Repository navigation
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
sproutskips Sprout commitments, a build withsproutcannot open that database until its tip is past NU7 and the last checkpoint (#2204). - Update peer-cache metric queries.
zcash.net.peers.cachedrops theremote_iplabel and records cached-peer counts in one series per cache file. Update dashboards and alerts that select or group byremote_ip(#2477). - Handle corrected block RPC errors.
getblockheaderandgetblockat 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).
getblocktemplateremoves 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_peersaccepts 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.totalcounts bans (#2444). - Bundled Testnet checkpoints extend to height 4,618,666, so default builds without
sproutcan 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_peersnow 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_backand
peer.inbound.held_seats.refusedcount 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.totaland
address_book.redial.reconnected.totalcount these peers
(#2497). - A DNS name in
network.unconditional_peerscan 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 legacyblockandtxmessages 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 newBlock::zcash_deserialize_bounded
andTransaction::zcash_deserialize_bounded.ZcashReader::with_limitis
removed (#1491). -
Builds without the
sproutfeature no longer build the Sprout note
commitment tree. The first block whose Sprout commitments they skip marks
the database, and a build withsproutrefuses 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). -
getblocktemplatebuilds 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.totalmetric counts bans
(#2444). -
zakura-network: misbehavior reporting moves to a newMisbehaviorReporter.
zakura_network::initandinit_with_zakurareturn it instead of an
mpsc::Sender<(PeerSocketAddr, u32)>, and the handshake builder takes it
withwith_misbehaviorinstead ofwith_bans.MetaAddr::misbehavior,
MetaAddr::new_misbehavior,MetaAddrChange::UpdateMisbehaviorand
MetaAddrChange::misbehavior_scoreare gone
(#2444). -
zcash.net.peers.cacheno longer has aremote_iplabel: 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 byremote_ip
(#2477). -
zakura-rpc: the indexer client'sBlockAndHash::decodeand
BlockAndHeight::decodenow 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
getblocktemplaterequest 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
getheadersrequest with an empty
headersmessage instead of no reply, so a zcashd-compat sidecar peer no
longer times out and churns through reconnects at tip-parity
(#1180). - Fixed
getblockheaderandgetblock(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 thesproutfeature 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.