Skip to main content

Hydra Team Update

· 4 min read
Noon van der Silk
Noon van der Silk
Software Engineering Lead

High-level Summary​

The user-friendly hydra-node configuration file has been merged, letting operators run a head from a single YAML file and diff configs against peers at a glance. The rest of the period was mostly hardening: deposit and recovery chain observations are now scoped to the head that produced them (closing a long-standing cross-head contamination issue), the Blockfrost backend retries on HTTP errors instead of crashing and no longer chokes on some inline datums, and the event store's rotation logic got two fixes - a race between VACUUM and streaming client history, and a deadlock in the network layer's outbound message queue that could silently halt broadcasting.

What did the team achieve?​

  • Merged the user-friendly hydra-node configuration file: nodes can now be configured via YAML, making it easy to spot differences between peers' setups #2581, #2296
  • Scoped deposit and recover chain observations to the current head: Open/Closed states only record OnDepositTx events for their own headId and Idle silently drops the rest, while recovery of a deposit from a previous head keeps working in every state, including mid-fanout and while a new head is running #2743, #2606
  • Made the Blockfrost chain backend more resilient: HTTP errors are now retried instead of crashing the node #2729, #2477, and a follow-up fixed inline datum decoding that could throw on some UTxOs #2754, #2751
  • Fixed a race between the event log's VACUUM-based rotation and the API server streaming client history over the same SQLite connection, which could take a node down on reconnect #2755
  • Fixed a deadlock in the network layer's PersistentQueue: a false Eq result in popPersistentQueue could leave the broadcast queue stuck at capacity with no log output; it now pops unconditionally by index and emits PersistentQueueFull / PersistentQueueLoadFailed for visibility, and a failed file delete no longer crashes the network component #2742
  • Event log rotation now archives the outgoing database to a numbered old-state/hydra-<logId>.db file instead of deleting it, restoring the old file-based persistence's pre-rotation backup behaviour #2748, #2747
  • Added a fuel verification key option to hydra-tui so the funds tab can show both the head's ada and a separate fuel UTxO #2731
  • Moved the spec back into the hydra repo, dropping the separate flake input #2733

What's next?​

  • Merge some in-progress performance work: computing accumulator commitments via the rust-accumulator FFI instead of off-chain PlutusTx, batching etcd broadcasts onto shared connections, and raising the per-snapshot transaction cap - early results show sustained throughput roughly doubling on a 1k-UTxO head #2752
  • Make hydra-chain-observer version-aware by detecting script hashes, so a single process can watch heads from any historical Hydra version instead of running one per version #2740
  • Enforce --deposit-period consistency across head nodes by embedding it in the on-chain OpenDatum #2734
  • Accept PaymentExtendedKey (HD wallet keys) #2583
  • Work on the updated version of the spec (typst+Agda)