github redis/go-redis v9.24.0-beta.1
9.24.0-beta.1

pre-release33 minutes ago

This is a beta release. The release adds:

  • The experimental MultiDB client for Active-Active (Redis Enterprise) failover.
  • Metadata-driven cluster routing and client-side caching (CSC) eligibility that comes from the COMMAND metadata.
  • CSC OPT-IN / OPT-OUT tracking modes, CSC on ClusterClient, and CSC with EphemeralConn.
  • A per-client OTelRecorder, and MultiDB metrics in redisotel-native.
  • New full-duplex options (FullDuplexPipes, PipelineMinIdleConns) and a cluster recovery fix for full-duplex sessions.
  • Commands for Redis 8.10 and for the 8.12 release candidates (BACKUP, set member expiration, BITRANGEINTARRAY, ZMOVE, CF.COMPACT, BLESS, scalar-quantized HNSW).

🚀 Highlights

MultiDB client (experimental)

NewMultiDBClient routes commands to one active database out of several standalone or cluster members (e.g. Redis Enterprise Active-Active) and fails over based on circuit breakers, health checks and a weight-based failover strategy, with auto-fallback to higher-weight members. The multidb package adds ping and lag-aware health checks and probe policies. See example/multidb (#3954) by @ndyakov. Defaults:

  • MultiDB owns retries: CommandRetries (0 = 2, -1 = none) retries on the database active at each attempt, and every attempt is counted by the breaker and failure detector. Standalone members get no retries of their own (MaxRetries 0 becomes -1); an explicit MaxRetries is kept.
  • Health checks: HEALTHY_ALL, 3 probes 500ms apart per check, honoring multidb.WithProbes / WithDelay / WithTimeout; HealthCheckTimeout (3s) caps the pass. Probes that cannot fit it are rejected at construction and by AddDatabase. Under the default policy, SetActiveDatabase and ProbeTargetBeforeFailover use one probe per check.
  • Startup: InitialDBStateMajorityAvailable needs half the members, rounded up (1 of 2). A failed start matches ErrInsufficientHealthyDatabases; with no healthy member it also matches ErrPermanentlyNotAvailable.
  • Weights are in (0, 1] (0 = 1.0); -1 disables a member. The last enabled member cannot be disabled.
  • The default failure detector (1000 failures and 10% within 2s) is set with FailureDetectorConfig, whose fields are used as given; start from DefaultMultiDBFailureDetectorConfig().
  • redisotel-native reports MultiDB events (failover, active database change, circuit breaker state change, health check) in a new multidb metric group, MetricGroupFlagMultiDB, which MetricGroupAll includes (#4007) by @ndyakov.
  • A member ProcessHook that returns nil without calling next is no longer counted as a healthy database response (#4271) by @harshitdeora. A local result of a member hook in a cluster pipeline is kept when another node fails (#4279) by @harshitdeora. Full-duplex outcomes are recorded in the failure window after the active member is selected again (#4283) by @xixi-box.

Experimental: the MultiDB API can change in a minor release. User hooks and health checks must not call MultiDB control methods or Close.

Metadata-driven cluster routing

Cluster routing and CSC use the same resolved COMMAND record for each command, so flags, key positions and policies cannot drift between two metadata sources (#3960, #4093) by @ofekshenawa.

