Release candidate for 1.9.0. Static transaction files change backups and rollback; experimental relay gains safeguards. 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
- State format v32 and disk space. First start upgrades to state format v32; downgrading needs a fresh sync or pre-upgrade backup. Leave about 11 GB free on Mainnet archive nodes for the move. If a static-file write fails, free space and restart
zakurad(#2146). - Include static files in backups. RocksDB checkpoints no longer contain finalized transaction bodies. Include
static_files: copy the whole state directory withzakuradstopped or take a filesystem snapshot. Don't hard link static files into a restore (#2146). - Coordinate read-only RPC secondaries. Upgrade secondaries together with their primary: a v31 secondary cannot follow a v32 primary. Stop secondaries around
zakurad rollback-state, which shortens static files in place (#2146). - Pruned nodes stop serving old bodies. After the first static-file move, pruned nodes that started as archives treat transaction bodies below their pruning marker as pruned. Check RPC consumers that rely on those bodies (#2146).
Highlights
- Security: Experimental:
tx-relayshrinks peer request windows for orphan transactions and latenotfoundreplies, and limits fetches to 250 of 500 pool slots so peers cannot crowd out other submissions (#2634). - Experimental:
tx-relaylimits a new connection to its peer's 544 KB opening allowance plus what 256 KB/s earns since the handshake, instead of about 44 MB (#2637). - Experimental: core mempool cleanup after a new block scales with the block's size instead of scanning the whole pool for conflicting and expired transactions (#2636).
Full changelog
Changed
- Finalized transaction bodies now move out of RocksDB into append-only
static files in the state directory, in the background and at least 50
blocks at a time. On first start the node upgrades its database in place
to state format v32. Older releases can't open the upgraded database, so
going back means a fresh sync or restoring a backup from before the
upgrade (#2146). - A Mainnet archive node grows by about 11 GB at the peak of the move,
because the new static files sit next to the old RocksDB data until
RocksDB frees it. Leave that much free disk. If a static file write
fails, for example on a full disk, moves stop until zakurad restarts, so
free space and restart it
(#2146). - Back up the
static_filesdirectory along with RocksDB, because a
RocksDB checkpoint no longer holds finalized transaction bodies. Copy the
whole state directory with zakurad stopped or take a filesystem snapshot,
and don't hard link static files into a restore
(#2146). - Upgrade read-only RPC secondaries together with their primary, and stop
them aroundzakurad rollback-state. A v31 secondary can't follow a v32
primary, and rollback shortens static files in place
(#2146). - A pruned node that started as an archive used to keep serving old
transaction bodies below its pruning marker. After the first static file
move it treats them as pruned
(#2146). - With the experimental core mempool, a new block no longer scans the whole
pool for conflicting and expired transactions, so its cost follows the
block's size instead of the pool's
(#2636). - With the experimental
tx-relayenabled, the
zakura.p2p.tx_relay.inv_idscounter is now
zakura.p2p.tx_relay.have_ids, because the announcement message is now
calledHave. Update dashboards that read the old name. The wire format
is unchanged (#2665).
Fixed
- Header sync now takes equal peer offers in arrival order, so peers that
reconnect under fresh keys or overstate their chain work can't keep an
honest peer's headers from being requested
(#2638). - With the experimental
tx-relayenabled, a new connection accepts at
most its peer's 544 KBHaveopening allowance plus what 256 KB/s earns
since the handshake, instead of about 44 MB
(#2637).
Security
- With the experimental
tx-relayenabled, orphan transactions and late
notfoundanswers now shrink a peer's request window, and tx-relay
fetches can hold at most 250 of the pool core's 500 slots, so peers can't
crowd out local submissions, legacy peers or re-admissions after a reset
(#2634).
Also in CHANGELOG.md.