github zakura-core/zakura v1.9.0-rc0

pre-release4 hours ago

Release candidate for 1.9.0. Faster block relay, optional download hedging, and mining fixes. Check your config before upgrading.

Supported on Mainnet until block 3,643,309, around December 3, 2026. Upgrade warnings start about three days earlier.

Before you upgrade

  • Update trace consumers. Update trace parsers from peer_id to connection_id. Addresses in legacy_peer_request.jsonl now follow network.expose_peer_addresses; on macOS, raw trace requests for Unix sockets return HTTP 403 (#2455, #2534).

Highlights

  • Mining nodes push new blocks to initial announcement peers advertising NODE_BLOCK_PUSH, saving one request round trip; opt-in network.push_relayed_blocks also pushes committed relayed blocks but can send duplicates (#2537).
  • Enable rpc.enable_trusted_block_submission to relay through /submittrustedblock before local validation; submit only final block bytes verified by a trusted full node, or peers may ban your node (#2534).
  • Opt-in [network.block_download_score_and_hedge] with enabled = true selects legacy peers using recent response times and can send one backup request when another peer looks faster or the first request stalls (#2572, #2595, #2596).
  • The internal miner switches from coinbase-only templates to templates with transactions on the same tip and stops solving replaced templates after an Equihash attempt finds no solutions (#2586).
  • Windows nodes reuse the existing state database after a major database format upgrade, and pruned nodes stop downloading already-finalized old blocks when peers announce them (#2563, #2612).
  • Experimental: [mempool] implementation = "core" selects the new pool core, with metrics for pooled transactions and reset recovery; mempool.implementation reports the selected pool. default and classic keep the current pool (#2502, #2506).
Full changelog

Added

  • Mining nodes push newly mined blocks to peers in the initial announcement
    group that advertise NODE_BLOCK_PUSH, saving one request round trip.
    The opt-in network.push_relayed_blocks setting also pushes blocks this
    node relays after it commits them, which can send an extra copy to a
    peer that already has the block
    (#2537).
  • Added the binary /submittrustedblock endpoint, enabled through
    rpc.enable_trusted_block_submission. It can start relaying a new block
    before local validation and follows the RPC listener's cookie-authentication
    setting. Submit only final block bytes verified by a trusted full node,
    because peers can ban a relay that sends an invalid block. Zakura withdraws
    pending announcements and responses when a submission is rejected or
    times out. The endpoint returns inconclusive when it cannot establish
    the result, so the caller can retry
    (#2534).
  • Added the opt-in [network.block_download_score_and_hedge] setting.
    With enabled = true, Zakura picks a legacy peer using recent response
    times and can send one backup request when a later announcer is likely
    to deliver sooner or the first request stalls. Peer and request events
    can trigger the backup. If you tried the feature in a development build,
    replace [network.block_source_score] and remove its second_request
    key
    (#2572,
    #2595,
    #2596).
  • Added the [mempool] implementation setting. default and classic keep
    the pool nodes run today. core runs the new pool core, which is
    experimental
    (#2502).
  • Added tx-relay, an experimental v2 P2P service that moves transactions
    between Zakura peers and hands them to the pool core. It is off by
    default. It runs only with [network.zakura.tx_relay] enabled = true, the
    v2 stack on, and [mempool] implementation = "core". While a peer has a
    live tx-relay session, the legacy adapter leaves that peer's
    transactions to it. New zakura.p2p.tx_relay.* metrics count what it
    moves, and the existing zakura.p2p.legacy.inbound.refused counter gains
    a type label. The experimental advert_delay (2 s),
    advert_window_full_tx (4) and advert_window_small_tx (16) keys tune
    its fan-out
    (#2503).
  • Added mempool metrics for canaries: the pool core exports how many
    transactions it pools and how many a reset took out that are not back
    yet, and mempool.implementation shows which pool the node resolved to
    (#2506).
  • Added zakura.p2p.tx_relay.registered, which says whether the node
    registered tx-relay
    (#2506).

Changed

  • With trace_dir set, the legacy request rows in legacy_peer_request.jsonl now follow
    network.expose_peer_addresses: they hold the peer's address when it is set and stay
    redacted when it is not. Each row also carries the peer's user agent, and a peer upgraded
    to Zakura gets one peer_upgraded row in handshake.jsonl with its trace label, its
    address under the same setting, and its user agent. The rows' peer_id column is now
    connection_id, the same number the connection's legacy_gossip.jsonl rows carry
    (#2455).
  • Serving a block to several legacy peers uses less CPU by reusing its
    serialized bytes and sharing its checksum across requests
    (#2556,
    #2557).
  • When Zakura is still verifying a locally submitted block, it waits to
    download a peer's copy until local verification fails or five seconds
    pass. The wait leaves the peer-download slot available. Its duration
    and outcome are recorded in metrics and, when enabled, in
    commit_state.jsonl
    (#2562).
  • The experimental core mempool builds each snapshot from the previous one,
    reducing CPU use as the pool fills
    (#2591).
  • The pool.route_inv.hedge.not_yet.count and
    pool.route_inv.hedge.no_candidate.count metrics count decisions to send
    a backup block request. pool.route_inv.hedge.held shows how many
    downloads are waiting for one
    (#2596).
  • A stalled v2 frame read now resets its session or one-shot request. Sibling
    services can keep using the connection. Header sync counts an unanswered
    subscription Close from the completion of its write. A second header-sync or
    block-sync stall on the same connection can close it without a peer fault
    (#2604).
  • Inbound v2 handshakes now share a node-wide rate of 24 starts a second for
    each source validation state. The last 8 tokens go only to an address within
    its share of 3 starts a second, so a few busy addresses can't refuse the rest.
    The 10-second handshake deadline now starts when the connection is accepted.
    Every TLS start keeps its node and address charges
    (#2604).
  • Discovery dispatch uses a wall-clock budget even when its receive queue fills.
    Header sync meters peer Status messages and locator walks for new
    subscriptions per connection, so reopening a session gets no fresh burst
    (#2604).
  • The experimental tx-relay service uses less memory with many connected
    peers. It sizes each peer's request window based on recent deliveries
    and announces new transactions to the lowest-latency peers first
    (#2623).

