Skip to content

Raft upgrade compatibility: member rejoining, WAL page sizes and legacy HTTP headers #299

Description

@guillaume-chervet

Two blockers encountered while validating a 6.6.0 → 6.7.2 Raft upgrade

@sakno

Downstream tracking: SlimPlanet/SlimFaas#402. Tested on macOS ARM64 (.NET SDK 10.0.300, 16 KiB system pages), with three real Raft nodes and with focused tests against develop (f10ace6).

1. A removed live member cannot rejoin

Create a three-node HTTP Raft cluster, remove one live follower, append another entry, and add the follower again. Catch-up applies the removal before the node receives the re-addition. Subsequent AppendEntries returns HTTP 500 / QuorumUnreachableException, and the follower does not apply the re-addition.

MoveToStandbyState(resumable: false) faults the election task. The Leader setter subsequently accesses that completed task's Result without checking success. In addition, UnfreezeAsync short-circuits on the previously completed readiness probe, and the old leadership task remains faulted.

The existing downstream membership test passes on 6.6.0 and fails repeatedly on 6.7.2. A focused upstream HTTP cluster test reproduces the failure without downstream application code.

2. Existing WAL metadata pages are interpreted using a different size

WAL metadata pages created with 6.6.0 have a 4096-byte layout. Newer constructors use max(4096, Environment.SystemPageSize), which is 16384 on this host. After correctly restoring the snapshot, reopening the compacted WAL fails with:

WriteAheadLog.InternalException: WAL page 0 doesn't exist on the disk
  WriteAheadLog.MetadataPageManager.GetView
  WriteAheadLog.ApplyAsync
  WriteAheadLog.InitializeAsync

A rolling upgrade fails at its first follower; all 180 synthetic sets were verified on every node before upgrading. The same saved snapshot/WAL restores successfully with 6.6.0. The fixture and test are in SlimPlanet/SlimFaas#403.

The persisted page size needs to survive reopening, including logs already created with the newer larger-page layout. Invalid/mixed page sizes should fail before files are opened or resized. Private-memory buffers also need an alignment compatible with legacy pages smaller than the OS page size.

A proposed fix and tests are being prepared against develop. No causal link to the original downstream staging incident is claimed.

Additional rolling-upgrade blocker: HTTP headers

After correcting WAL restoration, the next real 6.6.0 → 6.7.2 native rolling-upgrade attempt fails because X-Raft-State-Version is required on incoming requests. Legacy nodes have no state version header (implicit version zero). Conversely, a new leader requires X-Raft-Last-Index on AppendEntries responses, which older followers do not emit. Both absent headers need legacy-compatible defaults while keeping malformed present values rejected. Three regression cases fail before this compatibility fix. Tracked with the other fixes in #300.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions