Repository navigation
Erigon v3.8.2 — Watery Whiskers
Erigon 3.8.2 fixes valid blocks rejected by parallel execution and parallel commitment, speeds up block execution and Caplin head updates, and improves Caplin peering, sync committee and data column gossip. It also fixes several security issues in peer, RPC and Engine API handling.
This release is recommended for all users. Nodes can upgrade from 3.8.1 without a re-sync. Engine API monitoring, seg du scripts and Caplin operators should read the breaking changes and upgrade notes below.
Highlights
- Valid blocks no longer rejected. Parallel execution no longer rejects a valid block with
invalid state root hashwhen one transaction touches an empty account that holds storage and a later transaction in the same block funds it. Parallel commitment no longer rejects a valid block withempty branch data read during unfold(#25375, #25814) — by @chfast, @awskii - Faster block execution. Parallel execution no longer re-runs transactions from the same sender again and again. On mainnet bursts of 1,400–2,700-transaction blocks, p95 time per block drops from about 920 ms to 100–180 ms (#24500) — by @awskii
- Faster Caplin head updates. The next-epoch state is precomputed outside the fork-choice lock, so the first block of each epoch no longer blocks head updates for about 0.5 s.
headevents are sent about 90 ms earlier, andattestation_dataanswers from fork choice at once instead of waiting up to 1 s (#25473, #25543, #25359) — by @domiwei - More validator votes reach the network. One per-peer gossip limit was shared by all sync subnets, so Caplin ignored and did not relay about 69% of sync committee messages; each subnet now has its own allowance. Gossip attestations and aggregates are checked against the committees of their target epoch, so valid votes from another branch are no longer rejected and their senders are no longer penalised (#25593, #25208) — by @lystopad, @domiwei
- Data columns from the local EL. From Fulu, Caplin builds a block's custody data columns from the blobs the execution layer already has (
engine_getBlobsV2), instead of waiting for all columns to arrive over gossip (#25447) — by @lystopad - Engine API during catch-up. When execution lags a payload by more than one state step can run,
engine_newPayloadanswersACCEPTEDinstead of a JSON-RPC error, andlatestValidHashis null unless the status isVALIDorINVALID(#25382) — by @awskii - Reorg notifications.
eth_subscribenow notifies on reorgs:newHeadsandtransactionReceiptsget the new block, andlogssubscribers getremoved: truefor dropped logs (#25344) — by @domiwei
Breaking Changes
- Engine API batch limit.
--rpc.batch.limit(default 100) now also applies on the authenticated Engine API port; a client that sends larger Engine batches gets a batch-limit error (#25872) — by @lupin012 - Engine API
/healthrequires a JWT./healthon the Engine API port answers403without a valid JWT; update monitoring that probes it without a token./healthrequest bodies are limited to 4 KiB. The public HTTP port is unchanged (#25871) — by @lupin012 - Trace output.
trace_callandtrace_callManyreturn-32001for unknown blocks, reward traces carry a nulltransactionHashandtransactionPosition, andvmTracedrops the Erigon-onlyidxfield (#25030) — by @banteg payload_attributesSSE event. Caplin uses the Beacon API encoding (snake_case fields, quoted decimals) for all forks (#25292) — by @domiweiseg duoutput. Estimates use the downloader's file selection, so they match what a node of each prune mode downloads. The JSON fielddetected_modeand the detected mode in the header are removed;estimatesis null when headers or bodies cannot be read or are not fully indexed; the archive estimate no longer counts commitment history, which a node keeps only with execution proofs; commitment accessor files count under "commitment hist". Scripts that readdetected_modeorestimatesneed an update (#25960) — by @awskii
Upgrade notes
- Nodes stuck on an Amsterdam block after snapshot sync. Full and minimal nodes now keep the history of the block that execution resumes in after snapshot sync; without it, initial sync could stop on an Amsterdam block on every start. A node that is already stuck there must run
erigon snapshots resetor start from a new datadir (#25833) — by @chfast traces_toindex. Rebuilding it with the custom trace stage now includes coinbase and uncle reward recipients; an index rebuilt this way on 3.8 must be rebuilt again (#25476) — by @Ravan002- Caplin disk use. Non-archive Caplin keeps the blocks of the spec window plus every block since finality; on mainnet, Sepolia and Hoodi this is about 56,800 more slots of disk (#25412) — by @domiwei
- Caplin gossip mesh. The gossipsub mesh degree is the spec value (8/6/12): block-to-head p50 drops from 156 to 132 ms, while Caplin traffic grows from 129/107 to 152/184 KB/s in/out (#25674) — by @domiwei
- Caplin outbound peers. 20% of peer slots (26 at the default 128) are reserved for outbound peers that Caplin chooses itself. After the upgrade, a node with more than 114 inbound peers drops inbound peers down to 102 in one step (#25943) — by @domiwei
Fixes and Improvements
Execution and Engine API
- Parallel execution and unwind correctness. A rare race could validate a stale account balance, nonce or code hash and end with a different state than serial execution; the value is now re-read. After an unwind, a self-destruct and re-create in the same window no longer skips a live storage slot (#24874, #25667) — by @chfast, @AskAlexSharov
- Receipt snapshots. New receipt-cache history snapshots replace published files in which some receipts had
cumulativeGasUsedandlogIndexreset to 0 at step edges (#25271) — by @AskAlexSharov - Engine API. During catch-up,