  • Cluster routing shares command metadata with client-side caching for key extraction and read-only status. Existing default routing and cluster-wide DBSize, ScriptLoad, ScriptFlush, and ScriptExists helpers are preserved. Install NewCommandMetadataResolver() or client.NewDynamicResolver() with SetCommandInfoResolver to enable metadata-based fan-out and aggregation. CommandMetadata configuration alone does not enable them. Cross-slot fan-out is not atomic.
  • With metadata routing enabled, cross-slot MGET follows redirects for each affected slot during resharding instead of returning a missing value for a key that moved. Successful slot groups are not replayed.
  • Explicit pipelines and transactions retain node-local command execution; ordered AsyncAutoPipeline rejects fan-out before dispatch. Same-slot metadata commands remain batchable. Automatic fan-out is available through the blocking face or unordered half-duplex mode. Existing direct helpers and special routing retain their separate execution path. SCAN, INFO, and known diagnostics remain node-local. Single-slot raw commands and binary marshaler values are accepted when no fan-out is needed.
  • Numeric sum/min/max fan-out preserves integer replies as exact int64 results, including raw Do results; integer overflow returns an error. Aggregates containing floating-point replies return float64.
  • Resolved command key specifications take precedence over SetFirstKeyPos. Missing snapshot commands retain the legacy key-position fallback; configure CommandMetadataPreferLive or CommandMetadataConfig.Overrides for full metadata. The default CommandMetadataStatic mode performs no network lookup unless resolution reaches an explicitly installed NewDynamicResolver. After activation, reconnects, topology changes and a configured refresh interval also verify server identity in Static mode.
  • Embedded command wrappers retain raw-reply safety checks and receive typed aggregate results through their promoted SetVal methods.
  • Metadata identity probes have bounded internal deadlines, including when application socket timeouts are disabled. Waiting callers can cancel independently of an ongoing refresh. Dynamic resolution waits for one initial attempt; after failure or invalidation, commands use conservative metadata immediately while background verification retries with backoff.
  • Cluster routing decisions are stored on commands without using commands as map keys, and concurrent topology reloads share one in-flight load.
  • Topology loading prefers CLUSTER SHARDS and falls back to CLUSTER SLOTS for unsupported commands or unusable replies, including missing or unknown node health. Connection failures try another node without repeating SLOTS on the failed connection.
  • Keyless routing prefers healthy nodes, but temporary local failure/loading markers no longer exclude every candidate. Nodes explicitly marked offline by topology remain excluded. ErrClusterTopologyUnhealthy supports errors.Is checks. Unknown response policies return an error.
  • Fan-out responses containing MOVED or ASK refresh topology for later commands without replaying already successful operations.
  • FT.CURSOR READ is no longer automatically retried after a lost reply, on standalone and cluster clients, because the server may already have advanced the cursor. Cursor ownership handling is not part of this release.
  • A SCAN iterator on a ClusterClient now visits every master. Before, it sent each page to a random node with the cursor of another node and could end early without an error (#4200) by @ndyakov.
  • With the dynamic resolver, DEL, EXISTS, TOUCH and UNLINK on several keys of one slot returned the per-slot count multiplied by the number of keys in that slot. They now count each slot once (#4141) by @lazerg.

Client-side caching

Experimental: the CSC options can change in a minor release.

  • OPT-IN and OPT-OUT tracking modes: Options.ClientSideCacheTrackingMode selects TrackingModePlain (default), TrackingModeOptIn or TrackingModeOptOut. Options.ClientSideCacheable is a predicate that selects the eligible unmarked calls, and WithCache(ctx) / WithoutCache(ctx) mark one call. The mode works with miss coalescing, refresh on invalidate, pipelines, autopipeline and cluster fan-out. With CSC on, the native full-duplex engine leases only the untracked pipeline pool and does not use the main pool (#4193) by @ofekshenawa.
  • CSC on ClusterClient: each node client has its own cache, shared by its connection pool, with per-node capacity limits. ClusterOptions, UniversalOptions and NewFailoverClusterClient accept the CSC options, and CSCStats() and CSCRefreshStats() add up the registered nodes. The client does not store a reply that cannot prove that every key it depends on exists. A MOVED or ASK redirect bypasses the cache, and a topology change flushes only the caches of nodes that enter or leave routing. Replica reads, ordinary pipelines, transactions and full-duplex batches stay uncached (#4080) by @ofekshenawa.
  • Ephemeral connections with client-side caching: with CSC active, EphemeralConn now lazily dials a fresh, untracked connection. Session commands are allowed without disrupting the parent cache, and reads always go to Redis. The connection takes no PoolSize turn, does not support maintenance handoff, and must be closed after use. Without CSC, idle-connection reuse is unchanged (#4174) by @aditya-rai-5.

