Zebra 7.2.0-beta.1 is a pre-release from the next branch for testing. It contains everything in 7.1.0 and main, most of the mining and performance work first tested in 7.1.0-alpha.0, and what next has added since: it follows ZIP 2009 by requiring the original ZIP 317 fees on Mainnet until its fee reduction height, expected around 15:20 UTC on 19 November 2026; it turns on compact block relay with compatible peers, and it verifies Orchard, Ironwood and Sapling faster with unreleased cryptography revisions and Pasta assembly. It also adds opt-in block-body pruning and opt-in fast tree sync, and it refuses to start with custom Testnet or Regtest parameters that 7.1.0 ignored or panicked on. Since 7.2.0-beta.0, pruned nodes advertise NODE_NETWORK_LIMITED instead of NODE_NETWORK with NODE_PRUNED, failed transaction downloads are retried from other announcing peers, and submitblock is faster. It adds no network upgrade. It upgrades the state format from 29.0.0 to 29.2.0, which takes about a minute on Mainnet, and a node that doesn't enable pruning or fast tree sync can switch back to 7.1.0 without resyncing. Its end of support is the same as 7.1.0's: it warns from around 3 December 2026 and stops around 10 December 2026.
This is a beta for testing. Run it on a node you can switch back to 7.1.0, and keep production nodes on 7.1.0. Its crates aren't published to crates.io, it never becomes the Latest release, and its Docker image is tagged 7.2.0-beta.1 and edge, not latest. If you run a pruned 7.2.0-beta.0 node, upgrade it to this beta: see "Changes since 7.2.0-beta.0". If you run a wallet, mining, or RPC service on it on Mainnet before height 3,590,000, read the fee change under Breaking Changes first. If you run a configured Testnet or Regtest, check its [network.testnet_parameters] against the configuration changes under Breaking Changes before you upgrade.
Documentation for this beta is in the Zebra book for the next branch, at zebra.zfnd.org/next, which has been rewritten for this release. The /next book follows the branch, so it may also describe changes made after this beta.
Update Priority
| User Type | Priority | Reason |
|---|---|---|
Testers and operators running edge
| Medium, on a test node | State upgrade to 29.2.0, ZIP 2009 fee policy, compact block relay on by default, faster verification |
| Pruned 7.2.0-beta.0 nodes | High, on a test node | They advertise NODE_NETWORK with NODE_PRUNED, so Zebra 7.1.0 and zcashd peers take them for archive nodes; this beta advertises NODE_NETWORK_LIMITED instead
|
| Mining pools and solo miners | Medium, on a test node | Faster submitblock, templates served before their proposal check, compact block relay, pruned mining nodes, preciousblock, ZMQ, mining metrics
|
| Custom Testnet and Regtest operators | Medium, before upgrading | Invalid funding streams, lockbox disbursements and recipients, and Regtest values Zebra used to ignore, now stop zebrad from starting
|
| Wallet, RPC, and lightwalletd operators | Low, on a test node | Faster getaddressutxos, getblock and getrawtransaction; on Mainnet, reduced-fee transactions aren't relayed until height 3,590,000, around 15:20 UTC on 19 November 2026
|
| Production node operators | None | Stay on 7.1.0 |
Changes since 7.2.0-beta.0
For testers already running 7.2.0-beta.0. Both betas use state format 29.2.0, so upgrading from 7.2.0-beta.0 runs no state upgrade, and a 7.2.0-beta.0 config loads unchanged.
- Pruned nodes advertise
NODE_NETWORK_LIMITED. In 7.2.0-beta.0, a pruned node advertisedNODE_NETWORKwith the experimentalNODE_PRUNEDbit (bit 25). Zebra 7.1.0 and zcashd don't know bit 25, so they took it for an archive node and could ask it for historic blocks it had deleted. This beta replaces that: a pruned node never advertisesNODE_NETWORKorNODE_PRUNED, and advertisesNODE_NETWORK_LIMITED(bit 10) only once it holds the full 2,304-block window ending at its tip. It answersnotfoundfor blocks more than 2,304 below its tip. When Zebra estimates that it is within 1,152 blocks of the network tip, it asks limited peers for the block hashes that extend its own tip, and downloads from a limited peer only blocks that peer announced within 2,304 blocks of its tip. Zebra still treats peers that advertiseNODE_NETWORKwithNODE_PRUNED, such as pruned Zakura nodes and pruned 7.2.0-beta.0 nodes, as non-archive peers: it doesn't connect to them as new outbound peers while syncing, or ask them for historic blocks (#12537). - Transaction download retries. When a peer answers
notfoundor fails for a gossiped transaction, Zebra retries the download up to twice from other peers that announced it, after 200 milliseconds, instead of leaving it to the mempool crawler, which could take up to 73 seconds. Nodes rebuild compact blocks and include new transactions in templates sooner (#12532). - Optional CPU pinning. The new
sync.pin_parallel_cpu_threadskey, off by default, pins the verification workers to distinct physical cores on Linux. Zebra 7.1.0 and 7.2.0-beta.0 reject it, and a config generated by this beta contains it, so remove it before you switch back (#12533). - Faster
submitblockand templates. The state validates a submitted block's chain context while the block verifier is still checking it, so the commit only publishes the result. Blocks without note commitments commit on the block writer's own thread, and the block verifier checks proof of work, and hashes blocks of up to 64 KiB and 32 transactions, on its own task instead of a blocking thread. In thestate.block_writer.phase.duration_secondsmetrics, thecontextual_commitphase of a submitted block includes that earlier validation, so it can be longer than the block verifier'sstate_commitphase inzebra.consensus.block.phase.duration_seconds(#12542). Verified blocks share their created outputs with cached mined commits instead of copying them (#12540). Template precomputation no longer starts a child coinbase proof that is already cached, which delayed the next template for miners with a shielded address (#12539).
