Felix 0.6.0-preview.5
Released 2026-10-10 as a preview. The full history is in CHANGELOG.md.
The fifth preview of 0.6.0, mostly fixes to shard restores and replication.
A restore that could not finish no longer holds the only move slot for the
30-minute move timeout: one whose new copy stops moving for
FELIX_SHARD_RESTORE_STALL_MS (default two minutes) is dropped, and dropped
shards take turns at the slot. A restore of an empty shard is now seated, so a
cluster whose brokers registered a moment apart no longer stays on two of
three brokers. A Quorum cache’s writes no longer stall for seconds after one
follower dies. Token delegation now takes only tokens minted for the
delegating gateway: token exchange accepts an RFC 8693 actor_token, and
gateways must send it on the exchanges whose tokens they delegate
(FELIX_CONTROLPLANE_DELEGATE_UNBOUND_TOKENS=true keeps the old behaviour
during an upgrade).
Changed
Section titled “Changed”POST /v1/tenants/{tenant_id}/token/exchangetakes an optional RFC 8693actor_token(withactor_token_type): a gateway’s ownfelix-controlplanetoken holdingtoken.delegateon the tenant. The broker token it mints, and every token refreshed from it, carriesmay_act: {"sub": "<gateway>"}. The refresh record keeps the actor as a Raft metadata version 6 field, so under the Raft store an exchange withactor_tokenis a409until every member is upgraded.- An exchange whose
actor_tokendoes not verify for the tenant or lackstoken.delegateanswers403with codeactor_refusedinstead offorbidden, and counts asreason="actor_refused"infelix_controlplane_auth_rejected_total. A gateway can now tell its own token being refused from a user without permissions, which staysforbidden. POST /v1/tenants/{tenant_id}/token/delegaterefuses a subject token whosemay_actdoes not name the caller. A token with nomay_actis refused too unlessFELIX_CONTROLPLANE_DELEGATE_UNBOUND_TOKENS=true. Gateways must sendactor_tokenon the exchanges whose tokens they delegate.
- A restore of a shard whose log was empty, such as a cache used only for
counters or one nothing had written to yet, was never seated: with nothing
to ship, the leader never reported the new copy’s position. It held the
cluster’s move slot until the 30-minute move timeout, so every other restore
waited behind it. A cluster whose brokers registered a moment apart placed
its shards on two of three brokers and stayed that way, and a
Quorumcache on such a pair refused every write once either broker died (#1133). A follower of an empty log is now reported level at offset zero. - A
Quorumcache’s writes stalled for seconds after one of its followers died, even with a majority alive: the leader shipped the counter log to every follower and waited for all of them before shipping the cache log, and again at the end of each pass. The counters now ride each follower’s exchange, so a follower that does not answer holds neither mark (#1137). - One restore that could not finish held the only move slot for the whole
30-minute move timeout, so every other shard short of its replication
factor stayed short (#1138). A restore whose new copy’s reported position
has not moved for
FELIX_SHARD_RESTORE_STALL_MS(shard_restore_stall_ms, default two minutes,0waits for the move timeout) is now dropped asstalledand its slot goes to the next shard. Shards whose restores were dropped take turns at the slot, oldest attempt first, instead of the first in key order retaking it every time. - Any holder of
token.delegatecould delegate any broker token of the tenant it got hold of, a stolen one included, and present it over its own certificate. Delegation now takes only tokens minted for the caller.
Documentation
Section titled “Documentation”- The shared-connection docs said each user’s limits kept one user at a limit from blocking the others. The users share a connection ceiling of four users’ worth by default, so four at their limits can starve a fifth; the docs now say so, and a test covers the per-user publish byte budget and the ceiling over QUIC.
docs/security/rbac.mdsays whattoken.delegatelets its holder do.
