github zakura-core/zakura v1.8.9

56 minutes ago

Metrics hardening, faster mining templates, and steadier peer reconnections. We recommend upgrading.

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

Before you upgrade

  • Database upgrade prevents rollback. Existing databases upgrade in place to 31.0.0; older releases cannot open them. After a build without sprout skips Sprout commitments, a build with sprout cannot open that database until its tip is past NU7 and the last checkpoint (#2204).
  • Update peer-cache metric queries. zcash.net.peers.cache drops the remote_ip label and records cached-peer counts in one series per cache file. Update dashboards and alerts that select or group by remote_ip (#2477).
  • Handle corrected block RPC errors. getblockheader and getblock at verbosity 1 and 2 now return -1 with the original message for state-service failures. Update clients that treat these failures as missing blocks; missing-block codes remain -5 for hashes and -8 for heights (#2482).
  • Oversized RPC inputs are rejected earlier. Blocks and transactions larger than the 2,000,000-byte block size limit are rejected before parsing, including RPC submissions. Ensure submitted payloads stay within the limit (#1491).

Highlights

  • Security: the metrics endpoint limits connections to 16, closes clients without request headers within 30 seconds or open longer than 120 seconds, and delays failed accepts so idle clients cannot exhaust file descriptors (#1615).
  • getblocktemplate removes a thread-handoff wait for transparent miner rewards, while long polling checks the mempool every 2 seconds instead of every 5 (#2342, #2526).
  • Peers connected inbound for 2 hours keep a seat for 3 minutes after disconnecting; outbound peers connected for 10 minutes get redial priority for 5 minutes, with new counters for both (#2493, #2497).
  • network.unconditional_peers accepts DNS names with a port and resolves them before each dial, following address changes without a restart and retrying the last address if lookup fails (#2478).
  • Legacy peer misbehavior reports are no longer lost: scores accumulate per IP across ports, bans apply immediately at the threshold, and address_book.misbehavior.banned.total counts bans (#2444).
  • Bundled Testnet checkpoints extend to height 4,618,666, so default builds without sprout can sync Testnet from genesis again (#2494).

Library users

Breaking changes in zakura-chain and zakura-network: use the new bounded block and transaction decoders for untrusted input; ZcashReader::with_limit is removed, and MisbehaviorReporter replaces the old reporting APIs (#1491, #2444).

Full changelog

Added

  • network.unconditional_peers now also takes a DNS name with a port, such as
    "peer.example.com:8233". Zakura looks the name up before each dial, and
    never while that peer is connected, so it follows a peer that comes back at a
    new address without a config change or a restart. It dials one address from
    the answer at a time, starting with the first it can dial and moving to the
    next on each later attempt, so each peer needs a name of its own. The address
    it dialed last is treated like a listed address. When a lookup fails, Zakura
    dials the last address again. A name without a port fails config load
    (#2478).
  • A listener now keeps the inbound seat of a peer that was connected for 2
    hours reserved for that peer's IP for 3 minutes after the connection
    closes, so a node that restarts can reconnect to a full listener. At most a
    tenth of the inbound limit is held at once. The new counters
    peer.inbound.held_seats.held, peer.inbound.held_seats.taken_back and
    peer.inbound.held_seats.refused count the holds
    (#2493).
  • A node now dials a peer before other candidates for 5 minutes after an
    outbound connection to it that was open for 10 minutes ends, so a node that
    restarts gets back the peers that had dialed it. Each redial still waits the
    119 s reconnection delay. The new counters
    address_book.redial.started.total and
    address_book.redial.reconnected.total count these peers
    (#2497).
  • A DNS name in network.unconditional_peers can be pinned to the peer's node
    ID, such as "<node id>@peer.example.com:8233". The name's address is then
    only exempt from scoring and bans while a P2P v2 connection proves that node
    ID, so a changed DNS answer, or someone else at the peer's old address, gets
    no exemption. A connection to the name that stays on the legacy protocol
    can't prove the node ID, so Zakura closes it. A pin needs
    network.p2p_stack = "dual", and config load fails
    without it. A node that runs both stacks warns about each unpinned name
    (#2505).

Changed

  • Reduced state-service overhead when processing new blocks by comparing the
    verified best chain by block hash instead of copying every retained header
    reference
    (#1444).

