What changed in Valkey from 8 to 9

Valkey 9.0.6 - Released Tue 01 September 2026

Upgrade urgency SECURITY: This release includes security fixes we recommend you apply as soon as possible.

Security Fixes

• GHSA-jcj7-v34w-v9vv: Fix a use-after-free in RDMA connection handling that could allow an authenticated client to crash the server using CLIENT KILL. Only affects servers built with USE_RDMA and configured with an RDMA listener (#4534)

Bug Fixes

• Fix crashes, assertion failures, and hangs when using RDMA together with IO threads by @quanyeyang (#3335)
• Fix a double free when a module timer callback stops its own timer with ValkeyModule_StopTimer by @quanyeyang (#4211)
• Fix torn RESP3 push frames when a client publishes a large message to a channel it is also subscribed to by @quanyeyang (#4253)
• RESET now clears CLIENT IMPORT-SOURCE state so reused pooled connections regain normal key expiration semantics by @tjade273 (#3973)
• Truncated AOF files now discard an incomplete MULTI block entirely, preventing loss of later writes after another restart by @chzhoo (#4342)
• Fix an ACL bypass in GEORADIUS and GEORADIUSBYMEMBER where duplicate STORE options checked only the first destination key by @tjade273 (#3971)
• Fix a use-after-free crash when a cluster message arrives for a message type registered by an unloaded module by @enjoy-binbin (#4360)
• Always deep-validate payloads on RDB load and RESTORE, preventing deferred assertion crashes; sanitize-dump-payload is now a deprecated no-op by @jjuleslasarte (#3721)
• Fix out-of-bounds memory access when registering or receiving cluster module messages of type 255, which is now a valid type by @enjoy-binbin (#4410)
• AOF loading no longer applies ACL checks, preventing silent data loss when replaying commands with a disabled default user by @lukepalmer (#3984)
• Fix a client memory accounting leak that inflated the mem_clients_normal INFO field on replicas after primary disconnects by @enjoy-binbin (#4395)
• Fix a permanent client hang when a blocking command such as BLPOP is pipelined with a partially received next command by @foobar (#4531)
• HGETEX now requires write permission on the key, so read-only ACL users can no longer change field TTLs or delete fields by @ranshid (#4576)
• Compare the full TLS certificate CN when authenticating, so an embedded NUL cannot impersonate a truncated ACL username by @madolson (#4577)
• Restore read performance with IO threads on TCP/TLS by applying extra read-completion handling only to RDMA connections by @quanyeyang (#4414)
• Restore write performance with IO threads on TCP/TLS by limiting post-write safety checks to RDMA connections by @quanyeyang (#4452)
• Fix atomic slot migration protocol errors with IO threads by not offloading export connection writes while snapshotting by @satheeshaGowda (#4104)
• Reject invalid slot import ranges when loading an RDB, preventing corrupted files from creating bad migration state by @enjoy-binbin (#4229)
• Reject RDB slot-import records with an invalid job name length, fixing an out-of-bounds read during startup by @quanyeyang (#4210)
• Fix a crash when COPY ends with a bare DB token during slot migration, and block cross-DB COPY regardless of option order by @madolson (#4301)
• HPERSIST, HTTL, HPTTL, HEXPIRETIME, and HPEXPIRETIME now return a syntax error when the FIELDS keyword is missing by @cjx-zar (#4300)
• Fix a TLS and IO threads race that could leave slot migration export jobs stuck until timeout by @jjuleslasarte (#4320)
• Fix a server crash when hash field expirations are set near the maximum timestamp, for example via HPEXPIREAT by @ranshid (#4312)
• Fix a stack overflow crash on TLS connections when retrying a failed write of large replies by @murphyjacob4 (#4307)
• Validate cluster bus PUBLISH and MODULE packet payload lengths, preventing a remote crash from forged length fields by @tjade273 (#3972)
• Fix a use-after-free crash when serving blocked clients if handling one client frees another blocked on the same key by @quanyeyang (#4212)
• Fix a server panic with IO threads when pipelined commands with a wrong number of arguments reached the key prefetcher by @madolson (#4302)
• Reject crafted stream RESTORE and RDB payloads with inconsistent lengths or negative field counts that could crash the server by @madolson (#3922)
• Reject stream payloads with mismatched live and deleted entry counts that could make XDEL destroy live entries by @roshkhatri (#4381)
• Fix CLUSTER SLOT-STATS ORDERBY returning wrong ordering once cumulative slot counters differ by more than 2^31 by @jzy1688 (#4459)
• Fix atomic slot migration failures with TLS and IO threads by not offloading export connection reads while snapshotting by @satheeshaGowda (#4559)

Full Changelog: https://github.com/valkey-io/valkey/compare/9.0.5...9.0.6

Valkey 9.1.2 - Released Mon 31 August 2026

Upgrade urgency SECURITY: This release includes security fixes we recommend you apply as soon as possible.