Highlights
For miners and pools
- Most of 7.1.0-alpha.0. No timer waits between new transactions or blocks and new work, faster
submitblockfor validated proposals, earlier relay of solved blocks, legacy JSON-RPC batches for pool software, ZMQhashblocknotifications,preciousblock, and mining metrics. See the 7.1.0-alpha.0 notes. The alpha's 5-second peer misbehavior scoring isn't in this beta: scores below the ban threshold apply every 30 seconds again, as in 7.1.0. - Templates are served before their proposal check. A template is served as soon as it's built and checked as a block proposal in the background. If the check fails, Zebra restores valid work for that parent or withdraws the work. Early coinbase-only work on a newly received block now expires after 1 second instead of 30 (#12507).
- Compact block relay is on by default. Peers that both negotiate it, through a service bit from the draft Compact Block Relay ZIP, send compact blocks rebuilt from the receiver's mempool, and can relay authenticated blocks before full verification. Other peers keep full-block relay. Set
network.compact_block_relay = falseto turn it off (#11932, #12507). - Pruned mining nodes.
state.prune_blocks_older_thankeeps only recent block bodies. It's opt-in, and an existing archive must be moved by hand while the node is stopped; see the pruning guide. A pruned node never advertisesNODE_NETWORKor the experimentalNODE_PRUNEDbit. It advertisesNODE_NETWORK_LIMITED(bit 10) only once it holds the full 2,304-block window ending at its tip, and answersnotfoundfor blocks more than 2,304 below its tip. Zebra 7.1.0 doesn't know bit 10: it ignores and doesn't relay gossiped addresses withoutNODE_NETWORK, refuses them as outbound peers while syncing, and otherwise asks them only for blocks they announced or that noNODE_NETWORKpeer can send. zcashd treats them as clients: it doesn't sync headers or download blocks from them, so blocks a pruned miner finds reach zcashd nodes only through other peers. Both still accept their connections. Keep some Zebra peers connected, for example withnetwork.persistent_peers, so your blocks reach the network quickly. Zebra no longer stalls its initial sync on peers that advertiseNODE_NETWORKwithNODE_PRUNED, such as pruned Zakura nodes (#12087, #12088, #12090, #12507, #12521, #12537). - ZMQ needs a source build. The release binaries and Docker images are built without the
zmqfeature, and they refuse to start whennotify.zmqpubhashblockis set. Build from source with--features zmqto use it.
For wallet, RPC, and lightwalletd services
- ZIP 2009 fee policy. Until next-block height 3,590,000 on Mainnet, expected around 15:20 UTC on 19 November 2026, this beta requires the original ZIP 317 fees: transactions paying the reduced 1,000-zatoshi fee per logical action, which 7.1.0 accepts, aren't relayed or mined until then. From that height it accepts the same fees as 7.1.0. Public Testnet is already past its height, 4,465,026, so there it accepts the same fees as 7.1.0 (#12365).
- Faster address and block RPCs.
getaddressutxosis about 50 times faster for addresses with a long spent history, after a one-time state upgrade (#12390).getblockandgetrawtransactionwith verbosity 0 return stored bytes, JSON-RPC 2.0 responses pass through without re-encoding, and verbose responses render scriptasmfaster (#12384, #12386, #12387, #12494).
For all node operators
- Faster shielded verification. Release binaries are built with unreleased Pasta, Halo2, RedDSA, orchard, Sinsemilla and sapling-crypto revisions, pinned by commit, and with Pasta assembly, which detects BMI2/ADX at runtime and falls back on older x86-64 CPUs. An unreleased
zcash_scriptrevision is also pinned for faster script rendering. (#12507, #12516, #12520, #12527). - Opt-in fast tree sync. With
state.fast_tree_sync = true, checkpoint sync commits blocks with note commitment tree roots supplied by peers at 3 distinct IP addresses. A fast-synced database can't be used by 7.1.0. Every node now advertisesNODE_TREE_ROOTS(service bit 3) and answersgettreeroots, even with the option off; these messages are drafts pending a ZIP (#12507). - Faster commits at the chain tip. Chain copies now share their per-block indexes, so each commit and finalization no longer copies them (#12508).