Fixed

  • Count inbound TLS attempts and control handshakes against one pending budget,
    so outbound dials cannot exceed the limit while inbound TLS attempts remain
    pending (#2337).
  • Limit the entire v2 control handshake to ten seconds in both directions, so
    peers cannot extend an inbound claim or a dial's pending-handshake permit by
    pacing successive control steps
    (#2337).
  • Close v2 connections neutrally when their handshake captures an obsolete
    capability generation, so an in-flight handshake cannot miss header-sync
    renegotiation (#2337).
  • Limit failed inbound v2 handshake attempts per address with a refundable
    cadence derived from the connection limit and dialer backoff: sixteen tokens
    and sixteen tokens per second by default. Validated and unvalidated sources
    have separate ledgers, so spoofed Initials can't lock out an honest
    address. An empty bucket answers an unvalidated source with Retry and a
    validated one with Refuse, counted by zakura.p2p.conn.retried.cadence
    and zakura.p2p.conn.rejected.cadence
    (#2337).
  • Drop a closed v2 QUIC connection's state within three seconds, so a peer
    with a long or delayed return path cannot hold a closing-connection slot for
    three of its probe timeouts
    (#2337).
  • Nodes start successfully when ZAKURA_TRACE_PEER_NAMES names fleet peers for
    the gossip trace, instead of rejecting it as an unknown configuration field
    (#2455).
  • The mempool no longer remembers a transaction as rejected when its
    verification failed on this node, such as a failed median-time lookup, so
    resending it before the next block can admit it
    (#2502).
  • On macOS, the raw trace endpoint returns HTTP 403 when a requested table is a
    Unix socket (#2534).
  • Preserve a concurrent block announcement from another peer when a pushed
    body is rejected
    (#2537).
  • Fixed Windows nodes ignoring the existing state database after a major
    database format upgrade and syncing from genesis instead
    (#2563).
  • The internal miner now switches from a coinbase-only block template to one with
    transactions on the same tip, instead of mining an empty block whenever the first
    template after a new tip had no transactions
    (#2586).
  • The internal miner no longer keeps solving a replaced block template when the
    update arrives during an Equihash attempt that finds no solutions
    (#2586).
  • Stop Zakura block sync from requesting a block it already verified after a
    reorg moves the verified tip past queued downloads. The reset now drops that
    retired work instead of leaving it queued until the next block commits
    (#2597).
  • Block sync keeps an in-flight download when a competing block commits at the
    same height and the requested block is still on the selected header chain.
    While the chains diverge, peers are offered blocks through their shared parent.
    (#2598).
  • The block-sync liveness deadline allows the next body a peer owes to transfer
    at 256 KiB/s. Returned ranges and new requests add no time, so re-issuing work
    cannot move it (#2604).
  • Fixed a delayed block-verification verdict penalizing a peer's replacement
    connection; the verdict now applies to the connection that supplied the block
    (#2604).
  • Fixed a peer's answer from another chain after a reorg closing its whole v2
    connection; the node now resets only the block-sync session and requeues the
    heights (#2604).
  • Fixed the discovery dial history growing without a cap; it now keeps at most
    32,768 entries (#2604).
  • Fixed a replaced v2 connection releasing its per-IP slot while the old
    connection still ran
    (#2604).
  • Fixed nodes with pruned storage downloading old blocks they had already
    finalized each time a peer announced them
    (#2612).
  • Fixed legacy transaction downloads failing with notfound when the peer that
    advertised the transaction was busy: the download now waits up to 2 s for that
    peer instead of asking a peer that doesn't have the transaction yet
    (#2617).
  • With [mempool] implementation = "core", a legacy peer can no longer hold
    more than five undecided transactions by sending children whose parents
    are missing: the pool core now ends an admission at the peer's 73 s
    deadline before its slot is freed. A peer re-advertising ids that are
    already being fetched no longer loses its new ids for lack of a slot
    (#2619).
  • With [mempool] implementation = "core", a transaction waiting on missing
    parents is checked again at once for the first two parents to arrive per
    block, then at most every 10 s, instead of once per arriving parent
    (#2619).

Security

  • Header sync counts a filled page reservation as progress only after the round
    stages a fair share of headers (250), so a peer that advertises one-header
    pages can't hold its round with one header per request timeout. A round that
    starts its validation gets a fresh request deadline
    (#2604).

  • Header sync now stops starting Equihash checks after the first invalid header
    and checks a page's credit before hashing it, and discovery checks a Hello's
    network and chain before its signature
    (#2604).

Also in CHANGELOG.md.

Don't miss a new zakura release

NewReleases is sending notifications on new releases.