Full-duplex autopipelining

  • AutoPipelineOptions.FullDuplexPipes sets the number of full-duplex pipes (held connections) on a standalone Client. 0 means 1. NumShards > 1 under FullDuplex still selects that number when FullDuplexPipes is 0, but FullDuplexPipes is the preferred name. Config().FullDuplexPipes reports the effective count. On a cluster, a full-duplex autopipeliner keeps one pipe for each master (#4107) by @ndyakov.
  • Warm spare connection: Options.PipelineMinIdleConns (default 0, pipeline_min_idle_conns in URLs) sets the minimum idle connections of the dedicated pipeline pool, and is available on ClusterOptions, RingOptions, FailoverOptions and UniversalOptions. A full-duplex engine keeps one idle spare, so a replacement for a failed held connection does not wait for a TCP or TLS dial. The spare is dialed but not initialized: HELLO and AUTH still run on the first lease (#4101) by @ndyakov.
  • Cluster recovery: when the session of a node engine drops, the engine used to replay the unacknowledged commands only on a new connection to the same node, and failed them if that node was closed or replaced. The commands now go through the cluster router, so they reach the new owner of the slot (#4111) by @ndyakov.
  • OTel metrics no longer cost throughput: with redisotel-native on, full duplex lost about 90% of its throughput. In one test with 256 callers and 64 commands in flight on a local Redis, GET throughput went from about 190k/s on v9.23.0 to about 2.1M/s now (about 10x; about 2.0M/s with OTel off). The gain depends on RTT, the number of commands in flight and the CPU, so expect a smaller effect on a remote Redis. The full-duplex reader reads its goroutine id once for each session, and RecordOperationDuration uses a bounded attribute cache and a queue for the success path (#4269) by @ndyakov.

Per-client OTelRecorder

Each client can have its own metrics recorder. Set OTelRecorder in Options, ClusterOptions, RingOptions, FailoverOptions, UniversalOptions or MultiDBOptions. The client sends its commands, pipelines, errors, cluster redirects, Pub/Sub messages, stream lag, and pool and connection events to it. Cluster nodes, ring shards, sentinel clients and MultiDB members use the recorder of their parent. Clients without it use the recorder from redis.SetOTelRecorder, as before. In redisotel-native, NewRecorder makes a recorder with its own WithCustomAttributes, which take precedence over the attributes of the global Init config (#4179) by @vladvildanov.

inst := redisotel.GetObservabilityInstance()
_ = inst.Init(redisotel.NewConfig().WithEnabled(true).WithMeterProvider(provider))
defer inst.Shutdown()

acmeRec, _ := inst.NewRecorder(redisotel.NewConfig().
	WithCustomAttributes(attribute.String("tenant", "acme")))
acme := redis.NewClient(&redis.Options{Addr: "acme:6379", OTelRecorder: acmeRec})

Redis 8.10 and 8.12 commands

The 8.12 commands target the 8.12 release candidates (the CI image is 8.12-rc3). The commands and options can change before 8.12 is released.

  • Set member expiration: SExpire, SPExpire, SExpireAt, SPExpireAt (each with WithArgs), STTL, SPTTL, SExpireTime, SPExpireTime, SPersist, SAddEX and SAddEXWithArgs (#4165) by @ndyakov.
  • BITRANGEINTARRAY (#4242) by @ndyakov.
  • ZMOVE, BZMOVE and ZMOVEMEMBER (#4289) by @ndyakov.
  • CF.COMPACT as CFCompact (#4243) by @ndyakov.
  • CELL_SIZE for CMS.INITBYDIM and CMS.INITBYPROB (CMSInitByDimCellSize, CMSInitByProbCellSize) (#4023) by @vlady-kotsev.
  • Scalar-quantized HNSW indexes in FT.CREATE, and the two new FT.INFO fields (#4077) by @vlady-kotsev.
  • BLESS GET, BLESS SET, BLESS CLEAR and BLESS SCAN for the NO-EVICT flag (#4041) by @vlady-kotsev. BlessScan is node-local: scan each master of a cluster or ring.
  • BACKUP START, SEAL, ABORT, CLEANUP and STATUS for Redis 8.10 (#4286) by @Purv-Kabaria.
  • Alignment with 8.12-rc2: RERANK for FT.HYBRID, the BITRANGEINTARRAY options, and new command metadata (#4277) by @ndyakov.