- Selected peers and public addresses.
network.persistent_peerskeeps chosen peers connected in reserved outbound slots. Without a configured external or public listen address, Zebra advertises the public IP that its outbound peers agree on (#12459). - Clean shutdown. SIGINT and SIGTERM now close listeners and connections and stop a managed zcashd before exiting (#12367).
What changes for operators
- Configuration on Mainnet and public Testnet. No configuration key is removed or renamed since 7.1.0, so a 7.1.0 config still loads. If you ran 7.1.0-alpha.0, remove
network.experimental_early_block_relay, which this beta rejects. Compact block relay is then on; setnetwork.compact_block_relay = falseto keep it off. - Configuration on a configured Testnet or Regtest. These now stop
zebradfrom starting instead of being ignored or causing a panic later: lockbox recipients that aren't P2SH or P2PKH addresses, such as TEX addresses; funding streams or lockbox disbursements that are invalid, including Regtest funding streams with too few addresses unlessextend_funding_stream_addresses_as_required = true; and, on Regtest,network_magic,slow_start_interval,target_difficulty_limit,disable_pow,genesis_hash,pre_blossom_halving_intervalortemporary_orchard_disabling_soft_fork_heightset to anything other than Regtest's value (#12048, #12049). - Metrics.
rpc.request.duration_secondsandzcash.net.peer.handshake.duration_secondsare now histograms with_bucketseries instead ofquantileseries.zcash.net.peers.missing_services,zcash.net.peers.initial, andzcash.net.peers.cachehave different labels. Instate.requests, raw block and transaction reads are now labelledraw_blockandany_chain_raw_transactioninstead ofblockandany_chain_transaction, and peer request metrics label served block downloadsRawBlocksByHash,EarlyRawBlocksByHashandRawBlocksinstead ofBlocksByHash. Update dashboards and alerts that select on them. Zebra also exports standard process metrics and a# HELPline for every metric, and sends every metric to Sentry whenSENTRY_DSNis set (#12437, #12439). - Example Docker Compose files. They no longer publish the RPC and metrics ports to the host, the observability stack turns on RPC cookie authentication and needs
GF_SECURITY_ADMIN_PASSWORDset to start, Prometheus no longer enables its admin and lifecycle APIs, and the example errors dashboard is removed. The mining stack readsEXTRA_COINBASE_DATA, literal UTF-8 of at most 86 bytes, instead ofMINER_DATA, so rename it in your.env. If you run a copy of these files, publish the ports you need again yourself (#12039, #12040, #12041).
Breaking Changes
- Reject Regtest configurations whose
[network.testnet_parameters]setnetwork_magic,slow_start_interval,target_difficulty_limit,disable_pow,genesis_hash,pre_blossom_halving_interval, ortemporary_orchard_disabling_soft_fork_heightto a value other than the one Regtest uses, so the node doesn't start; Zebra previously ignored these values. Remove the field, or use a configured Testnet to change it.network_nameis still ignored (#12049). - Fail configuration loading with an error on invalid
[network.testnet_parameters]funding streams and lockbox disbursements, instead of panicking at startup or in subsidy RPCs. A Regtest node whose funding streams have too few addresses, which previously started and panicked only when a subsidy was computed, no longer starts; add the missing addresses or setextend_funding_stream_addresses_as_required = true(#12048). - Follow ZIP 2009 by requiring the original ZIP 317 fees until next-block height 3,590,000 on Mainnet or 4,465,026 on public Testnet. Transactions paying the reduced 1,000-zatoshi fee per logical action, which Zebra 7.1.0 accepts at every height, are not relayed or mined until those heights. Reduced fees and higher mining weights apply from height 1 on Regtest and custom testnets.
- Reject configured Testnet and Regtest lockbox recipients without P2SH or P2PKH payment scripts; replace TEX recipients with supported transparent addresses.
