zakura-core/zakura v1.8.8

on GitHub

Added

• With network.zakura.trace_dir set, legacy peer connections write
  legacy_gossip.jsonl: each block announcement sent to a peer, the peer's
  request for that block, and the block being sent to it or found missing.
  They also write the announcements each peer sends, up to 16 a minute for
  each connection. Requests for blocks this node did not announce to that
  peer are not written, so what a peer sends cannot make the file grow
  without bound. Rows never hold a peer's address. The ZAKURA_TRACE_PEER_NAMES
  environment variable (name=ip pairs separated by commas) names the peers an
  operator wants to tell apart
  (#2446).

Changed

• network.unconditional_peers now also keeps its peers connected. Zakura
  dials every entry that names one address, and dials it again whenever no
  connection to that peer is open in either direction. An entry can give the
  port to dial, such as "192.0.2.1:8233". An entry without a port is dialed
  on the network's default port, and a CIDR range is only dialed when it covers
  a single address. Each dialed address adds one outbound and one inbound
  connection slot on top of the connection limits. The inbound slots are a pool
  that only listed peers can use, so a node whose other inbound slots are full
  still accepts a listed peer that dials it. One IP holds at most
  max_connections_per_ip of these slots. Zakura's other outbound dials leave
  a listed address's IP the room the listed peer needs. Zakura also
  lets these peers reconnect inbound without its usual wait between connections
  from one IP. IP addresses that an existing config
  already lists start being dialed after the upgrade, and a listed peer that
  accepts no connections costs one failed dial every ten minutes once its delay
  has grown
  (#2389).
• Start up to 50 legacy startup peer connection attempts at once, and start the
  next one as soon as an attempt completes, instead of spacing attempts 100 ms
  apart. The total startup peer target and per-IP limits still apply
  (#2396).
• Reduced the CPU a node spends keeping its peer address metrics up to date,
  which was the largest single CPU cost of an idle node at the chain tip
  (#2409).
• Reduced CPU use when encoding blocks and transactions with Orchard, Ironwood
  or Sapling bundles, and when serving blocks to peers, by no longer copying
  each bundle before writing it
  (#2421).
• Legacy connections now reuse recently calculated large-block checksums,
  reducing repeated hashing when several peers request the same block.
  Concurrent sends can still each calculate a checksum before it is remembered
  (#2424).
• The error label of zakura.net.in.errors and the reason label of
  mempool.failed.verify.tasks.total now name the error kind, such as io,
  parse, invalid or download_failed, instead of the full error message.
  Peers could otherwise create an unbounded number of metric series. Update
  dashboards and alerts that match full error messages on these labels
  (#2439).
• The start time in a trace row's process_trace_id is now the moment the
  row's ts counts from, so adding the two gives the row's wall-clock time.
  Before, the two were read at different moments and could be seconds apart
  (#2446).
• With [network.zakura] trace_dir set, zakurad now moves to the file at a
  trace table's path within about a second when its own file was renamed,
  deleted or replaced, so trace files can be rotated by renaming them without
  losing rows
  (#2449).
• The continuous profiler finishes reading a renamed trace table across bounded
  reads before switching files, truncating it, or rotating again, preserving rows
  in large backlogs even while the node is delayed before opening a replacement.
• Legacy P2P message checksums are calculated with ring's SHA-256, which
  uses the vector instructions of CPUs that lack SHA instructions. On such a
  mainnet node, serving old blocks to syncing peers used 32% less CPU per MB
  and one connection delivered blocks about 50% faster. Checksum values are
  unchanged
  (#2450).
• Calculate block-header hashes, transaction Merkle-tree hashes, and transaction
  IDs for versions one through four with ring, preserving the existing hash
  values (#2457).
• Made Sprout JoinSplit proof and signature verification opt-in with the sprout
  Cargo feature. Builds using post-NU7 snapshots do not need it; verifying
  pre-NU7 blocks above the last checkpoint, including Testnet sync from genesis
  with the current checkpoints, requires --features sprout
  (#2466).
• Sped up mempool cleanup after new blocks by finding conflicting transactions
  through spend-owner indexes instead of scanning the entire pool
  (#2467).
• Sped up mempool expiry cleanup after new blocks by indexing expiry heights
  instead of scanning transactions that are not due to expire
  (#2468).
• Reduced mempool cleanup work after new blocks by batching transaction removals,
  dependency-order updates and metric reports while retaining unmined children
  of confirmed transactions
  (#2469).
• Sped up pending transaction gossip cleanup after mining or invalidation by
  removing changed IDs directly instead of scanning the advertisement backlog
  (#2470).
• Removed the retired Sprout nullifier index from mempool storage. Sprout
  JoinSplits are now rejected by mempool policy even on custom pre-NU7 networks;
  historical block verification remains available with the sprout feature
  (#2471).
• Reduced mempool ordering maintenance during admission and block cleanup by
  caching ordering scores and refreshing only surviving transactions whose
  priorities change
  (#2473).
• Reduced how often the state service exports RocksDB metrics from every 30
  seconds to every 5 minutes. The zakura.state.rocksdb.* gauges and the
  size_on_disk RPC estimate now refresh at that interval
  (#2475).
• Retained the approved v1.8.8 node patch version for the pending operational,
  performance, tracing, and build changes. Builds that verify pre-NU7 blocks
  above the last checkpoint still require the documented sprout feature
  (#2466,
  #2479).

Fixed

• Scale NU7 reissuance to custom halving intervals so compatible Zebra and Zakura nodes
  calculate the same subsidy. Existing custom NU7 chains need coordinated deployment
  and a reset or explicit migration if they used the previous payout rule
  (#1591).
• Restored NODE_NETWORK advertising for pruned nodes so peers can discover,
  reconnect to, and download recent blocks from them. Pruned nodes additionally
  advertise an experimental NODE_PRUNED flag. While an updated node is still
  syncing, it downloads blocks from peers without the flag and waits for one of
  them when they are all busy. It asks a pruned peer only when it is connected
  to no such peer, or when those peers leave block requests unanswered. It
  sends getblocks to peers without the flag too, because pruned peers leave
  a getblocks for deleted blocks unanswered. Near
  the network tip it treats pruned peers like any other peer. getnetworkinfo
  now reports the node's actual
  localservices, including these flags
  (#2382).
• Cancel spawned legacy handshakes when their caller is dropped, so a connection
  timeout releases the stream and connection reservation immediately
  (#2396).
• A restarted node now dials the peers in its legacy peer cache when every seed
  DNS lookup fails, instead of retrying DNS forever without connecting
  (#2399).
• The legacy peer cache now keeps only peers that responded to this node in
  the last 3 hours, not addresses learned from gossip, so a restart spends its
  first connection attempts on peers that recently accepted a connection
  (#2404).
• Restored the published consensus crate's Halo2 error type and conversion,
  preserving compatibility with its current stable release
  (#2413).
• Read gossip trace peer names from network.zakura.trace_peer_names instead
  of an environment variable that could prevent the node from starting. The
  fleet installer writes the mapping and removes the obsolete environment
  override after verifying the configuration
  (#2474).

Security

• A legacy peer's getblocks reply to a FindBlocks request now keeps at
  most 500 block hashes, the most zakurad and zcashd serve, so one peer can
  no longer make the syncer process up to 50,000 hashes per response
  (#2437).