Security Fixes

• GHSA-jcj7-v34w-v9vv: Fix a use-after-free in RDMA connection handling that could allow an authenticated client to crash the server using CLIENT KILL. Only affects servers built with USE_RDMA and configured with an RDMA listener (#4534)
• GHSA-fq2f-crmw-q97r: Fix an unauthenticated use-after-free of the Lua interpreter state, caused by a process-global script debugger command table that cached a raw pointer to a freed interpreter and was never invalidated (#4574)

Bug Fixes

• Fix a double-free crash when a module timer callback stops its own timer with ValkeyModule_StopTimer by @quanyeyang (#4211)
• Fix torn RESP3 push frames when a client publishes to a channel it is also subscribed to, which could desync client libraries by @quanyeyang (#4253)
• Listpacks are now always validated on RDB load and RESTORE, preventing deferred assertion crashes; sanitize-dump-payload and its ACL flags become no-ops by @jjuleslasarte (#3721)
• Fix crashes, hangs, and CPU spinning when the RDMA transport is used together with I/O threads by @quanyeyang (#3611)
• RESET now clears the CLIENT IMPORT-SOURCE flag, so reused pooled connections return to normal expiration semantics by @tjade273 (#3973)
• Truncate a partially written MULTI block from the AOF on short read, preventing loss of newer writes after a later restart by @chzhoo (#4342)
• Fix an ACL bypass where duplicate STORE/STOREDIST options let GEORADIUS write or delete keys outside the user's permitted patterns by @tjade273 (#3971)
• Fix command log redaction leaking between commands in a MULTI transaction and missing for commands executed from scripts by @madolson (#4323)
• Fix a use-after-free crash when a module's cluster message type is received after the module is unloaded by @enjoy-binbin (#4360)
• Fix out-of-bounds access for cluster module message type 255, which is now a valid, dispatchable message type by @enjoy-binbin (#4410)
• AOF loading no longer performs ACL checks on replayed commands, preventing silent data loss when the default user is disabled by @lukepalmer (#3984)
• Fix a client memory accounting leak on replicas that inflated the mem_clients_normal INFO field after primary disconnections by @enjoy-binbin (#4395)
• Fix a permanent client deadlock when a blocking command like BLPOP is followed by a partially delivered pipelined command by @foobar (#4531)
• HGETEX now requires write permission on the key, closing an ACL gap that let read-only users change field TTLs or delete fields by @ranshid (#4576)
• Compare the whole TLS certificate CN during authentication, so an embedded NUL can no longer impersonate another ACL user by @madolson (#4577)
• Fix atomic slot migration failures with I/O threads by not offloading the export job's writes while snapshotting by @satheeshaGowda (#4104)
• Reject invalid slot import ranges when loading an RDB, so corrupted files can no longer create bad migration jobs by @enjoy-binbin (#4229)
• Reject RDB slot import records with an invalid job name length, preventing an out-of-bounds read at startup by @quanyeyang (#4210)
• MOVE and COPY now check ACL access to the current database, so users can no longer exfiltrate keys from an unauthorized DB by @cjx-zar (#4155)
• Fix a crash on COPY with a trailing DB option during slot migration, and block cross-DB COPY regardless of option order by @madolson (#4301)
• Fix a server panic when pipelined commands with invalid arity reach the key prefetcher with I/O threads enabled by @madolson (#4302)
• HPERSIST, HTTL, HPTTL, HEXPIRETIME, and HPEXPIRETIME now return a syntax error when the FIELDS keyword is missing by @cjx-zar (#4300)
• Fix a race between TLS I/O-thread writes and reads that could leave slot migration export jobs stuck until timeout by @jjuleslasarte (#4320)
• Fix a signed overflow that let very large hash field expiration times (e.g. via HPEXPIREAT) crash the server by @ranshid (#4312)
• Fix a frozen monotonic clock on hosts with unsynchronized TSC that stopped background tasks and key expiration by @quanyeyang (#4346)
• Fix a stack overflow crash when retrying a failed TLS write with a large reply by @murphyjacob4 (#4307)
• Fix the --check-system clocksource check to skip hosts using a hardware clock and suggest only actually available clocksources by @quanyeyang (#4272)
• Fix an assertion failure with I/O threads when a blocked client's pending command was processed again before unblocking by @quanyeyang (#4376)
• Sentinel no longer loads the built-in Lua scripting engine, removing a spurious warning at startup by @enjoy-binbin (#4327)
• Validate channel, message, and module payload lengths in cluster bus packets, preventing forged packets from crashing nodes by @tjade273 (#3972)
• Harden stream validation on RDB load and RESTORE so crafted payloads can no longer crash the server on later commands by @madolson (#3922)
• Reject stream payloads with mismatched live/deleted record counts, preventing XDEL from destroying unaccounted entries by @roshkhatri (#4381)
• Skip unnecessary post-read processing with I/O threads on socket and TLS connections, restoring small-payload throughput by @quanyeyang (#4401)
• Fix a use-after-free crash when serving clients blocked on the same key if one client is freed during processing by @quanyeyang (#4212)
• Avoid an unneeded client lookup per write completion with I/O threads on socket and TLS connections, improving pipelined throughput by @dgershko (#4440)
• Fix CLUSTER SLOT-STATS ORDERBY returning wrong ordering when slot counters differ by more than 2^31 by @jzy1688 (#4459)
• Fix slot migration failures with I/O threads and TLS by keeping the export job's ACK reads on the main thread while snapshotting by @satheeshaGowda (#4559)