  • Blocks and transactions from peers and RPC requests that are larger than the
    2,000,000-byte block size limit are now rejected before they are parsed,
    including legacy block and tx messages whose bodies exceed the limit
    (#1491).

  • zakura-chain: the block and transaction parsers no longer enforce the block
    size limit. Decode untrusted input with the new Block::zcash_deserialize_bounded
    and Transaction::zcash_deserialize_bounded. ZcashReader::with_limit is
    removed (#1491).

  • Builds without the sprout feature no longer build the Sprout note
    commitment tree. The first block whose Sprout commitments they skip marks
    the database, and a build with sprout refuses to open a marked database
    until its tip is past NU7 and the last checkpoint. The state database
    format moves to 31.0.0; existing databases are reused in place, and older
    releases cannot open them
    (#2204).

  • getblocktemplate builds a template for a miner paid to a transparent
    address without handing the build to another thread, which removes a wait
    that was longer than the build itself when a new block arrives. Templates
    that need a proof for a shielded miner reward are built as before, and
    templates are still built one at a time
    (#2342).

  • Reduced CPU use at the chain tip by no longer re-checking the keys of
    already authenticated peers on every peer-discovery message
    (#2407).

  • Reduced CPU use under transaction load by no longer copying the chain tip
    block's transaction list on every chain tip read
    (#2419).

  • Reduced the CPU a node at the chain tip spends on each new block by no longer
    visiting every block and every index partition when it copies its in-memory
    chains (#2433).

  • Peer misbehavior reports on the legacy P2P stack are no longer batched or
    dropped. Scores add up per IP address across ports, and a peer is banned as
    soon as its score reaches the threshold, instead of up to five seconds
    later. Previously, reports that arrived while the address book was busy
    could be lost, so some misbehaving peers were never banned. The new
    address_book.misbehavior.banned.total metric counts bans
    (#2444).

  • zakura-network: misbehavior reporting moves to a new MisbehaviorReporter.
    zakura_network::init and init_with_zakura return it instead of an
    mpsc::Sender<(PeerSocketAddr, u32)>, and the handshake builder takes it
    with with_misbehavior instead of with_bans. MetaAddr::misbehavior,
    MetaAddr::new_misbehavior, MetaAddrChange::UpdateMisbehavior and
    MetaAddrChange::misbehavior_score are gone
    (#2444).

  • zcash.net.peers.cache no longer has a remote_ip label: each peer cache
    write adds the number of cached peers to one series per cache file. The
    per-address series grew with uptime and were never removed, so long-running
    nodes served a much larger metrics page. Update dashboards and alerts that
    select or group this metric by remote_ip
    (#2477).

  • zakura-rpc: the indexer client's BlockAndHash::decode and
    BlockAndHeight::decode now reject blocks larger than the 2,000,000-byte
    block size limit before parsing them
    (#2489).

  • Reduced CPU use during block sync by no longer re-checking the keys of
    already authenticated peers each time block-sync candidates are published
    (#2490).

  • A long-polling getblocktemplate request now checks the mempool every
    2 seconds instead of every 5, so new transactions reach miners' templates
    sooner (#2526).

Fixed

  • The inbound service now answers an empty getheaders request with an empty
    headers message instead of no reply, so a zcashd-compat sidecar peer no
    longer times out and churns through reconnects at tip-parity
    (#1180).
  • Fixed getblockheader and getblock (verbosity 1 and 2), so that they do
    not report a failure of the state service as a missing block. They now
    return -1 with the original error message. A block that is not in the best
    chain still gives -5 for a hash and -8 for a height
    (#2482).
  • Extended the bundled Testnet checkpoints past NU7 activation to height
    4,618,666, so default builds without the sprout feature can sync Testnet
    from genesis again (#2494).

Security

  • The Prometheus metrics endpoint now serves at most 16 connections at once,
    closes connections that send no request headers within 30 seconds or stay
    open longer than 120 seconds, and waits before retrying failed accepts, so
    idle clients can no longer exhaust the node's file descriptors
    (#1615).

Also in CHANGELOG.md.

Don't miss a new zakura release

NewReleases is sending notifications on new releases.