feat: High availabilty via RAFT

[alpe]

Implement failover via RAFT

• Improve Cache startup/shutdown with parallelization
• Publish to RAFT cluster in executor
• Sync DB after each block created in executor
• Add new RaftReceiver to sync when in aggregator follower mode
• Introduce failoverState to switch between follower/leader mode
• Provide RAFT node details via http endpoint

[github-actions]

The latest Buf updates on your PR. Results from workflow CI / buf-check (pull_request).

Build	Format	Lint	Breaking	Updated (UTC)
✅ passed	⏩ skipped	✅ passed	✅ passed	Dec 3, 2025, 4:47 PM

[codecov]

Codecov Report

❌ Patch coverage is 41.48816% with 519 lines in your changes missing coverage. Please review.
✅ Project coverage is 62.05%. Comparing base (ded4f34) to head (695324e).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
pkg/raft/node.go	12.50%	168 Missing ⚠️
pkg/raft/node_mock.go	45.08%	74 Missing and 21 partials ⚠️
block/internal/syncing/raft_retriever.go	0.00%	60 Missing ⚠️
node/full.go	32.81%	36 Missing and 7 partials ⚠️
node/failover.go	74.45%	22 Missing and 13 partials ⚠️
block/internal/syncing/syncer.go	28.88%	30 Missing and 2 partials ⚠️
block/internal/executing/executor.go	6.89%	23 Missing and 4 partials ⚠️
pkg/raft/election.go	79.26%	12 Missing and 5 partials ⚠️
pkg/rpc/server/http.go	6.66%	13 Missing and 1 partial ⚠️
block/internal/syncing/assert.go	57.89%	4 Missing and 4 partials ⚠️
... and 9 more

Additional details and impacted files

@@ Coverage Diff @@## main #2836 +/- ##
==========================================
- Coverage 64.53% 62.05% -2.48% 
==========================================
Files 81 88 +7 Lines 7382 8083 +701 ==========================================
+ Hits 4764 5016 +252 - Misses 2076 2490 +414 - Partials 542 577 +35 

Flag	Coverage Δ
combined	62.05% <41.48%> (-2.48%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[alpe]

@claude review again

[claude]

Claude finished @alpe's task —— View job

---

Comprehensive Code Review

I've analyzed this high-availability RAFT implementation PR. Here's my detailed review across multiple dimensions:

---

1. Architecture & Design Patterns

✅ Strengths

Modular RAFT Integration: The separation of concerns is well-executed:

• pkg/raft/node.go: Core RAFT node abstraction
• pkg/raft/election.go: Leader election orchestration with dynamic role switching
• node/failover.go: Failover state management decoupled from business logic
• Integration points are well-defined through interfaces (e.g., Runnable, RaftNode)

Clean State Machine: The DynamicLeaderElection.Run() in pkg/raft/election.go:54-152 implements a robust state machine with clear transitions between leader/follower modes.

Factory Pattern: The use of factory functions for leader/follower initialization (leaderFactory, followerFactory in node/full.go:97-107) enables clean separation and testability.

⚠️ Issues & Recommendations

1. JSON Serialization Performance Concern (pkg/raft/node.go:211)

data, err:=json.Marshal(state) // todo:use protobuf

Critical Issue: At 100ms block times, JSON serialization overhead can impact consensus performance significantly. This TODO should be prioritized.

Recommendation:

• Implement protobuf serialization immediately
• Add benchmarks comparing JSON vs protobuf performance
• Estimate impact: JSON typically 2-5x slower than protobuf for structured data

2. Silent Message Drops (pkg/raft/node.go:291-297)

select {
casef.applyCh<-RaftApplyMsg{Index: log.Index, State: &state}:
default:
f.logger.Warn().Msg("apply channel full, dropping message")
}

Issue: Dropped messages could lead to nodes falling behind without recovery mechanism.