Full Changelog: https://github.com/valkey-io/valkey/compare/9.1.1...9.1.2

Valkey 9.1.1 - Released Tue 21 July 2026

Upgrade urgency SECURITY: This release includes security fixes we recommend you apply as soon as possible.

Security Fixes

• CVE-2026-56684: Fix a use-after-free in TLS connection handling that could allow an authenticated client to achieve remote code execution using CLIENT KILL (#4234)
• CVE-2026-63639: Reject corrupt stream RDB files containing a shared NACK across consumers, which could allow remote code execution. Reported by @z0v3r1n and @lifip. (#4073)

Bug Fixes

• Omit the implicit alldbs ACL rule from ACL LIST, ACL SAVE and CONFIG REWRITE so older versions can parse the output by @dvkashapov (#3964)
• Improve throughput when IO threads are enabled by offloading object deallocation from the main thread by @roshkhatri (#3938)
• Fix use-after-free crash when ACL LOAD removes a user whose authenticated client has its close deferred by @ranshid (#3800)
• Enforce db= ACL permissions on every DB clause of COPY, closing a bypass with REPLACE or repeated DB tokens by @enjoy-binbin (#3801)
• Enforce database-level ACLs for CLUSTER FLUSHSLOT, which removes keys from all databases by @enjoy-binbin (#3806)
• Fix use-after-free in the module API when unregistering the first registered cluster message receiver by @eifrah-aws (#3846)
• Fix HRANDFIELD with a positive count looping forever when non-expired fields are fewer than the requested count by @cjx-zar (#4047)
• Fix clients left on the wrong database after module keyspace notifications for MOVE and COPY by @enjoy-binbin (#4024)
• Fix Sentinel crash during coordinated failover when the command link to the old primary disconnects by @lukepalmer (#4068)
• Fix crash when active hash field expiration leaves a single-entry expiry bucket whose last field is later removed by @ranshid (#3950)
• Fix assertion in HEXPIRE, HGETDEL and HPERSIST when a module blocks the client in a keyspace notification callback by @enjoy-binbin (#3743)
• Fix undefined behavior in the failover delay calculation when cluster-node-timeout is below 30 milliseconds by @enjoy-binbin (#3941)
• Reject zipmap RESTORE/RDB payloads with overflowing length fields that could cause out-of-bounds access on 32-bit builds by @madolson (#3920)
• Reject NAN scores in listpack and ziplist encoded sorted sets on RDB/RESTORE load, preventing a crash on skiplist conversion by @madolson (#3921)
• Fix corrupted replies (dropped leading bytes) caused by a reply buffer race when IO threads are enabled by @nanyan0312 (#4060)
• Fix startup crash on 32-bit systems where time_t is 64-bit (such as Alpine 3.23) when generating INFO output by @chenshi5012 (#3787)
• HGETDEL now returns a syntax error when the FIELDS keyword is missing or misplaced by @lcxn123 (#4049)
• COMMAND INFO in RESP3 now returns the subcommands field as an array instead of a set for commands without subcommands by @rickrams (#3939)
• Send the replica version on the dual-channel RDB connection so full syncs with newer encodings like hash field TTLs succeed by @hpatro (#4105)
• Fix duplicate failure handling and an invalid reply sequence in cluster slot migration by @chx9 (#3723)
• Reject control characters in SENTINEL SET values to prevent config-file injection via Sentinel config rewrite by @eifrah-aws (#3847)
• Reject control characters and delimiters in cluster AUX fields and validate cluster-announce-ip to prevent nodes.conf injection by @eifrah-aws (#3848)
• Redact key names and user data from more server log messages when hide-user-data-from-log is enabled by @zackcam (#3872)
• ACL LOG now reports the denied database ID for COPY instead of the command name when db= access is denied by @enjoy-binbin (#3888)
• Fix garbled shard IDs in the cluster UPDATE message log line by @enjoy-binbin (#3942)
• Fix negative master_sync_total_bytes in INFO replication during disk-based sync when the RDB exceeds 2GB by @chx9 (#3811)
• Increase the maximum process title length from 255 to 1024 characters to avoid truncation with long paths by @pkhartsk (#3843)

