Skip to content

Felix 0.6.0-preview.9

Released 2026-10-17 as a preview. The full history is in CHANGELOG.md.

The ninth preview of 0.6.0, mostly about what a crash does to a broker’s storage identity. A process crash keeps the storage id and only a machine crash rotates it, a volume restored from before the node’s last registration rotates it, a wiped node can no longer vote for a deposed leader, and a fence counts an answer only from the copy the failover’s report named. docker compose up now starts a local stack, and the control plane no longer serves a Swagger UI.

Upgrade notes.

  • A Raft control plane writes a fence answer’s storage identity only once every member is at metadata level 16. Upgrade every member first.
  • A Postgres control plane runs migration 0034 on start.
  • A broker now writes <data dir>/registration and, after a start under a relaxed fsync mode, run-state version 2. A downgrade after a crash rotates the storage id.
  • /docs (Swagger UI) on the control plane is gone; use /v1/openapi.json.
  • compose.yaml at the repository root runs a control plane, a broker and development credentials from the released images, so docker compose up (or podman compose up) is enough to publish and subscribe with felixctl (#1345). The credentials are dev-only and every port is bound to loopback. CI does a round trip through the stack on every change.
  • On Linux, under periodic or none fsync, a broker whose process crashed (SIGKILL, an OOM kill, a panic) keeps its storage id, so its followers’ copies and its own leadership survive the restart without a re-copy (#1367). Only a machine crash rotates it: run-state now records the machine’s boot id (/proc/sys/kernel/random/boot_id), and a start in a different boot, or where either boot is unknown, takes a new id as before. Keeping the id first calls syncfs on every filesystem under the data directory and rotates if that fails. macOS and other platforms still rotate after every relaxed crash. A failed fsync forgets the boot, so a crash after one still rotates. A VM or container memory-snapshot restore keeps the boot id; use on_commit or delete storage-id for those workflows. A clean stop is written in the layout older releases read, so a downgrade after one keeps the id.
  • A release now asks getfelix.dev to show it as soon as it is published, through a release-published dispatch that needs the new SITE_DISPATCH_TOKEN secret.
  • The control plane no longer serves a Swagger UI at /docs. /v1/openapi.json is unchanged; point any OpenAPI viewer at it (docs: Viewing the API document). This removes the zip build dependency behind RUSTSEC-2026-0367 and RUSTSEC-2026-0368 and the Swagger UI download at build time.
  • A replica holding no fence promise on disk, such as one on a wiped data directory, no longer answers a fence for a shard its routing view does not have yet, or a leader other than the one the view names at that generation (#1267). A wipe took the promises it had given, so a deposed leader could count it toward a majority its successor had already fenced, and acknowledge a write the successor never held. A new shard’s first fence may now wait for a follower’s assignment watch.
  • A broker whose data directory was restored from a volume snapshot or copied in no longer keeps its storage id when the node has registered since the copy was taken, under any fsync mode (#1334). The broker records each registration’s node incarnation in <data dir>/registration, and a cluster member asks the control plane for its current incarnation before any store opens: when the control plane is ahead, it writes a new storage id, so peers and placement stop counting the copy on positions it no longer holds. A cluster member with durable storage now waits for the control plane to answer before it opens storage. A copy taken during the node’s current registration is not detected, so the backup runbook still deletes storage-id for one.
  • A promoted leader’s fence no longer counts a replica that came back under a new storage identity (#1335). Under Periodic or None fsync, a follower that crashed lost records it had already answered for. If the leader was then replaced, the fence could count that short copy toward its majority, and a record acknowledged on the old leader and that follower was lost. A failover’s assignment now carries the storage identity the replica report named for each replica (fence_storage_ids). The fence counts an answer only when it comes over a connection whose handshake gave that identity. Any other answer still takes the promise, but counts for nothing, and the fence waits for a copy that can vouch for what it holds. Under Raft the field waits for metadata level 16. A member of the old set the report could not vouch for is named with an empty identity and does not count. A replica outside the old set, or one from a build that predates identities, counts as before. With the old leader gone for good and no majority that counts, the shard waits. The way out is to stop the promoted broker and run felixctl placement abandon. That keeps every copy in the new replica set and takes the furthest of them. Under Periodic, the default, this includes a leader and a follower whose machines go down together. The follower comes back under a new identity even when nothing was lost, so the shard no longer serves when it returns. It waits for the operator instead. A process killed on Linux keeps its identity in the same boot (#1367), so its answer still counts.
  • The TLA+ model now checks wipes, follower crashes and leader crashes on the fenced, follower-acked path the broker ships, not only on the report-gated one (#1333). Three new configurations pass (FelixShardFencedWipe, FelixShardFencedCrash, FelixShardFencedLeaderCrash), and each has a twin that drops one fix and finds the lost record. FelixShardFencedCrash is the first passing check of a follower crash with a rotated storage identity.