⚠️ Behavior changes

Check these when you upgrade from 9.23.0:

  • ContextDeadlineError for socket timeouts that come from a context deadline (#4263) by @aditya-rai-5. With ContextTimeoutEnabled: true, a socket timeout caused by the deadline of the caller's context now matches context.DeadlineExceeded with errors.Is, and still matches net.Error with Timeout() == true. The complete original transport error stays in the chain. The error type is internal to the pool, so use errors.Is. Such an error is no longer retried, because the context has already expired, and its message starts with context deadline exceeded:. Earlier socket timeouts and hard read deadlines are unchanged.
  • WithTimeout keeps hooks: a client returned by WithTimeout now runs the hooks of its parent. Before, it dropped them, so hooks added with AddHook did not run for commands, pipelines and transactions on that client (#4095) by @CaliLuke.
  • CSC eligibility comes from the COMMAND metadata and fails closed (#3960) by @ofekshenawa. An unknown, malformed, blocking, nondeterministic or script-based command runs as usual but is not cached. JSON.MGET and TS.INFO are no longer cached by default. With CSC enabled, more eligible commands can be cached in the default Static metadata mode (the BF.*, CF.*, TOPK.*, TDIGEST.*, V* and AR* families, CMS.QUERY, EXPIRETIME, PEXPIRETIME, HEXPIRETIME, HPEXPIRETIME, PFCOUNT, SDIFFCARD, SUNIONCARD, ZINTERCARD and others). Command() results now also contain subcommand entries named parent|child, for example client|id. When built-in CSC is on, a raw CLIENT CACHING command is rejected, like the other commands that change connection state (#4193) by @ofekshenawa.
  • Connection-pool limits: MaxActiveConns now counts all connections registered with the pool (pooled, excess, and dedicated NewConn connections) plus in-flight dial reservations. Previously, admission checked only the pooled count (poolSize). Clients with this limit configured can now receive ErrPoolExhausted from NewConn/EphemeralConn or during bursts that previously exceeded the limit (#4174) by @aditya-rai-5.
  • Dial concurrency and prewarming: NewConn now waits for the shared MaxConcurrentDials limit and honors context cancellation while waiting. Idle prewarming shares this limit without holding a pool turn while waiting for dial capacity. Refill checks resume after foreground dials and successful prewarm dials complete, and concurrent refill requests are preserved. These pool-limit and prewarming changes apply to all clients using the pool, not only CSC-enabled ephemeral sessions (#4174).
  • Struct scan on a bad field: Scan on HGETALL/HMGET results used to stop at the first field that failed to decode, and with HGETALL the set of filled fields was random. It now sets every field it can decode and still returns an error. When one field fails, that field's error is returned as before. When several fail, their errors are joined with errors.Join, in key order for HGETALL and argument order for HMGET, so errors.Is and errors.As match each of them. The underlying decoder error stays in the chain, so errors.Is(err, strconv.ErrSyntax) and errors.As with *strconv.NumError work. For the built-in types, a field that fails to decode keeps its previous value, and a nil pointer field stays nil. A custom decoder (ScanRedis, UnmarshalText or UnmarshalBinary) runs on the existing field and can leave a failed field changed, so do not rely on unchanged values after an error from one. Code that reuses the struct after an error will see more fields set (#4136 by @lazerg, #2255 by @cpoole, #4294 by @Avv123).
  • URL configuration validation: ParseURL, ParseClusterURL, and ParseFailoverURL now reject negative dialer_retry_timeout, read_buffer_size, and write_buffer_size values. Cluster and failover URLs that previously accepted negative retry timeouts now return an error; replace those values with 0 for no delay, or omit the parameter for the default 100 ms delay. An explicit dialer_retry_timeout=0 now disables the retry delay (#4144) by @Samuel-Sorial. ParseURL and ParseFailoverURL now also reject a negative database number (#4275) by @ump45nose, and a plain-second duration that overflows int64 nanoseconds (#4276) by @ump45nose. A disabled dial_timeout (0 or a negative value in the URL) now means no dial deadline. Before, every dial failed at once (#4168) by @PerryLink.
  • Empty password with a username (opt-in): a Username with an empty Password sends no AUTH, the same as 9.23.0. To authenticate an ACL user created with nopass, set AuthEmptyPassword: true (or auth_empty_password=true in the URL); for sentinel connections use SentinelAuthEmptyPassword (sentinel_auth_empty_password=true). The rule also applies to credentials from CredentialsProvider, CredentialsProviderContext and StreamingCredentialsProvider used when a connection is set up; re-authentication on a credentials refresh is unchanged. Redis Enterprise and Redis Cloud databases with unauthenticated access can reject an empty password, which is why this is opt-in. The client-side cache keeps a separate namespace for a username that sends no AUTH (#4161 by @saddamr3e, #4194 by @ndyakov).
  • Ring telemetry: redisotel.InstrumentTracing and InstrumentMetrics now instrument existing shards. Remove any per-shard instrumentation workaround when instrumenting the ring to avoid duplicate spans and metrics (#4227) by @nk-batman21.
  • Full-duplex clients keep one idle spare connection: a full-duplex engine raises MinIdleConns of its pipeline pool to 1, so the pool holds one extra idle connection for each client (for each node on a cluster). Set PipelineMinIdleConns to a negative value to opt out (#4101) by @ndyakov.
  • Replies nested deeper than 1000 levels fail: the RESP reader returns an error for a reply with more than 1000 levels of nested aggregates or attributes, instead of exhausting the stack (#4296) by @xixi-box.
  • Script.Load keeps the client-computed SHA-1: a Script made with NewScript no longer takes the digest from the SCRIPT LOAD reply. A server that replies with another digest cannot redirect later Run/RunRO calls. NewScriptServerSHA still uses the server digest (#3943) by @saddamr3e.
  • PubSub.Channel misuse does not panic: calling Channel and ChannelWithSubscriptions on one PubSub now returns a closed channel and logs a warning. Ring.Subscribe, PSubscribe and SSubscribe with no channels, or a failed shard lookup, return a PubSub that carries the error and do not panic (#3968) by @VedantMadane.
  • Return values: XInfoStreamFull now reports a -1 consumer-group lag when Redis replies with a nil lag, like XInfoGroups (#4217) by @LindseyZ1205. ClusterClient.ScriptFlush now returns OK (#4252) by @LindseyZ1205.
  • New interface methods: Cmdable and its sub-interfaces gain methods for the new commands (and BlessCmdable). This is the usual way go-redis adds commands. A custom type that implements these interfaces (for example a mock) must add the methods.