Full Changelog: https://github.com/valkey-io/valkey/compare/9.1.0...9.1.1

Valkey 9.0.5 - Released Tue 21 July 2026

Upgrade urgency SECURITY: This release includes security fixes we recommend you apply as soon as possible.

Security Fixes

• CVE-2026-56684: Fix a use-after-free in TLS connection handling that could allow an authenticated client to achieve remote code execution using CLIENT KILL (#4234)
• CVE-2026-63639: Reject corrupt stream RDB files containing a shared NACK across consumers, which could allow remote code execution. Reported by @z0v3r1n and @lifip. (#4073)

Bug Fixes

• Strictly validate CRLF terminators when parsing the RESP protocol; malformed requests now get a protocol error instead of being misparsed by @enjoy-binbin (#2872)
• Fix a use-after-free crash when creating slot import jobs during manual slot migrations by @twooster (#3283)
• Fix a memory leak in ZDIFF and ZDIFFSTORE when the result set becomes empty before all inputs are processed by @sarthakaggarwal97 (#3342)
• Fix HPERSIST sending a malformed reply that desynchronized the connection when used on a key of the wrong type by @madolson (#3516)
• Fix a crash from a race between IO threads and asynchronous client freeing by @deepakrn (#3458)
• Fix a double free when loading a stream with corrupt consumer PEL data from RDB or RESTORE by @enjoy-binbin (#3498)
• Fix listpack corruption and a subsequent crash when XTRIM marks the last entry of a stream listpack node as deleted by @smkher (#3591)
• Fix malformed replies when module callbacks build deferred-length arrays while a client's deferred reply buffer is active by @eifrah-aws (#3578)
• Fix a NULL pointer crash in TLS pending-data handling by @zuiderkwast (#3641)
• Fix a server crash when multiple RDMA clients disconnect at the same time by @quanyeyang (#3448)
• Fix a use-after-free when ACL LOAD deletes a user whose clients cannot be freed immediately by @ranshid (#3800)
• Fix a use-after-free when a module unregisters the first registered cluster message receiver for a message type by @eifrah-aws (#3846)
• Fix HRANDFIELD looping forever when a hash has fewer non-expired fields than the requested count by @cjx-zar (#4047)
• Fix clients being left on the wrong database after module keyspace notifications for commands like MOVE and COPY by @enjoy-binbin (#4024)
• Fix a Sentinel crash during coordinated failover when the connection to the old primary is disconnected by @lukepalmer (#4068)
• Fix underestimation of client output buffer memory when replies reference shared objects, so buffer limits are enforced correctly by @dvkashapov (#3306)
• Fix a crash on ARM/aarch64 caused by memory-ordering races in the IO thread job queue by @jjuleslasarte (#3878)
• Fix a crash when active hash field expiration leaves a single entry in a large expiration time-bucket by @ranshid (#3950)
• Fix a file descriptor leak when a blocking connection attempt, such as MIGRATE to an unreachable host, times out by @madolson (#3541)
• Fix a potential crash from a dangling slot migration job reference when the migration client is reset by @murphyjacob4 (#3554)
• Remove cached EVAL scripts when their scripting engine is unregistered, preventing dangling engine references by @eifrah-aws (#3503)
• Fix a memory leak in GEOSEARCH BYPOLYGON when argument parsing fails, such as on an invalid COUNT by @bandalgomsu (#3568)
• Fix a crash when a slot migration target node is removed from the cluster before the migration connects by @chenshi5012 (#3596)
• Fix a crash when the module GetLRU/SetLRU/GetLFU/SetLFU APIs are called with a NULL key by @yaronsananes (#3610)
• Fix an assertion failure in hash field expiration commands when a module blocks the client in a keyspace notification by @enjoy-binbin (#3743)
• Fix a cluster UPDATE log message reading shard IDs past their fixed-length buffer by @enjoy-binbin (#3942)
• Fix undefined behavior in the failover delay calculation when cluster-node-timeout is set below 30 milliseconds by @enjoy-binbin (#3941)
• Reject zipmap RESTORE payloads with overflowing length fields that could cause out-of-bounds access on 32-bit builds by @madolson (#3920)
• Reject NAN scores in listpack- and ziplist-encoded sorted sets on RDB/RESTORE load, preventing a later crash on skiplist conversion by @madolson (#3921)
• Fix a startup crash on 32-bit systems with 64-bit time_t, such as Alpine 3.23, caused by time value formatting mismatches by @chenshi5012 (#3787)
• Fix corrupted client replies when IO threads are enabled, caused by a race between in-flight writes and reply buffer reuse by @nanyan0312 (#4060)
• COMMAND INFO in RESP3 now returns the subcommands field as an Array instead of a Set for commands without subcommands by @rickrams (#3939)
• The dual-channel replication RDB connection now announces the configured replica-announce-ip, avoiding stale replica entries behind NAT by @jdheyburn (#2846)
• Prevent replicas from processing stale cluster packets and incorrectly promoting themselves to an empty primary within a shard by @zhijun42 (#2811)
• Send the replica version on the dual-channel RDB connection so full syncs of data like hash field TTLs no longer fail by @hpatro (#4105)
• Fix slot migration failure handling running twice on ownership changes and an out-of-order error reply in the internal SYNCSLOTS FINISH command by @chx9 (#3723)
• Allow slot-migration-max-failover-repl-bytes to be set to -1 to disable the limit, as documented by @enjoy-binbin (#3443)
• Fix CONFIG REWRITE producing negative values for memory configs such as maxmemory when set to very large values by @enjoy-binbin (#3440)
• Reject SENTINEL SET values containing control characters and safely quote Sentinel config values to prevent config file injection by @eifrah-aws (#3847)
• Reject control characters and delimiters in cluster AUX fields and validate cluster-announce-ip to prevent nodes.conf corruption or injection by @eifrah-aws (#3848)
• Fix changes to lua-enable-insecure-api via CONFIG SET not taking effect when the option was set at startup by @enjoy-binbin (#4182)
• Fix incorrect memory overhead reported for watched keys in client memory usage tracking by @enjoy-binbin (#3359)
• Replica logs now report 'Connection reset by peer' instead of the misleading 'Success' when the primary closes the connection by @abmathur-ie (#3580)
• Redact key names and user data from more log messages when hide-user-data-from-log is enabled by @zackcam (#3872)
• Fix INFO replication reporting negative sync transfer sizes when the RDB exceeds 2GB during disk-based sync by @chx9 (#3811)
• Increase the maximum process title length from 255 to 1024 characters to avoid truncation with long installation paths by @pkhartsk (#3843)
• valkey-cli --cluster del-node can now remove unreachable or failed nodes instead of failing with 'No such node ID' by @yang-z-o (#3209)
• Fix valkey-cli crashing after --eval script execution on jemalloc/tcmalloc builds by @bandalgomsu (#3281)
• valkey-cli --cluster fix now spreads uncovered slots randomly across primaries instead of assigning them all to one node by @abmathur-ie (#3586)

