Skip to content

Felix 0.6.0-preview.6

Released 2026-10-10 as a preview. The full history is in CHANGELOG.md.

The sixth preview of 0.6.0, fixes to shard placement, broker readiness and the control plane. A cluster started all at once now places its shards on all its brokers: while its first broker registered less than FELIX_PLACEMENT_SETTLE_MS ago (10 s by default), a new shard waits for as many live brokers as its factor, and a restore that only tops up a nearly empty shard no longer takes the move slot. A shard placed while its leader was the only live broker now gets its other copies once more brokers join. A broker fetches every tenant’s signing keys before it reports ready, so after a restart it no longer refuses forwarded writes, such as cache puts, while the control plane is down. A stopped control-plane Raft member no longer claims leadership, and a second SIGTERM sent just as the control plane starts draining is no longer missed. The docs now give failover timings per case with the default settings.

  • A cluster started all at once now places its new shards on all the brokers it starts with. Placement ran as soon as the first one or two brokers had registered, so each shard got fewer copies than its replication factor, and restores added the rest one shard at a time behind the single move slot: eleven shards spent about 90 s short, and stopping a broker then could leave some without a majority. While the cluster’s first broker registered less than FELIX_PLACEMENT_SETTLE_MS ago (10 s by default), a new shard waits for as many live brokers as its factor, and a broker registering runs placement at once. A restore that only tops up a short shard whose leader holds at most FELIX_SHARD_MOVE_FENCE_MAX_LAG_RECORDS records no longer takes a move slot, so a cluster whose shards did go short fills them together. The control plane logs when shards go short of their factor and when they are full again (#1151).
  • A second SIGTERM sent to the control plane just after its readiness flipped to draining could be missed, so the pre-drain hold-off ran in full. The listener for it is now installed before readiness changes.
  • A shard placed while its leader was the only live broker now gets the copies its replication factor asks for once other brokers join. Placement adds a copy only after the leader has reported at the shard’s generation, and a leader with no followers never reported, so the shard stayed at one copy with under_replicated: true and nothing logged. A leader with no followers now reports itself, after its promotion fence as before (#1153).
  • A broker fetches the JWKS of every tenant in its catalog when its first catalog sync lands, before it reports ready. It used to fetch them only when a client of that tenant connected, so a broker that restarted and then lost the control plane refused every write a peer forwarded to it, such as a cache put sent to a broker that does not own the key, until the control plane came back (#1148).
  • A control-plane Raft member that had shut down still reported itself as leader and answered leader-only requests (heartbeats, expiry) from its last term’s view. It now reports no leadership and refuses them with the same unavailable answer an instance without a leader gives, so callers move on to a live member (#1144).
  • The failover docs said a lost leader is replaced in about a second, which is the cluster harness’s 1 s expiry window. They now give each case with the defaults: a crashed Leader-stream or cache leader is marked down 18.75 s after its last heartbeat and replaced on the next placement pass, a SIGTERM hands leaderships off as planned moves, and a lease-free Quorum stream fails over about 5 s after its leader goes quiet (#1152).