bls: BlsManager stores context.Background() — DKG gRPC calls have no cancellation or timeout #908

Closed @Mayveskii opened 2026-03-17 23:21 UTC 1 comment Updated 2026-04-28 18:11 UTC

Summary

NewBlsManager stores context.Background() as a struct field bm.ctx. This means two gRPC calls in the DKG dealer path run without any timeout and cannot be cancelled on node shutdown.

Affected Code

decentralized-api/internal/bls/manager.go:133

func NewBlsManager(recorder cosmosclient.InferenceCosmosClient) *BlsManager {
    return &BlsManager{
        ctx: context.Background(), // Use background context for chain queries
        ...
    }
}

decentralized-api/internal/bls/dealer.go:392,404 — no timeout wrapper:

// No WithTimeout — hangs indefinitely if chain RPC is slow/unavailable
grantees, err := queryClient.GranteesByMessageType(bm.ctx, ...)
participant, err := queryClient.InferenceParticipant(bm.ctx, ...)

Contrast with manager.go:165 which correctly adds a 60s timeout:

ctx, cancel := context.WithTimeout(bm.ctx, 60*time.Second)
defer cancel()

Impact

1. DKG epoch block: If the chain RPC node is slow or unreachable during EventKeyGenerationInitiated, the worker goroutine inside event_listener blocks indefinitely on GranteesByMessageType or InferenceParticipant. This causes the node to miss the entire DKG window and drop out of consensus for that epoch.

2. No graceful shutdown: On SIGINT, main.go cancels the root ctx via defer cancel(), which propagates to listener.Start(ctx). However bm.ctx = context.Background() is independent — in-flight BLS gRPC calls are never interrupted, delaying process exit.

3. No Stop/Close on BlsManager: There is no shutdown method to cancel pending operations.

Root Cause

The context is stored at construction time as Background() instead of being passed from the caller. main.go already has a properly-scoped ctx with cancel:

// main.go:145
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
// ...
blsManager := bls.NewBlsManager(*recorder)  // ctx NOT passed
listener := event_listener.NewEventListener(..., cancel, blsManager)

Fix

// manager.go
func NewBlsManager(ctx context.Context, recorder cosmosclient.InferenceCosmosClient) *BlsManager {
    return &BlsManager{
        ctx: ctx,  // propagate caller context
        ...
    }
}

// main.go
blsManager := bls.NewBlsManager(ctx, *recorder)

This ensures: - GranteesByMessageType and InferenceParticipant respect the node lifecycle context - On SIGINT, pending BLS queries are cancelled immediately - Future callers can add per-call timeouts via context.WithTimeout(bm.ctx, ...)

Verified

Statically verified via AST analysis of commit history (904 commits). No existing fix found in HEAD. The dealer.go calls at lines 392 and 404 have no timeout wrapper, unlike the correctly-handled call in manager.go:165.

Severity: Medium-High — affects node availability during DKG under degraded RPC conditions.

---

💬 Comments (1)

@x0152 commented 2026-04-28 18:11 UTC

for the same reason as #909

---

🔄 Auto-synced from Issue #908 every hour.