engine_newPayloadno longer answersVALIDfor canonical blocks that execution has not reached yet. Re-sent payloads that were already validated are answered without an unwind or a second execution, and a bad block found during block download is remembered so the next forkchoice update does not download the same chain again (#25566, #25109, #25116, #25589) — by @AskAlexSharov, @awskii
RPC
eth_simulateV1. On a historical block, writes resolve against the simulated state instead of the head;traceTransfersno longer leaves gaps in log indices after a failed or reverted call; unknownblockOverridesandstateOverrideskeys and a post-mergedifficultyoverride are rejected with-32602(#25687, #24932, #25018, #25017) — by @AskAlexSharov, @MysticRyuujin, @lupin012- Historical reads and tracing. Range reads no longer miss data at the end of snapshot files,
debug_trace*no longer reports storage forSLOAD/SSTOREsteps that halted before running, andeth_getLogsreturns an error instead of silently dropping the logs of a transaction it cannot read (#25662, #24987, #25320) — by @AskAlexSharov, @MysticRyuujin, @lupin012 - Stricter inputs. Call requests that set
maxFeePerBlobGaswithout blob hashes are rejected, anddebug_getModifiedAccountsByHashrejects non-canonical block hashes (#24990, #25574) — by @banteg, @AskAlexSharov
Database and snapshots
- Robustness. Error paths no longer leak temp files, compression workers or open segments; a malformed preverified TOML no longer crashes the node; shutdown during block retire no longer hangs; a corrupt or truncated
.segfile, or one with a truncated dictionary, returns an error instead of crashing or hanging later; chaindata written by a newer Erigon prints an actionable error (#25666, #25413, #25561, #25906, #25501) — by @AskAlexSharov, @awskii, @Ravan002 - Caplin torrents. Block merges no longer remove the torrents of Caplin block and blob sidecar files (#24998) — by @awskii
Consensus layer (Caplin)
- Fork choice and block import performance. Vote weights are kept across
get_headcalls and votes are applied when they are recorded; Fulu column availability is read before taking the fork-choice lock; BLS public key aggregation is about 13–16 ms faster per mainnet block; participation lists are hashed in parallel; each beacon state uses about 27–53 MB less memory on mainnet; the chain-tip wait ends as soon as a block is indexed (#24985, #25595, #25594, #25542, #25596, #25647, #26330) — by @awskii, @domiwei - Fork choice correctness. Attestations inside synced and self-proposed blocks are counted,
is_head_weaktakes its committees from the head epoch, anINVALIDverdict with alatestValidHashalso invalidates the optimistic ancestors back to it, and beacon blocks are persisted only after fork choice accepts them (#25281, #25988, #25989, #25442) — by @domiwei, @yperbasis - Gossip and peering. Sync committee messages are relayed only after their signature is verified; per-peer gossip rate limits no longer reset every 10 s; seen attestations are recorded per target epoch for every validator; the attestation subnet search keeps looking for peers until each subnet has 6 (the gossipsub mesh low watermark) instead of 4; discovery dials only nodes that advertise our fork digest; a chain-tip block with unknown ancestors fetches them from the peer that served it; a forward sync batch cancelled at shutdown no longer bans its peers; large data column responses are no longer reset at the 5 s stream deadline (#25605, #25039, #26275, #25795, #25894, #25893, #25765, #26209, #26288) — by @lystopad, @awskii, @domiwei
- Block production. While a local validator's proposal is near, Caplin subscribes to every attestation subnet so its block can include votes from every subnet. Block building no longer produces empty blocks again and again when the top pending transactions all have a nonce too high, a forkchoice update rejected by a busy execution layer is no longer treated as applied, and block production for a past slot is refused (#26250, #25334, #25395, #25189) — by @lystopad, @domiwei
- Data columns. Columns built from the EL's blobs are stored and published to gossip as the Fulu spec requires, skip re-verifying their cell proofs, have their cells computed in parallel, and are written in one batch with parallel fsync, as are the columns of a self-built block and of a posted envelope. A self-built block stores only its custody columns after import, and a failed column write no longer prevents its import. A block that waits for its columns is retried as soon as they are stored. Supernodes skip FK20 recovery when the matrix is complete and fetch missing columns by root before recovering blocks with 9 or more blobs. Valid sidecars outside custody are relayed and reported in
data_column_sidecarevents, and custody columns of side branches and of branches that become canonical after a reorg are backfilled (#25522, #25510, #25514, #25926, #26347, #25540, #25654, #26050, #26331, #25587, #25681, #26072, #25931, #26287, #25300) — by @lystopad, @domiwei, @awskii - Beacon API. Archive nodes return correct epoch-2 committees, duties, participation and rewards; pending deposit, partial withdrawal and consolidation endpoints no longer return
404for recent finalized slots; every invalid block or state id is answered with400, and slot ids of 2^32 or more no longer alias lower slots; full headvalidatorsresponses no longer delay head-state updates, and requests filtered by many ids are much faster; plus fixes to state fork, expected withdrawals, RANDAO, config spec, finalized slot, epoch root and reward bounds,validator_inclusionandheadevent endpoints (#25296, #25324, #26144, #25660, #25619, #26145, #26150, #26151, #26152, #26161, #26184, #26185, #26160, #26155) — by @domiwei, @lystopad, @Snan0M, @awskii - Light client.