Added
- Added the
queue,contextual_commit,update_latest_chain_channelsandfinalizationphases of blocks committed after checkpoint sync to thestate.block_writer.phase.duration_secondshistograms. - Added compact block relay, enabled by default with peers that also enable it, using service bit 16 (
NODE_COMPACT_BLOCKS) from the draft Compact Block Relay ZIP. After NU5, a block extending the best tip can be relayed to those peers once its proof of work and transaction commitments have been checked, before full validation. Receivers rebuild compact blocks from their mempool, request missing transactions once, fall back to the full block, and still validate every block fully. Other peers keep ordinary relay after commit. Setnetwork.compact_block_relay = falseto disable it (#11932). - Added current network difficulty to
getmininginfoand advertised peer starting heights togetpeerinfofor pool compatibility. - Added optional ZMQ
hashblocknotifications for mining pools, configured withnotify.zmqpubhashblock. They need azebradbuilt with thezmqCargo feature, which release binaries and Docker images don't include; other builds don't start when the option is set. - Added opt-in historical block-body pruning with
state.prune_blocks_older_than; archives remain the default instate/v29and retain rollback compatibility with Zebra 7.1.0. Zebra 7.1.0 doesn't start with the newnetwork.compact_block_relay,network.persistent_peers,state.prune_blocks_older_than,state.fast_tree_sync,sync.pin_parallel_cpu_threadsornotify.zmqpubhashblockkeys, so remove them from the configuration before rolling back. Existing archives require an explicit stopped-node move into an absentstate-pruned/v29directory before enabling pruning. Stop every writer and direct reader and keep a separate archive backup first; startup never performs this activation move. New pruning caches start isolated directly, using the same format version and upgrade sequence. Inaccessible or coexisting archive/pruned paths fail without fallback or overwrite. Disabling pruning keeps the isolated directory if all bodies remain; after deletion, restore an archive snapshot or fully resync in a separate cache. Retention counts finalized bodies, keeps genesis, and requires at least 10,000 bodies on Mainnet/Testnet or 1,001 on Regtest. Increasing retention affects future deletion only. Catch-up pruning compacts deleted bodies after clearing the backlog; normal compaction handles ongoing deletion. Pruned nodes never advertiseNODE_NETWORKorNODE_PRUNED, and advertiseNODE_NETWORK_LIMITED(bit 10) only once they hold the full 2,304-block window ending at their tip; Regtest retention below 2,304 advertises neither. Zebra 7.1.0 and zcashd don't recognize bit 10, so they treat pruned nodes as unable to serve blocks. Pruning rejects zcashd sidecars. Limited nodes answernotfoundfor blocks more than 2,304 below their tip, and pruned nodes for unavailable bodies; block inventory stops at missing history while headers remain available. Pruning retains a restart witness and waits without deleting bodies when a bounded search finds none; valid branchless suffixes do not stop the running node. Monitor the retained floor, deleted heights, and batch latency throughstate.pruning.retained.height,state.pruning.deleted.height.count, andstate.pruning.batch.duration_seconds. - Added Pasta assembly to default release builds on AArch64 and supported x86_64 targets, with runtime BMI2/ADX detection, portable fallback on older x86_64 CPUs, and a startup log naming the selected backend.
- Added
preciousblockto prefer a known non-finalized tip among equal-work chains, with the last call winning and preferences lost on restart. - Added the
transparent_spend,anchor_reads,pickup,chain_clone,trees,chain_updatesandhistorysub-phases of each committed block to thestate.block_writer.phase.duration_secondshistograms. - Added experimental opt-in
state.fast_tree_sync, which speeds up checkpoint sync by taking authenticated note commitment tree roots from peers instead of computing the Sapling, Orchard and Ironwood trees, up to the trees embedded in Zebra. Fast sync starts only after peers at 3 distinct IP addresses have supplied roots; until then, blocks commit with tree updates. Every node advertisesNODE_TREE_ROOTS(service bit 3) and answersgettreerootsfrom its finalized chain, even withstate.fast_tree_syncoff. Fast-syncing a database is a one-way upgrade: older Zebra versions and direct readers cannot use it; retain a pre-fast-sync backup for rollback, or delete the state cache and resync with the option disabled. Missing roots stop Zebra after 120 seconds with recovery instructions, including after restarting with the option disabled. At fast-synced heights,getblock,getblockheaderand lightwalletd compact blocks report the stored tree sizes and roots,z_gettreestatereturns error -20 "tree state unavailable: fast-synced height", and the lightwalletdGetTreeStateandGetLatestTreeStatemethods returnUNAVAILABLE. New legacy messages and service bit 3 remain draft pending a ZIP. - Added the standard Prometheus process metrics, such as
process_resident_memory_bytes,process_open_fds, andprocess_cpu_seconds_total, which Zebra samples every 15 seconds when the metrics endpoint is open or Sentry is active, except on Windows. - Added a description and unit for every Zebra metric, which the Prometheus endpoint shows as
# HELPlines, and the metric catalog indocker/observability/metrics-catalog.json, which lists each metric's kind, unit, description, labels, and Prometheus series names. - Maintain selected peers with
network.persistent_peers, reserving outbound slots and retrying failures with fresh DNS lookups. - Added the
commitlabel tozebrad_build_info, and metrics that time each phase of block verification, checkpoint commits, block decoding, gossiped block downloads, block template building, the first template after a new tip, mined block announcements, and mempool cleanup, withsubmitblockoutcome counts. They are exported as Prometheus histograms, so they can be aggregated across nodes. - Added
sync.pin_parallel_cpu_threads(defaultfalse), which pins Rayon verification workers to distinct physical CPU cores on Linux, one worker per core available to the process. A nonzerosync.parallel_cpu_threadscaps the worker count, and the core count takes precedence overRAYON_NUM_THREADS. Workers stay unpinned on other systems or if the core topology can't be detected (#12533).
Changed
Mining and block templates
- Avoid repeated transaction and contextual validation when submitted blocks exactly match validated mining proposals apart from nonce and solution. Retain at most eight proposals within a conservative 576 MiB content charge, including spent-output scripts, rather than measured heap usage. Eight representative full-size proposals of 104-byte one-input/one-P2PKH-output payments fit; smaller valid transactions or large spent-output scripts can cause earlier eviction and normal revalidation. (#11658)
- The submitblock RPC now accepts any JSON value for its ignored second argument, matching zcashd and allowing mining pools to submit blocks with a string argument.