Recommendation:

• Add metrics to track dropped messages
• Implement backpressure mechanism
• Consider increasing buffer size or making it configurable
• Document expected behavior when messages are dropped

3. Bootstrap Configuration Validation (node/full.go:90-123)

case!nodeConfig.Node.Aggregator&&!nodeConfig.Raft.Enable:
// sync mode without raft
default:
returnnil, fmt.Errorf("raft config must be used in sequencer setup only")

Issue: Error message is unclear - doesn't explain invalid configuration.

Recommendation: Improve error message to explain valid configurations.

---

2. Error Handling & Resilience

⚠️ Critical Issues

1. Leadership Transfer Failure Handling (pkg/raft/election.go:99-102)

iferr:=d.node.leadershipTransfer(); err!=nil&&!errors.Is(err, raft.ErrNotLeader) {
returnerr// Node stops entirely
}

Issue: If leadership transfer fails (e.g., no suitable leader), the node stops. This could cause cluster instability.

Recommendation:

• Implement retry logic with exponential backoff
• Add configurable timeout for leadership transfer
• Consider degraded mode instead of full stop
• Log detailed error context for debugging

2. Missing Context Propagation (block/internal/syncing/raft_retriever.go:103-134)

func (r*raftRetriever) consumeRaftBlock(ctx context.Context, state*raft.RaftBlockState) error {
// ctx is passed but not checked in tight loops
}

Issue: Long-running operations don't check context cancellation, potentially blocking shutdown.

Recommendation: Add context checks in processing loops.

3. Database Flush Without Error Handling (block/internal/executing/executor.go:434-439)

ife.raftNode!=nil {
headerBytes, err:=header.MarshalBinary()
iferr!=nil {
returnfmt.Errorf("failed to marshal header: %w", err)
}

Missing: The code after this (from your review) should verify database flush completed before RAFT broadcast. If RAFT broadcast succeeds but DB flush fails, state inconsistency occurs.

Recommendation:

• Ensure DB flush completes before RAFT broadcast
• Add explicit error handling for flush operations
• Consider transaction boundaries

4. Validation Gap in RAFT State (pkg/raft/types.go:17-24)

func (sRaftBlockState) assertValid(nextRaftBlockState) error {
ifs.Height>next.Height {
returnfmt.Errorf("invalid height: %d > %d", s.Height, next.Height)
}
// ...
}

Issue: Only checks monotonic increase, doesn't validate:

• Gap detection (height jumps)
• Hash chain continuity
• Timestamp reasonableness (could be far future)

Recommendation: Add comprehensive validation including gap detection and chain continuity checks.

---

3. Security Considerations

⚠️ Issues

1. No Mutual TLS for RAFT Communication (pkg/raft/node.go:87-89)

transport, err:=raft.NewTCPTransport(cfg.RaftAddr, addr, 3, 10*time.Second, os.Stderr)

Issue: RAFT cluster communication is unencrypted. In production, this exposes:

• Block data in transit
• Potential MITM attacks
• No node authentication beyond configuration

Recommendation:

• Implement TLS for RAFT transport
• Add certificate-based authentication
• Document security requirements in production
• Consider network isolation requirements

2. No Rate Limiting on RAFT Apply (pkg/raft/node.go:279-300)
Issue: No protection against DoS through excessive RAFT proposals.

Recommendation:

• Add rate limiting per node
• Implement backpressure mechanisms
• Monitor proposal rates

3. Peer Address Validation (pkg/raft/node.go:344-364)

// we can skip address validation as they come from a local configuration

Issue: Comment suggests trust in configuration, but misconfiguration can cause issues.

Recommendation: Add basic sanity checks (valid port ranges, non-empty addresses, reachability probes).

---

4. Concurrency & Resource Management

⚠️ Issues

1. Potential Goroutine Leak (pkg/raft/election.go:66-82)

startWorker:=func(namestring, workerFuncfunc(ctx context.Context) error) {
workerCancel()
workerCtx, cancel:=context.WithCancel(ctx)
workerCancel=cancelwg.Add(1)
gofunc(childCtx context.Context) {
deferwg.Done()
iferr:=workerFunc(childCtx); err!=nil&&!errors.Is(err, context.Canceled) {
select {
caseerrCh<-fmt.Errorf(name+" worker exited unexpectedly: %s", err):
default: // do not block
}
}
}(workerCtx)
}

Issue: If errCh is full and worker errors occur, errors are silently dropped. This could mask critical failures.

