Releases · goliajp/kevy

Minor release: COW persistence — snapshot/rewrite serialization no
longer stalls a shard for the disk write (an O(n)-shallow view freeze,
~8 ns/entry, replaces it), plus an internal steel-dedup pass (one
crash-safe reshard engine shared by server and embedded), an embedded
durability fix, and real INFO persistence fields. Workspace 1.15.0 →
1.16.0; kevy-embedded 1.1.19 → 1.1.20; kevy-client 1.7.15 → 1.7.16 (dep
refs only). Perfgate PASS on every unit (6/6 angles, lx64; see
"Changed" for the gate-methodology update).

Added

Background BGSAVE / BGREWRITEAOF: the shard freezes a
copy-on-write view of its keyspace (collection values are
refcount-shared; mutations copy on write while a snapshot is in
flight) and a per-shard background thread serializes it. +OK
returns at the freeze; the snapshot/rewritten log swaps in within a
tick (~100 ms) of the disk write finishing. One job in flight per
shard (the Redis single-bgsave discipline). SAVE keeps its
synchronous, blocking-durable contract — and is skipped with a log
line if it races an in-flight background job.
INFO persistence real fields: aof_rewrite_in_progress now
reports the answering shard's actual state (it was a stubbed 0),
and the new aof_rewrites_total counts completed rewrites — the
completion signal for the now-asynchronous BGREWRITEAOF. Refreshed
per reactor tick.
kevy_store::Store::collect_snapshot / SnapshotView (embedded /
library users): an O(n)-shallow, Send point-in-time view —
serialize on any thread while the store keeps mutating.
kevy_persist serializers accept either a live store or a view
(SnapshotSource).

Changed

BGSAVE resets the AOF at the snapshot point (replacing the old
save-then-truncate): the new log carries exactly the post-snapshot
writes, teed while the background save ran. Crash exposure is
unchanged — the old log keeps receiving every write until the swap,
and the snapshot-rename + log-swap commit happens in one adjacent
critical section.
Embedded re-shard output is server-identical: a shard-layout
migration now writes per-shard dump-{i}.rdb snapshots + fresh AOFs
(previously rewritten-in-place AOFs), and is crash-idempotent via the
same reshard.journal roll-forward the server uses — a crash
mid-migration previously lost the migrated state from disk. Backup
rename failures now propagate instead of being silently ignored.
Perfgate methodology (bench/perfgate.sh): each angle now
measures 3 fresh server instances and gates on the median across
instances (was 3 rounds against one instance). Instance-to-instance
spread is the dominant noise axis (±5%); the baseline was re-recorded
accordingly. Affects contributors only.

Fixed

Embedded Store::save_snapshot no longer double-applies history on
restart: it never reset the AOF, so a restart with both files
replayed the full log on top of the snapshot — duplicating
non-idempotent commands (RPUSH'd elements doubled). It now performs
the same tee'd log reset as BGSAVE; a save that races the
background auto-rewrite waits it out (bounded) instead of writing a
snapshot whose log would still double-apply.

Internal

One crash-safe reshard engine (kevy_persist::reshard) behind both
the server and embedded migration paths; per-shard persistence file
names have a single source of truth (kevy_persist::layout); the
epoll/io_uring reactors share one cross-core drain
(drain_inbound_core); the CLUSTER topology emitters share one
derivation.

Known limitations

BGSAVE / BGREWRITEAOF completion is asynchronous: poll
INFO persistence (aof_rewrite_in_progress / aof_rewrites_total)
rather than expecting files to have swapped when +OK arrives.
A collection first mutated while a snapshot is in flight is deep-
copied at that moment (copy-on-write granularity is the whole
collection) — a write touching a very large hash/zset during a
background save pays that copy once.
Tombstone-PEL, cross-shard XREADGROUP, and cross-slot multi-key
items carried from v1.15.0 (below).