ZF Engineering Update: 3 - 16 November 2026

NU7 activated on Mainnet on 6 November, at height 3,543,405. Thank you to every operator who upgraded to Zebra 7.0.0 in time.Most of our work over these two weeks went into stabilising the Zebra Ándale performance improvements for miners. We also helped nodes recover from lost peers and database problems, gave our own nodes better tools to explain a stall, and made our test runs faster and cheaper. Everything below was merged into Zebra’s main branch and is not yet part of a published release.

CI / DevOps

Our own nodes now send more detail to Sentry, the error tracking service we use, so we can compare next and main nodes block by block and see why a node stalls: how long each new block takes to become the tip, why sync restarted and which kind of block error caused it, block counts, mempool size and disk size, and stack traces that point at Zebra’s own code. Unit tests in the merge queue now run on a larger machine and take about 4.5 minutes instead of about 14. After a database format change, CI now starts from the previous version’s saved state instead of syncing from scratch, which took about 11 hours for the NU7 release. Failed integration tests on main are now reported separately from infrastructure failures. Release PRs now take their version numbers from the change notes each PR adds. We also stopped running tests on Windows: the Windows support code stays, but CI no longer tests it.

Other Zebra work

A node whose first peers all fail at startup now retries them, and its DNS seeders, at most once every 5 minutes, instead of staying without peers until someone restarts it. It also starts from its saved peers when DNS fails. Zebra no longer shares addresses it has never connected to as if they served blocks. When the state database fails to open, zebrad now shows the real error instead of always suggesting that another zebrad process is running, so a missing disk, a permission problem or a full disk is easier to spot. Commands that only read the database, such as zebrad tip-height and zebrad copy-state, no longer leave a temporary directory behind after each run. If stored block data is incomplete, Zebra now reports an error instead of returning part of a block.

We removed the experimental elasticsearch feature. It made syncing much slower and we found no one using it, so operators who built Zebra with it should remove its settings from their config. Zebra also does less work for each block: it computes the Merkle roots of large blocks in parallel, and it no longer copies whole blocks into its tracing data when it commits them. Finally, the Zebra book has a new look, including a proper dark mode.

Work in progress

Our main focus is performance. Upcoming Zebra releases will bring the speed improvements from Zebra Ándale, such as faster block templates, block relay and block verification, to every operator. Beyond that, we are working on helping Zebra sync faster from scratch, verify each block with less work, and optionally use less disk space. You can follow this work in our performance tracking issue.

Work on other repositories

Faster verification upstream. To help Zebra verify blocks faster, we contributed speedups to the cryptography libraries it uses: signature batches in reddsa, point decoding in pasta_curves, proof verification in halo2 and bellman, and transaction parsing in librustzcash, where a large block now parses in 7.5 ms instead of 21.8 ms in a benchmark. Several of these are adapted from work the Zakura team did first. Zebra will use them once they are released, in the order set out in this tracking issue.

Zeal. We rebuilt Zebra Launcher, our desktop app for running Zebra, and renamed it Zeal (a group of zebras is a zeal). It installs and looks after zebrad for you, and now includes a block explorer that only reads from your own node, an experimental light client server, and a choice of stable, beta or nightly zebrad. It has not been released yet.

Zinder. Our indexer, which serves light wallets through a lightwalletd compatible interface, now reads NU7 blocks and is tested against a Zebra 7 Regtest chain through NU7. We published several releases, the latest being v0.15.1.

Zeeder. Our DNS seeder deployments now use zfnd/zeeder as their main image name, the next step in its rename. The image is the same, and zfnd/dnsseeder still gets every release.

Thanks to external contributors

Thanks to @Ravan002 for making Zebra’s RPC server answer requests without valid credentials with a clear 401 error instead of an empty reply, and to @Anirudh-712 for making zebrad refuse a Testnet or Regtest config that activates a network upgrade it can’t follow when it starts, instead of crashing when the chain reaches that upgrade.

And also thanks to @robustfengbin and @zacheryasc for their contributions.

10 Likes