Recommendation:

• Increase error channel buffer size
• Log when errors can't be sent
• Consider using a more robust error aggregation pattern

2. Race Condition Risk (block/internal/syncing/syncer.go:177-189)

func (s*Syncer) isCatchingUpState() bool {
returnlen(s.heightInCh) !=0||func() bool {
currentHeight, err:=s.store.Height(s.ctx)
// ...returns.raftRetriever!=nil&&s.raftRetriever.Height() >currentHeight
}()
}

Issue: Multiple store height checks without synchronization could race with concurrent updates.

Recommendation: Add mutex or use atomic operations for height comparisons.

3. Shutdown Ordering (node/failover.go:177-232)

deferfunc() {
shutdownCtx, done:=context.WithTimeout(context.Background(), 3*time.Second)
deferdone()
_=f.rpcServer.Shutdown(shutdownCtx)
}()

Issue: Hard-coded 3-second timeout may not be sufficient for graceful shutdown under load.

Recommendation:

• Make shutdown timeout configurable
• Implement phased shutdown (stop accepting new requests → drain existing → force close)
• Log shutdown progress

4. BoltDB Concurrent Access (pkg/raft/node.go:67-74)

logStore, err:=raftboltdb.NewBoltStore(filepath.Join(cfg.RaftDir, "raft-log.db"))
stableStore, err:=raftboltdb.NewBoltStore(filepath.Join(cfg.RaftDir, "raft-stable.db"))

Note: BoltDB doesn't support concurrent writers. Ensure RAFT library handles this correctly (it likely does, but worth documenting).

---

5. Performance & Resource Efficiency

⚠️ Issues

1. Memory Allocation in Hot Path (pkg/raft/node.go:206-222)

func (n*Node) Broadcast(_ context.Context, state*RaftBlockState) error {
data, err:=json.Marshal(state) // Allocates every call// ...
}

Issue: JSON marshaling allocates on every block production (every 100ms).

Recommendation:

• Use sync.Pool for buffer reuse
• Switch to protobuf (lower allocation)
• Profile allocation hotspots

2. Unbounded Channel Operations (block/internal/syncing/raft_retriever.go:62)

applyCh:=make(chan raft.RaftApplyMsg, 1)

Issue: Buffer size of 1 is very small; could cause blocking on high throughput.

Recommendation:

• Make buffer size configurable based on expected block rate
• Monitor channel saturation
• Document expected throughput

3. Blocking Operations in Election Loop (pkg/raft/election.go:96)

raftSynced:=d.node.waitForMsgsLanded(d.node.Config().SendTimeout) ==nil

Issue: Blocks election loop, delaying leadership decisions.

Recommendation: Consider async status checks or timeout optimization.

---

6. Testing Coverage & Quality

✅ Strengths

• Good unit test coverage for helper functions (splitPeerAddr, deduplicateServers in pkg/raft/node_test.go)
• Comprehensive election state machine tests (pkg/raft/election_test.go)
• E2E failover test (test/e2e/failover_e2e_test.go)

⚠️ Gaps

1. Missing Integration Tests:

• No tests for RAFT network partitions (split-brain scenarios)
• No tests for concurrent leadership changes
• Missing crash recovery tests
• No tests for state validation edge cases

2. Low Coverage on Critical Paths (from codecov report):

• pkg/raft/node.go: 12.50% coverage (168 lines missing)
• block/internal/syncing/raft_retriever.go: 0.00% coverage (60 lines missing)
• node/failover.go: 74.45% coverage (35 lines missing)

Recommendation:

• Add chaos engineering tests (network delays, partitions, crashes)
• Test leadership transfer under various loads
• Test recovery from various failure modes
• Increase coverage for error paths

