When a late block arrives and the beacon is proposing the next block, we
perform several checks to allow for the next block to reorg the incoming
late block.
Among those checks, we check that the parent block has been heavily
attested (currently 160% of the committee size).
We perform this check in these circumstances:
- When the late block arrives
- At 10 seconds into the slot
- At 0 seconds into the next slot (at proposing time)
The problem is that for blocks that arrive between 4 seconds and 10
seconds, the parent block will not have yet this expected weight since
attestations from the current committee were not imported yet, and thus
Prysm will send an FCU with payload attributes anyway at this time.
What happens is that Prysm keeps the EL building different blocks based
on different parents at the same time, when later in the next slot it
calls to propose, it will reorg the late block anyway and the EL would
have been computing a second payload uselessly.
This PR enables this check only when calling `ShouldOverrideFCU` after
10 seconds into the slot which we do only after having imported the
current attestations. We may want to actually remove this check entirely
from `ShouldOverrideFCU` and only keep it in `ProposerHead`.
Shout out to Anthithesis for reporting an issue that led to this
discoverly.
* downgrade level db
* fix current issues
* update geth
* Fix zstd build. The patch is now longer needed now that https://github.com/bazelbuild/rules_go/issues/3411 is fixed.
* Revert "update geth"
This reverts commit 2a7c51a952676a36ed62f837e91bebce2413e731.
* change to hex
---------
Co-authored-by: Preston Van Loon <pvanloon@offchainlabs.com>
* employ a val reader to prevent constant copies
* clean it up and fix tests
* gaz
* radek's review
---------
Co-authored-by: Radosław Kapka <rkapka@wp.pl>
* avoid part path collisions with mem addr entropy
* Regression test
---------
Co-authored-by: Kasey Kirkham <kasey@users.noreply.github.com>
Co-authored-by: Preston Van Loon <pvanloon@offchainlabs.com>
* Improve package documentation.
* `processAttestations`: Improve logging.
* Add `Benchmark_checkSurroundVotes` benchmark.
* Implement `saveChunksToDisk` as remplacement of `saveUpdatedChunks`.
The idea is to open only on DB transaction for all validator chunk indexes instead of
one DB transaction per validator chunk index.
It saves the overhead due to transaction start/stop of the DB.
Result of `Benchmark_checkSurroundVotes`:
- Before this commit: 133 seconds
- After this commit: 5.05 seconds
* `LoadSlasherChunks` and `SaveSlasherChunks`: Batch.
* `loadChunks` ==> `loadChunksFromDisk`
* `updatedChunkByChunkIndex`: Don't update if `latestEpochWritten == currentEpoch `.
* `updatedChunkByChunkIndex`: Load all needed chunks once.
* `latestEpochWritten` ==> `latestEpochUpdated`.
* `checkSurroundVotes`: Dump to disk at most every `25_600` chunks.
* `SaveAttestationRecordsForValidators`: Batch.
* `batchSize`: Use as package const and add comment.
* pending broadcast err if missing blobs and 0 peers
* compute request first for len check
---------
Co-authored-by: Kasey Kirkham <kasey@users.noreply.github.com>
* `getChunk` ==> `getChunkFromDatabase`.
* `loadChunks`: Rename variables.
* `Update`: Use explicit arguments.
* `detect_attestations.go`: Reduce abstraction layers.
* `loadAndUpdateChunks`: Change arguments order.
* `updatedChunkByChunkIndex`: Update all known validators in the chunk.
* `LastEpochWrittenForValidators`: Avoid avoidable `for`loop.
* `chunks.go`: Ensure implementations respect the interface.
* `LastEpochWrittenForValidators`: Stop considering lack of epoch as genesis epoch.
* `updatedChunkByChunkIndex`: Don't update latest updated epoch.
And add a bunch of tests.
* Improve slasher cold boot duration.
Before this commit, on a slasher cold boot (aka, without any db),
the `updatedChunkByChunkIndex` function looped for all validators
AND for all epochs between the genesis epoch and the current epoch.
This could take several dozen of minutes, and it is useless since the
min/max spans are actually a circular buffer with a limited lenght.
Cells of min/max spans can be overwritten (with the same value)
plenty of times.
After this commit, the `updatedChunkByChunkIndex` function loops
for all validators AND AT most 'historyLength' lenght.
Every cell of min/max spans are written AT MOST once.
Time needed for slasher boot goes from `O(nm)` to "only" `O(m)`, where:
- `n` is the number of epochs since the genesis.
- `m` is the number of validators.
* First take at updating everything to v5
* Patch gRPC gateway to use prysm v5
Fix patch
* Update go ssz
---------
Co-authored-by: Preston Van Loon <pvanloon@offchainlabs.com>
* retrieve and save blobs during backfill
* Update beacon-chain/sync/backfill/batch.go
Co-authored-by: Preston Van Loon <pvanloon@offchainlabs.com>
* Update beacon-chain/sync/backfill/blobs.go
Co-authored-by: Preston Van Loon <pvanloon@offchainlabs.com>
* Update beacon-chain/sync/backfill/metrics.go
Co-authored-by: Preston Van Loon <pvanloon@offchainlabs.com>
* make blobSync initialization path a little safer
* use bytes.Equal to avoid extra allocation
* stop using log.Fatal and other little cleanups
* wrap up post blob sync actions in batch mutator
* unit test coverage on verifier
---------
Co-authored-by: Kasey Kirkham <kasey@users.noreply.github.com>
Co-authored-by: Preston Van Loon <pvanloon@offchainlabs.com>