✨ New Features

  • FT.AGGREGATE steps: FILTER and LIMIT are now steps in FTAggregateOptions.Steps, so you can filter before GROUPBY, filter more than once, and place LIMIT between steps. AggregateBuilder gets Limit(). The old tail fields Filter and LimitOffset/Limit still work, are sent after all steps, and are deprecated in favor of steps (#4076) by @ndyakov
  • Per-stream IDs for XRead: XReadArgs.IDs gives one ID for each stream, in the order of Streams, and takes precedence over ID (#4203) by @aozmekik
  • Native iterators for ScanIterator: ScanIterator.Vals(ctx) returns an iter.Seq[string] for use with range (#3348) by @khasanovbi
  • More URL options: ParseURL, ParseClusterURL and ParseFailoverURL accept dialer_retries, dialer_retry_timeout, context_timeout_enabled, read_buffer_size, write_buffer_size, disable_identity and identity_suffix where the options struct has the field (#4063) by @SulimanAbdulrazzaq
  • Context-aware credential subscriptions: the optional auth.ContextStreamingCredentialsProvider interface lets the client call SubscribeWithContext when it sets up a connection, so the context bounds the wait for the first credentials. Providers without the method keep Subscribe (#4249) by @PragalvaXFREZ
  • Per-node TLS verification for cluster URLs: ParseClusterURL accepts tls_verify_node_host=true on rediss:// URLs. The parameter clears TLSConfig.ServerName, so each node is verified against the host it is dialed at and not against the host of the first URL. Use it when each node has a certificate for its own host. The default is unchanged: every node is verified against the host of the first URL, so clusters that use one certificate and announce their nodes by IP continue to work. rediss:// cluster URLs now also set MinVersion: tls.VersionTLS12, as ParseURL and ParseFailoverURL do (#4145) by @PragalvaXFREZ
  • skip_verify in ParseClusterURL: the cluster parser accepts ?skip_verify=true on rediss:// URLs, like ParseURL and ParseFailoverURL (#4106) by @PragalvaXFREZ
  • TSCreateRuleOptions.AlignTimestamp: the field is now exported, so callers outside the package can set the bucket alignment for TSCreateRuleWithArgs (#4280) by @pcbeingused333

