Week 37 · 2026

Highlights

4 PRs merged, 3 opened and 36 commits across two repos. A consolidation week after the 2.4.0/2.4.1 pair: the merged work is a dependency bump, a test fix and last week's two leftovers, while the substance sits in the three PRs opened, two of which have since merged. Performance also closed out as a workstream (see M6). Ten issues were opened, among them five findings on unbounded memory growth and crash loops (#2862, #2863, #2865, #2866, #2867) and, on Saturday, a user's stuck deposit (#2871) taken up in W38; fifteen were closed, three by merges (#2848, #2126, #2167) and twelve in a sweep of the backlog on Monday and Tuesday, five of those as not planned.

Reactive. The etcd network component moved up a version on both sides (#2857). The bundled etcd goes from 3.5.25 to 3.6.6, and the gRPC client stack from grapesy 1.0.1 to 1.2.0 and http2 5.3.9 to 5.4.4. The point of the client bump is #2167, where putMessage blocked indefinitely after five to ten thousand large messages: the underlying http2 connection stall is still reproducible upstream when puts share a connection with the watch stream, but on the new stack it surfaces as a reconnectable ServerDisconnected error rather than blocking silently and forever. The dedicated, recycled, timeout-guarded broadcast connections added earlier for that issue remain the operative fix, re-verified here with repeated 8000 by 100KiB broadcast soaks. Operators upgrading should note that etcd supports rolling upgrades from 3.5.20 onward, but that the etcd data directory upgrade is one way. The same PR corrected the docs: hydra-node embeds etcd and uses it by default, so the tutorial's download step is gone and a separate install is only needed with --use-system-etcd (#2126). The nightly Blockfrost test also got a one-line fix (#2868): it waited on a bounded waitMatch for a tick past the contestation deadline, and that bound is shorter than the contestation period, so it now sleeps through the period first.

Both PRs left over from last week merged on the Monday this week opened, and were described in W36 and W35 respectively: the libs/ extraction (#2854) and the partial fanout chunk-size search (#2849).

Of the three opened, the fanout wedge is the one to note (#2858, for #2855). A PartialFanout naming the whole remaining set behaved differently depending on where it arrived: from a freshly closed head it was rerouted to a full fanout, but from FanoutProgress with nothing yet distributed the equivalent handler had no such guard and sent it out as a non-final step. That empties the head while the on-chain datum still says Closed, which leaves it both unfinalizable (the next step would have to be a zero-output final fanout, which the validator rejects) and unrevertable (the revert requires the distributed set to be empty, and it no longer is), so the tokens can never be burned. The fix moves the decision out of the client handlers into nextFanoutStep, which now returns the step to take, so every producer of a fanout step goes through the same choice and the guard cannot be missed at one entry point again. That also covers a path no client input reaches: a node that persisted a whole-remainder selection rebuilds it verbatim on replay, where a rollback re-post would have fired the same wedging step. Merged the Tuesday after the week closed.

Alongside it, adversarial model testing (#2864) adds random deposits and decommits under layer 2 transaction spam, divergent-fork rollbacks, per-node network delivery latency faithful to etcd's total order, and node restart and fail-recovery, all under liveness assertions. It found and fixed a real wedge of its own: a ReqSn arriving before a party had observed the deposit hard-errored and stalled the head, and is now parked with a TTL. It also fixed several fidelity bugs in the mock chain that the new tests exposed, namely non-atomic chain steps, queue ordering and silently dropped transactions. Merged the Wednesday after. Still open: two tests for minting inside a head with known and unknown minting policies, which also check that an error thrown during transaction construction does not bring the node down (#2860, merged in W40).

KPI & Outcome Achievement (M6). More than the usual bookkeeping this week. The W36 note published, and Performance (M2) is now marked done: decreased memory usage (M2.2) is recorded as delivered on 2026-08-17 by the CBOR database and binary API pair (#2767, #2762), joining the 10x TPS work (M2.1, delivered 2026-07-23) to close the workstream, both ahead of their deadlines. Three milestone dates also moved earlier: metrics exposed and queryable (M3.1) from 2026-11-25 to 2026-11-10, and the hydra-recipes repo (M3.2) and continued compatibility with cardano-node releases (M5.3) both from 2026-12-15 to 2026-11-30.