Week 44 · 2026

Highlights

Four PRs merged this week, expanding operational monitoring, adding demo recording tools and fixing two crash paths. There was no new Hydra release during the reporting period.

Operational Excellence (M3). The metrics overhaul merged Thursday, adding visibility into snapshot progress and missing signatures, input queues, broadcast and etcd delays, L1 operations and SQLite persistence. Updated Grafana dashboards and example Prometheus alerts help operators locate stalled work, including distinguishing a silent chain backend from a blocked processing loop. This completes M3.1, metrics exposed and queryable, delivered on 29 October ahead of its 10 November deadline.

The accompanying metrics reference explains what each measurement includes. Confirmation counters avoid counting repeated snapshot adoption, while restored state contributes no invented pre-restart latency. Existing metric names remain available as aliases. Confirmation rates are per node: adding rates across peers would count the same Head transactions multiple times. Alert thresholds and notification destinations still need configuring for each deployment.

Demo recording infrastructure also merged. A scripted three-party Head lifecycle checks that confirmation stops while a participant is suspended and resumes when it returns, then closes and fans out. The tooling records the Grafana dashboard alongside a timestamped scenario timeline and editable captions, making the behaviour easier to demonstrate. This exercises temporary unavailability, not recovery from a process crash.

Maintenance and DevX (M5). Two crash fixes landed. The Blockfrost fix handles tokens with valid empty asset names when the backend returns a null name, resolving the reported AssetNameMissing crash in 2.4.2. The deposit lookup fix passes the validated deposit input and output directly to incrementTx, removing the partial lookup that could crash on an empty UTxO or select an unintended output. It closes #2703.

Next. Four PRs opened this week remained open on 1 November. The UTxO capacity fix addresses a reported stall when a transaction takes the Head above its 4,095-output accumulator limit. Two further changes would reject underfunded deposits and prevent invalid contestation-period arithmetic. The SQLite and startup improvements address a deadlock when an event batch exceeds the write queue's capacity, reduce startup replay work and memory use, and make test filters fail when they select no tests.

A separate crash-consistency investigation opened to examine recovery across network delivery, persisted state and outgoing effects. The next step is to reproduce the suspected failure windows and establish which cases need automatic or operator-assisted recovery; the issue records source-level hypotheses, not completed crash experiments.