light_client/updatesreturns at mostmin(count, 128)updates and the earliest consecutive run of periods when one is missing; the best verified update of each period is kept in its attested header's fork and served from the head's chain before finalization; candidates are served only on the finalized chain; an update is built even when the finalized block is not in memory, andlight_client/finality_updateserves the newest update that has finality (#25032, #26021, #26114, #26147, #26190) — by @Ravan002, @lystopad - Gloas execution payload envelopes. A posted envelope publishes its data columns before storing them and verifies blob bundles per blob in parallel; with
broadcast_validation=gossipit is published before its columns; work goes on after the client disconnects, and a re-POST is answered from state instead of400; with consensus validation, an envelope that fails validation publishes nothing; history backfill bans peers that serve invalid envelopes (#25710, #25718, #25755, #25767, #25528) — by @domiwei, @yperbasis
Security
- Execution-layer peers. Each peer's outbound queue is bounded by bytes; the eth/69
Statusblock range is validated during the handshake; inboundGetBlockHeadersrequests are rate-limited; eachGetReceiptsrequest gets a re-execution budget (#24977, #25844, #25732, #25569) — by @yperbasis, @Bruce039, @AskAlexSharov. Reported by @mardgez13, @duongynhi000005-oss - Discovery endpoint prediction. The predictor that fills the local node record's IP and UDP port counts endpoint statements once per reporter network (IPv4 /24, IPv6 /48; LAN hosts individually), ignores statements about the other address family, and no longer scans every statement or handshake per packet. A prediction needs statements from 10 networks within 5 minutes; a node behind NAT without UPnP or NAT-PMP that sees fewer keeps its record without a predicted endpoint, and can set it with
--nat extip:<IP>(#24975) — by @mardgez13. Reported by @mardgez13 - Transaction pool. The blob fee is checked on every promotion to pending, blob pool limits are checked before KZG proof verification, and inbound transaction batches are flushed at 256 messages or 4 MiB (#25878, #25356, #25576) — by @lystopad, @AskAlexSharov, @joxerx. Reported by @duongynhi000005-oss
- Caplin peer requests and gossip. A request
countthat overflowed the rate-limit cost no longer bypasses the per-peer limit, and duplicate roots are served once;BeaconBlocksByRootrequests above 128 roots are rejected; a fixed-size request with an oversized length prefix is refused;DataColumnSidecarsByRangescans are bounded; Fulu data column sidecars are checked against the expected proposer; voluntary exits posted to the Beacon API are gossiped only after verification; a light client update is stored only after its block passes the state transition; the built-in execution path rejects RLP-wrapped typed transactions (#25809, #25723, #25918, #26001, #25684, #26279, #26021, #25529) — by @yperbasis, @Snan0M, @AskAlexSharov, @lystopad, @domiwei. Reported by @duongynhi000005-oss - RPC and Engine API. JSON-RPC batches over
--rpc.batch.limitare rejected before their messages are allocated;trace_rawTransactionbounds EIP-7702 authorization lists before decoding;debug_storageRangeAtwith a hugemaxResultno longer exhausts memory; with--graphql, the Engine API port no longer serves/graphqlbefore the JWT check; the embedded MCP server exposes only the namespaces enabled by--http.api;engine_newPayloadrejects a payload with more transactions than its gas limit allows before decoding or execution (#25873, #24966, #25348, #25352, #25264, #25176) — by @lupin012, @AskAlexSharov, @chfast - Downloader and dependencies. Snapshot names that escape the snapshot directory are rejected.
anacrolix/torrentv1.63.0 fixes a remote crash through theut_holepunchextension and a crash on hybrid torrents (GHSA-c5j3-3rp9-c4rv, GHSA-6qq7-r4c9-cf54) and a bencode recursion crash reachable from the downloader;golang.org/x/netis updated. The Docker image distrusts the CA roots behind the eMudhra intermediate key compromise (#25562, #26433, #26007, #25401) — by @AskAlexSharov, @anacrolix, @lystopad
Full Changelog: v3.8.1...v3.8.2