🐛 Bug Fixes

Connections, pool and transactions

  • Connection handshake and Options.Limiter: the handshake of a new connection (HELLO, AUTH, SELECT, CLIENT SETINFO, CLIENT MAINT_NOTIFICATIONS) no longer calls Limiter.Allow per command. The command that dials pays one admission and reports the handshake result once; that admission also covers the commands an OnConnect callback sends on the connection. A PubSub connection, a reconnect after a maintenance handoff, and a fresh full-duplex session each pay one admission for their handshake. Before, a circuit breaker in the half-open state spent its probe budget on the handshake and never closed (#4232) by @othoathydrahost. A panic in an OnConnect or credentials hook now settles the limiter admission, so a half-open breaker does not stay stuck (#4247) by @ump45nose
  • Silent reply mix-up after a maintenance handoff: with maintenance notifications on, a MOVING handoff under concurrent traffic could give one connection two owners. Both wrote on the same socket and read each other's replies, so a GET could return the "OK" of a SET, with no error. putConn no longer releases a connection that a hook handed to a handoff, the init of a handoff restores the owner's state, a delayed MOVING claims the idle connection before it queues the handoff, and a later handoff mark is not lost (#4110, #4239) by @ndyakov
  • Free re-run of retired idle connections: when the server closes idle TLS connections (for example with a timeout setting), the pre-command drain retires them, and each one used to cost a retry of MaxRetries. The command was never written, so the client now runs it again without spending retries or backoff. The pool size caps these re-runs (#4197) by @dunglas
  • Transactions: TxPipelined returns an error and drops the connection when EXEC replies with a different number of results than the number of queued commands. This happens when Redis rejects WATCH or MULTI inside MULTI (#4186) by @saddamr3e. A rejected MULTI drops the connection, so queued commands do not leak to another borrower (#4192) by @saddamr3e. Tx.Close requires an OK reply to UNWATCH, and removes the connection when the client cannot confirm a clean state (a refused UNWATCH, or an EXEC failure with an unknown outcome). Before, such a connection could return to the pool with WATCH or MULTI still active (#4316) by @xixi-box
  • Pipeline setup errors: when connection setup fails with a Redis error reply (for example WRONGPASS from AUTH), each queued command now gets the error. Before, Pipeline.Exec returned the error but every command had Err() == nil and a zero value (#4312) by @xh123637
  • Idle connection at MaxActiveConns: when an idle scan skipped an unusable connection and acquired a usable one on its last allowed attempt, the connection was removed from the idle list and never returned, so Get could fail with ErrPoolExhausted (#4322) by @xixi-box
  • TLS dial and context: a TLS dial now honors the caller's context deadline and cancellation during the handshake, and not only DialTimeout (#4143) by @Samuel-Sorial
  • Late PubSub connection: a PubSub connection that finished its setup after Client.Close drained the pool is now closed and returns ErrClosed, on standalone, Sentinel and cluster clients (#4157) by @aditya-rai-5
  • Maintenance notifications: a subscribed PubSub connection that gets a MOVING with a null endpoint stays under the control of the PubSub. Before, a handoff worker moved it to a new socket without SUBSCRIBE, and Channel() stopped delivering messages (#4246) by @ump45nose

Authentication and credentials

  • Pub/Sub re-authentication with StreamingCredentialsProvider on RESP3: credential rotations now re-authenticate established Pub/Sub connections in place. Before, long-lived subscriber connections kept the old credentials (#4195) by @AdamMagued. Re-authentication is serialized for each connection, so an older credential cannot be written last (#4317) by @Nacl-HCl, and it skips a connection that was replaced after the rotation (#4309) by @ump45nose. A close during subscription setup no longer re-enters initConn and panics (#4262) by @PragalvaXFREZ

  • Rejected re-authentication: the connection is now closed when the server rejects a re-auth with WRONGPASS or NOPERM. Before, it kept serving commands with the replaced credentials (#4058) by @saddamr3e

  • Pub/Sub connection leak: a closed Pub/Sub connection no longer stays in memory, with its buffers, when a StreamingCredentialsProvider is set (#4119) by @meruiden