- Mining work can extend newly received blocks before transaction and contextual validation finish. Early work is coinbase-only, expires after 1 second, and is disabled once NSM reissuance activates. A live candidate is not replaced by concurrent submissions, and expiry does not reopen its slot before verification finishes. Unpublished candidates release their slot after context computation finishes, without waiting for block commitment. A candidate is withdrawn when its parent is no longer the committed best tip. Long polls recheck cached child work immediately when its parent commits, and new candidates interrupt template retry backoff.
- Mining templates refresh after verified mempool additions and evictions, including removed dependent transactions, and wake long polls on publication. Steady arrivals use a 50 ms coalescing pause; bursts of at least 50 additions or lost notifications use one second. Slow builds extend the pause, while chain-tip changes rebuild immediately. Cached failures from an old tip do not interrupt mining (#11374).
- Mining templates remain available after backwards clock adjustments when their timestamp range can be narrowed safely (#11654).
- The internal miner discards solutions whose parent is no longer committed, pauses public-network mining while behind or without recently live peers, and waits for a different parent after accepting its own public-network block. Regtest and configured Testnet retain peerless mining and may extend their own accepted blocks (#11659).
- Relay submitted blocks before commit only when they match an exact successful local proposal verdict, except for nonce and Equihash solution. Early advertisements use ordinary gossip; all-peer advertisements remain commit-only. State-owned committed notifications survive RPC cancellation, and unavailable notification capacity rejects mined commits before mutation without blocking ordinary commits.
- The internal miner starts on new work immediately instead of sleeping for up to 20 seconds, and keeps its current candidate when a mempool-only template refresh allows old work.
- Block template transaction selection calculates coinbase resource usage without constructing a separate zero-fee coinbase, avoiding an additional shielded proof when a sizing coinbase would otherwise be needed.
- Reuse parsed transactions when submitted block bytes exactly match a validated mining proposal except for nonce and solution, and pass that match into verification without repeating body identity hashing. Misses retain full decoding and validation; proof-of-work, time and live-state checks still run on matches (#11739).
- Reduced block-template transaction selection time for large mempools while preserving ZIP 317 fee-weighted selection.
- Reduced mining memory use by sharing mempool transactions with the validated proposal cache.
- Precompute fee-bearing shielded coinbases while a mining candidate is being verified, reusing the proof only when the committed template has matching fees and subsidy context. Limit speculation to one fee proof per committed parent and skip mempool-reset boundaries. Keep the committed tip's and candidate child's coinbases across candidate expiry or withdrawal, and skip duplicate shielded proofs that would delay later mining templates.
- Correct founders reward payouts and pre-NU5 header commitments in external mining templates.
- Decode large submitted blocks in parallel under shared admission for
submitblockand proposal requests. Cancelled requests retain admission until their blocking decode finishes; inputs up to 4 KiB remain serial and inline (#12308). - Mining templates are served as soon as they are built and checked as block proposals in the background. Solved templates reuse successful checks; rejected work, including narrowed timestamps and RPC fallback builds, restores valid work for that parent or is withdrawn when no valid fallback exists. Rejection also revokes old-work authorization in
submitold; ordinary fee-only refreshes still permit old work. Restoring valid work preserves normal mempool-change and periodic refresh pacing. RPC proposals must extend the current tip at the next height on every network. Their 30-second response deadline includes readiness, and the two-request limit continues to count verification after timeout or client cancellation. - Mining to a Unified address skips Orchard receivers while Orchard payouts are disabled, using an available Sapling or transparent receiver instead. Addresses without an enabled receiver return a construction error (#11696).
- Reduced
submitblocklatency by validating a submitted block's chain context while the block verifier checks it, instead of afterwards (#12542).
Mempool
- Store and announce verified transactions without waiting for the five-second mempool queue check on idle nodes (#11376).
- Mempool transaction verification and downloads read state directly instead of waiting in the state write queue, and mempool metrics are updated incrementally.
- Reduced allocation and copying during mempool cleanup after a block is committed.
- Count same-effects transaction variants once per advertising IP in the inbound mempool queue without suppressing other peers' or local candidates (#12052).
- Reduce mempool expiration scans and redundant polls during transaction submission bursts.
- Reduce state round trips when admitting transactions to the mempool, which speeds up transparent transaction admission.
- Reduce mempool cleanup work when removing transactions and their dependents, without changing rejection tracking (#12341).
- Report submitted transactions removed as dependents during mempool eviction as rejected, without persistently rejecting collateral children that can be resubmitted after their parent is mined (#12370).
- Retry failed gossiped transaction downloads from other announcing peers before waiting for the mempool crawler, improving mining and compact-block reconstruction latency.
Block verification and sync
- Waiting for a missing transparent input during block verification no longer delays lookups of the transaction's other inputs (#11186).
- Lowered block and transaction verification latency by flushing proof and signature batches after 5 ms of idle input, while retaining the 100 ms maximum batch latency.
- Note commitment trees, including the Orchard and Ironwood trees, are updated faster for blocks with many shielded outputs, which speeds up block commits and checkpoint sync.
