Skip to content

Non-deterministic queries, unhandled settlement errors, epoch stats underflow #885

Closed @unameisfine opened 2026-03-13 19:29 UTC 1 comment Updated 2026-04-27 22:46 UTC

Summary

Found several bugs during code review of x/inference/keeper/:

1. Non-deterministic gRPC query responses (consensus risk)

Three query handlers iterate Go maps and return results directly without sorting:

  • GetParticipantsFullStats — maps.Values(participants) returns unstable order
  • InferencesAndTokensStatsByModels — iterates map[string]StatsSummary
  • DebugStatsDeveloperStats — iterates statByTime and statByEpoch maps

Go map iteration order is randomized per-process. This means different nodes return the same data in different order, which can cause issues with deterministic replay and client-side caching.

2. uint64 underflow in GetSummaryLastNEpochs / GetSummaryLastNEpochsByDeveloper

epochIdFrom := effectiveEpochIndex - uint64(n)

When n > effectiveEpochIndex (e.g. early chain with epoch=1, requesting last 5 epochs), this wraps to MaxUint64, creating an enormous iterator range. This is reproducible on any chain in its first few epochs.

3. Ignored errors in SettleAccounts

Four error return values are silently discarded in the settlement path:

  • GetBitcoinSettleAmounts error logged but settlement continues with potentially uninitialized data
  • AddTokenomicsData return value completely ignored
  • SetSettleAmountWithGovernanceTransfer return value completely ignored
  • TransferOldSettleAmountsToGovernance error logged but nil returned to caller

4. Missing error in log call

shareWorkWithValidators logs "Unable to update participant" but omits the actual error value, making debugging impossible.

Files

  • x/inference/keeper/accountsettle.go
  • x/inference/keeper/developer_stats_aggregation.go
  • x/inference/keeper/msg_server_validation.go
  • x/inference/keeper/query_developer_stats_aggregation.go
  • x/inference/keeper/query_get_participant_current_stats.go

💬 Comments (1)

@unameisfine commented 2026-04-27 22:46 UTC

Closing — PR #886 was closed as part of refocusing on larger scoped contributions.

🔄 Auto-synced from Issue #885 every hour.