getting penalised in gossipsub 1.1 scoring and saves us computation time in attempting to validate gossipsub messages which we will be unable to do with a non-sync'd chain.
## Issue Addressed
#1590
## Proposed Changes
This is a temporary workaround that prevents finalized chain sync from swapping chains. I'm merging this in now until the full solution is ready.
## Issue Addressed
#1172
## Proposed Changes
* updates the libp2p dependency
* small adaptions based on changes in libp2p
* report not just valid messages but also invalid and distinguish between `IGNORE`d messages and `REJECT`ed messages
Co-authored-by: Age Manning <Age@AgeManning.com>
## Description
There can be many head chains queued up to complete. Currently we try and process all of these to completion before we consider the node synced.
In a chaotic network, there can be many of these and processing them to completion can be very expensive and slow. This PR removes any non-syncing head chains from the queue, and re-status's the peers. If, after we have synced to head on one chain, there is still a valid head chain to download, it will be re-established once the status has been returned.
This should assist with getting nodes to sync on medalla faster.
## Description
Currently lighthouse load-balances across peers a single finalized chain. The chain is selected via the most peers. Once synced to the latest finalized epoch Lighthouse creates chains amongst its peers and syncs them all in parallel amongst each peer (grouped by their current head block).
This is typically fast and relatively efficient under normal operations. However if the chain has not finalized in a long time, the head chains can grow quite long. Peer's head chains will update every slot as new blocks are added to the head. Syncing all head chains in parallel is a bottleneck and highly inefficient in block duplication leads to RPC timeouts when attempting to handle all new heads chains at once.
This PR limits the parallelism of head syncing chains to 2. We now sync at most two head chains at a time. This allows for the possiblity of sync progressing alongside a peer being slow and holding up one chain via RPC timeouts.
The changes are somewhat simple but should solve two issues:
- When quickly changing between chains once and a second time back again, batchIds would collide and cause havoc.
- If we got an out of range response from a peer, sync would remain in syncing but without advancing
Changes:
- remove the batch id. Identify each batch (inside a chain) by its starting epoch. Target epochs for downloading and processing now advance by EPOCHS_PER_BATCH
- for the same reason, move the "to_be_downloaded_id" to be an epoch
- remove a sneaky line that dropped an out of range batch without downloading it
- bonus: put the chain_id in the log given to the chain. This is why explicitly logging the chain_id is removed
## Issue Addressed
NA
## Proposed Changes
Moves beacon block processing over to the newly-added `GossipProcessor`. This moves the task off the core executor onto the blocking one.
## Additional Info
- With this PR, gossip blocks are being ignored during sync.
## Issue Addressed
Sync was breaking occasionally. The root cause appears to be identify crashing as events we being sent to the protocol after nodes were banned. Have not been able to reproduce sync issues since this update.
## Proposed Changes
Only send messages to sub-behaviour protocols if the peer manager thinks the peer is connected. All other messages are dropped.
## Issue Addressed
Recurring sync loop and invalid batch downloading
## Proposed Changes
Shifts the batches to include the first slot of each epoch. This ensures the finalized is always downloaded once a chain has completed syncing.
Also add in logic to prevent re-dialing disconnected peers. Non-performant peers get disconnected during sync, this prevents re-connection to these during sync.
## Additional Info
N/A