Full Changelog: https://github.com/valkey-io/valkey/compare/9.0.4...9.0.5

Upgrade urgency LOW: This is the first stable release of Valkey 9.1.

Security fixes

• (CVE-2026-23479) Use-After-Free in unblock client flow
• (CVE-2026-25243) Invalid Memory Access in RESTORE command
• (CVE-2026-23631) Use-after-free when full sync occurs during a yielding Lua/function execution

New Features and enhanced behavior

• Add cluster bus network traffic usage metric in bytes by @hpatro (#3396)
• Reduce latency spikes during rehashing via incremental page release by @chzhoo (#3481)

Bug Fixes

• Fix(syncio): Set errno on EOF in syncRead and propagate to conn->last by @abmathur-ie (#3580)
• Fix GEOSEARCH BYPOLYGON leak on invalid COUNT by @bandalgomsu (#3568)
• Handle NULL pointer in streamTrim listpack delta calculation by @smkher (#3591)
• Fixes server crash when RDMA benchmark clients disconnect by @quanyeyang (#3448)
• Fix the memory leak in valkey-benchmark by @nmvk (#3643)

See also the release notes for 9.1.0-rc1 and 9.1.0-rc2.

Upgrade urgency SECURITY: This release includes security fixes we recommend you apply as soon as possible.

Security fixes

• (CVE-2026-23479) Use-After-Free in unblock client flow
• (CVE-2026-25243) Invalid Memory Access in RESTORE command
• (CVE-2026-23631) Use-after-free when full sync occurs during a yielding Lua/function execution