Week 33 · 2026

Highlights

8 PRs merged, 6 opened and 43 commits, with no issue activity. Only 6 of the 14 items carry a milestone this week; the rest is benchmark-variance and cleanup work gathered under Reactive.

Performance (M2). The Blockfrost backend got a dramatic speedup (#2821), replacing blind sleeps with condition-based polling: a full head lifecycle on preview drops from over an hour to about 13.5 minutes, publishing from ~10 minutes to ~32 seconds, and first seeding from ~30 minutes to ~40 seconds. A crash under sustained load was fixed (#2819): the node died on an uncaught EnhanceYourCalm from the etcd gRPC link, a regression from the W30 broadcast batching (#2778). And rejectLowDeposits merged from W32 (#2805), now a single pass instead of O(n²), with a property test over multi-output UTxOs.

Reactive. Two merges: protocol parameters tailored safely for 10 parties (#2815), and a tmpdir bug in the cluster tests (#2816), a follow-up to the withTempDir change below. Six PRs opened, most of them a push on benchmark reliability: an exploratory bench-variance PR (#2826) was closed in favour of two split-out pieces (#2827, #2828) that fix the bench's threadDelay units, pin seeds, compare e2e benchmarks as same-machine pairs against the merge base, and report raw series and GHC RTS work counters. Alongside those: library re-use fixes, including a latent metrics bug where mistyped Prometheus lookups silently did nothing (#2825), dead code removal (#2823), and binding deposits to the snapshots that authorize them (#2820).

Maintenance and DevX (M5). Both PRs opened in W32 merged: withTempDir now cleans up atomically on success (#2813), and the flaky etcd, TUI and chain-sync tests were fixed (#2811). The etcd one is worth a note: the test port sentinels are now close-on-exec, so spawned subprocesses no longer inherit and pin bound ports, a bug behind 16 of the last 25 hydra-node CI failures.

KPI & Outcome Achievement (M6). The rework of this tracker for Hydra merged (hydra-updates#1).