- Checkpoint sync commits blocks faster: authorizing data roots are computed before the database writer, and large blocks are read in parallel. RocksDB starts flushing column families backed by the oldest live WAL files when total WAL usage exceeds 4 GiB.
- Blocks with many transactions compute their authorizing data root in parallel, which shortens block commits.
- Checkpoint commit invariant panics stop the node even when the requesting caller has cancelled.
- Start partial shielded verification batches at the end of block transaction queueing rather than waiting for the idle timer.
- Start shielded transaction verification before transparent transaction verification to reduce mixed-block verification latency (#11805).
- Reduce signature verification overhead for empty and single-signature batches.
- Reduce proof-of-work verification cost with AVX2 and NEON leaf hashing on supported CPUs.
- Verify Sapling bundles in parallel on blocking workers, keeping proof preparation off asynchronous workers and flushing every used batch when block checks are queued.
- Build Halo2 verifying keys in the background at startup, avoiding first-use Orchard and Ironwood key construction after preload finishes without waiting for it before RPC or network readiness (#12008).
- Reduced the time to commit blocks at the chain tip, by starting each block's note commitment tree updates before its other contextual checks, by sharing the non-finalized chain's per-block indexes between chain copies instead of copying them on every commit and finalization, and by freeing the replaced non-finalized state snapshot after the commit is acknowledged.
- Prepare transaction verification on blocking threads, so blocks with many shielded transactions verify faster without stalling async workers.
- Reuse inbound blocks that commit while sync briefly waits, with bounded independent downloading if inbound verification stalls or fails (#12309).
- Block transaction verification no longer passes through an intermediate dispatch worker.
- Speed up checkpoint sync of shielded-heavy blocks by storing transactions without re-serializing them (#12419).
- Speed up syncing shielded-heavy blocks by updating note commitment trees on a thread pool the block writer doesn't share (#12420).
- Zebra caches successful transparent script verifications, so an input script verified when its transaction entered the mempool is not verified again when a block mines that transaction with the same spent outputs and network upgrade. Every other block and transaction check still runs. The cache is reported by the
zebra.consensus.cache.*metrics with thetransparent_scriptverifier label (#11939). - Reduce allocation and unused diagnostic bookkeeping overhead during checkpoint synchronization.
- Reduce transaction decoding CPU overhead with bulk byte-vector codecs.
Peer network
- Newly committed blocks are advertised to peers immediately instead of after up to 7 seconds, and mempool transactions are advertised every 2 seconds instead of every 7.
- Download newly advertised blocks promptly when a block-discovery request is pending, even if inbound gossip drops the advertisement. Gossip skips blocks the syncer is already downloading, but can retry a rejected block while its peer-attribution lookup is pending (#11710).
- Deliver queued block adverts even after gossip callers time out.
- Alternate listener addresses learned during handshakes are kept with their advertised capabilities instead of dropping non-serving listeners. The
getnetworkinforesponse'slocalservicesreports the services this node selected at startup instead of alwaysNODE_NETWORK. Handshakes can also advertiseNODE_COMPACT_BLOCKS, andNODE_NETWORK_LIMITEDonce a restored pruned node completes its recent 2,304-block window after startup;localservicesshows neither. Pruned listeners remain discoverable through gossip, but candidates that may serve archive history are dialed first and non-serving peers fill at mostpeerset_initial_target_sizeordinary outbound connections (persistent peers excluded), keeping the other slots for serving peers. - Recover one missing parent of an out-of-order gossiped block without waiting for the next sync round, reusing parents already being downloaded.
- Avoid delaying newly gossiped blocks until the next sync discovery deadline when their source is briefly busy.
- Serve ordinary block downloads from stored bytes on legacy and negotiated connections, avoiding transaction decoding on the serving node. Cached early bodies retain revocable serving authorization. The request/response metric command labels now include
RawBlocksByHash,EarlyRawBlocksByHashandRawBlocks; finalized state reads useraw_block. - Discover and advertise public listening IPs from agreeing outbound peers when no external or public listening address is configured. Configure
network.external_addrfor a different public address or port.
RPC
- RPC batches using a single version now accept versionless and JSON-RPC 1.0 calls and return responses in the same format, allowing mining pools such as s-nomp to run their startup checks. Converted requests and responses respect the existing body size limits, so a legacy response that only exceeds the response limit after conversion now returns a legacy-formatted oversized-response error instead of the full result, retaining the upstream error-envelope exception.