3. Mock Coverage (pkg/raft/node_mock.go):

• 45% coverage for mocks suggests tests aren't exercising all code paths

4. Missing Benchmarks:

• No performance benchmarks for RAFT broadcast latency
• No benchmarks for failover time
• No load testing for sustained throughput

---

7. Documentation & Code Quality

⚠️ Issues

1. Insufficient Documentation:

• SetApplyCallback (pkg/raft/node.go:274) has good docs, but most RAFT methods lack detailed documentation
• No architecture documentation for RAFT integration
• Missing operational runbook (how to recover from failures, add/remove nodes, monitor health)

2. Magic Numbers:

ticker:=time.NewTicker(300*time.Millisecond) // election.go:83ReadTimeout: 10*time.Second, // failover.go:158

Recommendation: Extract to named constants with explanatory comments.

3. Logging Inconsistency:

• Mix of Debug/Info/Warn/Error levels
• Some critical operations lack logging (e.g., database flushes)
• Consider structured logging for correlation across cluster

---

8. Specific Code Issues

🔴 Critical

1. Double Sign Risk (pkg/raft/election.go:95-104)

raftSynced:=d.node.waitForMsgsLanded(d.node.Config().SendTimeout) ==nilif!raftSynced||!runnable.IsSynced(d.node.GetState()) {
d.logger.Info().Msg("became leader, but not synced. Pass on leadership")
iferr:=d.node.leadershipTransfer(); err!=nil&&!errors.Is(err, raft.ErrNotLeader) {
returnerr
}
continue
}

Good: Checks sync status before becoming leader to prevent double signing.

Issue: If leadership transfer fails, the node might stop entirely, reducing cluster capacity.

Recommendation: Implement more sophisticated handling - perhaps stay as follower until properly synced.

2. Store Flush Timing (per review comment thread)
The discussion around Commit() and flush behavior is important. Current implementation flushes before RAFT broadcast, which is correct, but:

Recommendation:

• Document the critical ordering: DB flush → RAFT broadcast → P2P broadcast
• Add assertions/tests to verify this ordering
• Consider making SyncWrites configurable with clear performance trade-off documentation

⚠️ High Priority

3. Peer Duplication Check (pkg/raft/node.go:137-139)

ifsvrs:=deduplicateServers(cfg.Servers); len(svrs) !=len(cfg.Servers) {
returnfmt.Errorf("duplicate peers found in config: %v", cfg.Servers)
}

Issue: Deduplicates by ID only, but error message suggests checking Address too. Could have same ID with different addresses.

Recommendation: Check both ID and Address for duplicates, or clarify which is the key.

4. Error Propagation (block/internal/executing/executor.go:330-332)

ife.raftNode!=nil&&!e.raftNode.IsLeader() {
returnerrors.New("not raft leader")
}

Issue: Generic error message doesn't help with debugging.

Recommendation: Include current leader ID and node state in error.

---

9. Configuration & Operational Concerns

⚠️ Issues

1. No Default Validation (pkg/config/config.go:322-325)

iferr:=c.Raft.Validate(); err!=nil {
returnerr
}

Question: What does Raft.Validate() check? Need to review config validation for:

• Heartbeat timeout vs election timeout ratios
• Send timeout vs block time relationships
• Minimum cluster size

2. Missing Health Checks:

• No RAFT-specific health endpoint beyond basic RPC
• Should expose: current role, cluster size, sync status, last applied index

Recommendation: Add /raft/status endpoint with detailed cluster health.