getblockchaininforeportssize_on_diskfrom an estimate refreshed about every 30 seconds instead of measuring the database on every call.- Avoid panics in
getblocksubsidyand coinbase validation on custom Testnets whose block subsidy is not divisible by five (#12050). - Changed empty-state
getinfo,getblockchaininfoandgetmininginfodifficulty to1.0, matching zcashd. - Changed
gettxoutto answer from the UTXO set without requiring the creating transaction body; once that body is pruned,versionis null. - Report unavailable historical bodies explicitly in JSON-RPC and gRPC, preserve retained-index queries, and expose actual pruning availability through
getblockchaininfo. Keep historical tree-state headers on the requested fork when cached state is stale. Block-range and address-transaction streams deliver buffered results before terminal errors; stalled consumers receiveDEADLINE_EXCEEDEDrather than a successful end-of-stream. getblockandgetrawtransactionwith verbosity 0 return the stored bytes of finalized blocks and transactions instead of deserializing and re-serializing them; responses are unchanged. In the state request metrics these requests are now labelledraw_blockandany_chain_raw_transactioninstead ofblockandany_chain_transaction.- Serialize each transaction once when building verbose transaction objects for
getblockverbosity 2 and 3 and verbosegetrawtransaction, without changing their output. - The RPC compatibility middleware now forwards plain JSON-RPC 2.0 requests and their responses unchanged instead of parsing and re-encoding them, which removes up to tens of milliseconds per large
getblockresponse. Responses to JSON-RPC 1.0 requests fromlightwalletdandzcash-cliare converted by rewriting only their first and last bytes, so large results are no longer re-parsed either. - Verbose
getblockandgetrawtransactionresponses render scriptasmfaster, with unchanged output. - Convert large hex payloads in
submitblock,getblocktemplateand raw block and transaction RPCs several times faster (#11983). - Made
getaddressutxosabout 50 times faster for addresses whose outputs were all spent over a long history, such as the funding stream addresses. The first start runs a short state upgrade, about a minute on Mainnet (#12390). - Changed
z_gettreestateto return "the requested block is not in the main chain" instead of empty trees when the block is invalidated, or reorged away and pruned, while the request is answered.
State storage
- State copying checks source history and the target chain, body availability, and format before opening a writable target. Unavailable ranges, pruned targets, different-chain targets, and incompatible older-major targets are rejected without creating or changing the target. A target exactly one block behind now receives the remaining block rather than skipping it. Target pruning is disabled during replay and comparison.
- State-service startup reports initialization failures and their recovery instructions instead of a generic writable-directory, lock, or free-space hint (#12092).
Operations and monitoring
- Accept
ZEBRA_TARGET_environment overrides forcopy-statewithout failing to load the source configuration (#12037). - Apply top-level
--filterswhenstartis given explicitly, unless positionalstartfilters are provided (#12038). - SIGINT and SIGTERM now run node cleanup, close health listeners and connections, and stop a managed zcashd while its output is still logged. Signals during startup still cancel initialization promptly.
- Reduce retained heap memory on Linux glibc builds with an eight-arena default, preserving
MALLOC_ARENA_MAXandGLIBC_TUNABLES=glibc.malloc.arena_max=...overrides (#12320). - Reduce startup resident memory by loading only pinned Sapling verifying keys; bundled output proving parameters are loaded only for Sapling coinbase payouts.
- Reduced the work Zebra does when Sentry is enabled: it no longer builds a Sentry span for every INFO-level span, such as one per inbound peer message. Zebra doesn't send Sentry traces, so these spans were never sent.
- Stopped starting the Sentry session rotation and metrics flush threads when
SENTRY_DSNis unset or the Sentry client is disabled. - Changed
rpc.request.duration_secondsandzcash.net.peer.handshake.duration_secondsto Prometheus histograms: the endpoint now exports_bucketseries for them instead ofquantileseries, so they can be aggregated across nodes. The supplied dashboards and alert rules already query the_bucketseries. - Changed the
zcash.net.peers.missing_serviceslabel fromremote_services, the service bits the rejected peer sent, tomissing_service, the required service it lacks, such asNODE_NETWORK. Removed thecachelabel, the peer cache file path, fromzcash.net.peers.initialandzcash.net.peers.cache. Theseedlabel ofzcash.net.peers.initialnow names only Zebra's default DNS seeders, and labels every other configured initial peercustom. Update dashboards and alerts that select on these labels. - Changed the metrics sent to Sentry when
SENTRY_DSNis set: Zebra now sends every metric that the Prometheus endpoint exports, instead of a fixed list, once a minute. Counters send their increase, gauges their latest value when it changes and at least every 5 minutes, and histograms up to 20 values per set of labels as distributions. Each minute sends at most 1,500 metrics and at most 256 sets of labels per metric, and sends none while the Sentry SDK is still sending earlier data, so metrics never fill the queue they share with events, logs, and sessions. The newzebrad.sentry.metrics.dropped.totalcounter counts the values left out. Metrics also carrychain.near_tip, which istruewhen the node is near the chain tip. - Added the
chain_network_upgradelabel tozcash.chain.tip.block_delay_seconds, the network upgrade at the block's height. Sentry still receives each value's block height as thechain.heightattribute.