3. Metrics Gaps:

• No RAFT-specific metrics (proposals/sec, apply latency, leadership changes)
• Should track: failover count, leadership tenure, sync lag

---

10. Comparison with Review Comments

Addressing existing review comments:

1. ✅ Cache parallelization (per julienrbrt comment): Agreed this is a good improvement
2. ✅ P2P height tracking (syncer.go:310): Your explanation makes sense - RAFT provides the source of truth
3. ⚠️Protobuf TODO (node.go:211): As mentioned, this is critical for performance - should not wait
4. ✅ isCatchingUpState naming (per julienrbrt): Consider isRaftSynced() or isClusterSynced() to be more specific
5. ⚠️Commit/Flush behavior: The discussion about SyncWrites is important - current approach is safe but needs documentation

---

Summary & Recommendations

Must Fix Before Merge

1. Implement Protobuf serialization for RAFT state (pkg/raft/node.go:211)
2. Add error handling for dropped RAFT messages with metrics (pkg/raft/node.go:296)
3. Improve leadership transfer failure handling (pkg/raft/election.go:99-102)
4. Add TLS support for RAFT transport or document security requirements
5. Fix validation gaps in RaftBlockState.assertValid() (pkg/raft/types.go)

High Priority

6. Increase test coverage for critical paths (currently 41.48% patch coverage)
7. Add integration tests for failure scenarios (network partition, crash recovery)
8. Document operational procedures (recovery, scaling, monitoring)
9. Add comprehensive health checks and metrics
10. Review and test shutdown ordering under load

Nice to Have

11. Make timeouts configurable (currently hard-coded)
12. Add performance benchmarks for RAFT operations
13. Improve error messages with contextual information
14. Add structured logging with correlation IDs
15. Consider implementing backpressure for apply channel

Architecture Validation

The overall architecture is sound:

• ✅ Clean separation between RAFT consensus and business logic
• ✅ Well-designed state machine for leadership election
• ✅ Good use of interfaces for testability
• ✅ Proper handling of sync-before-leadership to prevent double signing

However, the implementation needs refinement in error handling, testing, and performance optimization before production readiness.

---

Overall Assessment: This is a solid foundation for HA via RAFT, but needs attention to the critical issues above, particularly around error resilience, security, and test coverage before production deployment. The architectural decisions are sound, but implementation details need hardening.

[julienrbrt]

first walkthrough, I'll go read about github.com/hashicorp/raft and come back to review election.go and node.go

[julienrbrt]

Why do we need to know the height of the p2p (go header) store? (I am still reviewing, this may get clearer). We can have the app height from the evolve store.

[alpe]

When the node switches from sync to aggregator mode, the internal state is key to prevent double signing.
The Syncer has a isCatchingUpState method now that checks the stores for any height > current.
it is called within the leader election loop to transfer leadership in case it is not fully synced, yet.

[julienrbrt]

nit: what is this for? the go doc is very light

[alpe]

The channel is passed by the syncer to receive first level state updates from within the raft cluster. This should be the fastest communication channel available.

[julienrbrt]

unrelated: i wonder how this will play with different sequencers.
In #2797 you can get to that path without node key (to sign). I suppose we'll need to add a condition for based sequencing.

[alpe]

yes, I was only preparing for single sequencer. Base would not work with raft as there are no aggregators.

[julienrbrt]

not sure I understand this. is the aggregator broadcasting to no one?

[tac0turtle]

the aggregator is required to broadcast to at least one node part of a larger mesh other wise p2p will not work

[alpe]

This is more who calls whom. The aggregator gets called not otherwise. Starting all nodes with p2p-peer setup makes sense though. When a ha cluster is setup, the raft leader gets the aggregator role and I clear the peers when the p2p stack is restarted.
There is an error thrown somewhere when peers are not empty.

[julienrbrt]

nit: we should be using DefaultConfig() value if empty.

[julienrbrt]

why the todo? size?

[tac0turtle]

we should migrate to protobuf here. json will cause overhead, at 100ms we need to minimise it as much as possible

[tac0turtle]

nit: in the flow of this function we decode once and encode twice this data. i wonder if we can make it only decode. This can be a follow up to not inflate this pr.

[tac0turtle]

what is reasoning behind moving this from the composing part to the constructor?

[alpe]

There is not strong reason for this other than make it consistent with full node constructor.
With full-nodes, the p2p client is setup in the failover.go so that it can be reset when the sync-node becomes leader and peers must be empty.

[tac0turtle]

overall looks good to me. what sort of latency does the node have for the switch?

[alpe]

obsolete by #2954