Docker and examples
- Fixed the example Grafana dashboards and Prometheus alert rules that queried series Zebra doesn't export. The
ZebraHighErrorRatealert fires on block verification errors fromsync.block.error.count. TheRPCHighLatencyalert ignoresgetblocktemplatelong polls, as do thep99 Latency by Methodpanels on the RPC and overview dashboards, which now show each method separately. TheSlowHandshakesalert fires when the p95 successful handshake takes over 2 seconds, below the 3 second handshake timeout. The RPC tracing dashboard queries the span metrics that the supplied Jaeger configuration exports, and its error panels show zero for methods with requests but no errors. Zebra exports the signature and proof failure counters, which count the items of batches that failed verification, at zero from startup, so the first failures show inrate()(#12041, #12438). - Made the example mining Compose stack start: it sets
ZEBRA_MINING__EXTRA_COINBASE_DATAinstead of the unknownZEBRA_MINING__MINER_DATA, and enables the loopback health listener its health check probes (#12040).
Dependencies
- Updated the
libz-sysbuild dependency to 1.1.33, clearing RUSTSEC-2026-0544. Zebra isn't affected: zebrad only uses it at build time, throughvergen-git2, and never callsgzprintf. - Build with unreleased Pasta, Halo2, RedDSA, orchard, Sinsemilla and sapling-crypto revisions, pinned by commit until the crates publish them. They speed up Orchard, Ironwood and Sapling proof, signature and hash verification, especially in blocks with many Orchard or Ironwood transactions, and parse large Sapling proof lists in parallel.
Removed
- Removed the example Grafana errors dashboard, which queried sighash error series Zebra doesn't export (#12041).
Fixed
- Fixed a panic when the mempool rejects a transaction again while an earlier rejection is still remembered.
- Fixed initial sync stalls on pruned Zakura peers (Zakura 1.8.8 and later) that advertise
NODE_NETWORKwithNODE_PRUNED: Zebra doesn't use them as new outbound peers while syncing, or as a source of historic blocks (#12521).
Security
- Stopped the example Compose files from publishing the Zebra RPC and metrics ports to the host. The observability stack enables RPC cookie authentication, disables the Prometheus admin and lifecycle APIs, and requires
GF_SECURITY_ADMIN_PASSWORDto start (#12039).
How to Upgrade
On first start, this beta upgrades a 7.1.0, 7.1.0-alpha.0 or 7.0.0 state from format 29.0.0 to 29.2.0 in place, which takes about a minute on Mainnet; it needs no resync. A 7.2.0-beta.0 state is already 29.2.0, so it opens without an upgrade. If you are upgrading from Zebra 6.x, also read the 7.0.0 release notes, because 7.0.0 migrates the state from v28 to v29.
No configuration key is removed or renamed since 7.1.0 or 7.2.0-beta.0, so a Mainnet or public Testnet config from either loads unchanged. Before upgrading a configured Testnet or Regtest, fix the [network.testnet_parameters] that this beta now rejects, listed under "What changes for operators" and Breaking Changes. Remove network.experimental_early_block_relay from a 7.1.0-alpha.0 config: this beta rejects it. Compact block relay is then on; set network.compact_block_relay = false to keep it off.
Rollback: 7.1.0 can open the state of a node that never enabled state.prune_blocks_older_than or state.fast_tree_sync, so you can switch back without resyncing. Remove the new network.compact_block_relay, network.persistent_peers, state.prune_blocks_older_than, state.fast_tree_sync, sync.pin_parallel_cpu_threads and notify.zmqpubhashblock fields from your config first, including any that a config generated by this beta contains, because 7.1.0 rejects unknown fields. To switch back to 7.2.0-beta.0 instead, remove only sync.pin_parallel_cpu_threads. Don't switch a pruned node back to 7.2.0-beta.0: it advertises NODE_NETWORK with NODE_PRUNED again (see Changes since 7.2.0-beta.0). A database that has fast-synced or deleted block bodies can't be used by 7.1.0: keep a backup from before you enable either option, or resync. The pruning guide has the steps for a pruning node.
Docker
docker pull zfnd/zebra:7.2.0-beta.1zfnd/zebra:edge points to the same image. zfnd/zebra:latest stays on 7.1.0.
Binaries
- Download the
zebradarchive for x86-64 or ARM64 Linux from the assets below, and check it againstSHA256SUMS. - Stop
zebradand wait for it to exit. - Replace the binary and start
zebradagain.
From source
git fetch --tags && git checkout v7.2.0-beta.1
cargo build --release --locked -p zebradOr install it with cargo install --locked --git https://github.com/ZcashFoundation/zebra --tag v7.2.0-beta.1 zebrad. Add --features zmq to the build to enable ZMQ notifications.
Compatibility
| Component | Supported |
|---|---|
| Zcash network upgrades | All through NU7, which activates on Mainnet at height 3,543,405 |
| Network protocol version | 170190, the same as 7.1.0 |
| State format | 29.2.0, upgraded in place from 7.1.0's 29.0.0, and the same as 7.2.0-beta.0 |
| Release binaries | Linux x86-64 and ARM64 |
| Minimum Rust (source builds) | 1.91 |
| End of support | The same as 7.1.0: warnings from around 3 December 2026, above block 3,638,752; stops around 10 December 2026, near block 3,662,944 |
Contributors
Thank you to everyone who contributed to this release:
@arcayne, @arya2, @gustavovalverde, @oxarbitrage and @upbqdn