  • Conn, EphemeralConn and Watch with StreamingCredentialsProvider: the derived clients now carry the credentials manager, so a cold connection no longer panics (#4155) by @TastyHeadphones

  • Failover client with StreamingCredentialsProvider: NewFailoverClient now creates the streaming credentials manager, so new connections no longer panic in initConn (#4182) by @lazerg

  • CSC cache and TLS client certificates: clients that authenticate with different TLS client certificates (tls-auth-clients-user) no longer share cache entries, so a restricted user cannot read entries that a privileged user fetched (#4261) by @saddamr3e
    Cluster, Sentinel and Ring

  • Redirects without a host: with cluster-preferred-endpoint-type unknown-endpoint, a null host in CLUSTER SLOTS is accepted and takes the host of the node that was asked (#4131) by @spicysharkco, and MOVED/ASK replies without a host reuse the host of the node that sent them (#4205) by @soujasK. After a maintenance handoff, the client uses the host of the serving connection (#4233) by @ump45nose

  • MOVED to the node's own address: the client drops the connection when MOVED points to the node that sent it, so the next attempt reconnects (#4229) by @SergeiSadov

  • Batch retry after a redirect: when the first reply of a node batch is retryable (for example TRYAGAIN) and a later command gets MOVED or ASK, the redirected command is no longer sent twice. A non-idempotent command such as INCR could run two times (#4258) by @lazerg

  • NewFailoverClusterClient and ConnMaxLifetimeJitter: the jitter and the failing timeout are now kept (#4196) by @nk-batman21. The onClose hook is registered for FailoverClusterClient (#4000) by @roganartu

  • Sentinel: the failover dialer uses DialTimeout, so Subscribe to an unreachable master does not hang until the kernel gives up (#4067) by @lazerg. Unused replica clients from empty or filtered discovery replies are closed (#4220) by @Grivn

  • Cluster commands: ClusterAddSlotsRange and ClusterDelSlotsRange return an error for an invalid range and do not panic (#4225) by @ump45nose. ClusterShards rejects an odd-length slots array (#4282) by @saddamr3e

Protocol and reply parsing

  • Error elements inside arrays: IntSliceCmd (TS.MADD, CMS.INCRBY), UintSliceCmd, FloatSliceCmd, StringSliceCmd and BoolSliceCmd now read the whole array when an element is a Redis error. Before, they stopped at that element, and in a pipeline the next command read the rest as its own reply (#4071 by @saddamr3e, #4137 by @lazerg). The reader also keeps typed errors (MOVED, ASK, LOADING) as array elements (#4274) by @ndyakov
  • Reply desync hardening: ReadSlice checks the type of the header (#4219) by @saddamr3e. Discard counts a RESP3 attribute and the element it decorates as one element (#4318) by @waw666waw666. VoidProcessor consumes push frames whose first element is not a string (#4046) by @saddamr3e. A buffered push drain stops at a frame boundary, so a slow handler no longer makes a healthy connection fail with i/o timeout and drops the CSC entries fetched on it (#4224) by @ndyakov
  • RESP3 numbers in Cmd: Cmd.Float64, Float32 and their slice forms accept RESP3 doubles (#4088) by @LindseyZ1205. Text, Int64, StringSlice and related accessors accept RESP3 doubles and big numbers (#4138) by @lazerg
  • 32-bit platforms: scanning a value that does not fit in uint returns an error (#4081) by @MokiMeow, and Cmd.Int() rejects an out-of-range integer reply (#4256) by @b84330808
  • CLIENT INFO values: the parser accepts = in values such as the client name and lib-name. ClientInfo() and ACLLog() failed on them (#4096) by @LindseyZ1205

Commands and search

  • RediSearch: FT.SEARCH on RESP2 reads the per-document score, payload and sort key correctly with NoContent (#4099) by @jayhemnani9910. EXPLAINSCORE and an empty RETURN parse correctly (#4140) by @lazerg. FTInfo reads fractional doubles on RESP2 (#4207) by @ump45nose and the CASESENSITIVE flag of a TAG field (#4291) by @Dv2Corn. FTCreate sends SORTABLE, UNF and NOINDEX after the field options (#4199) by @RaphaelFakhri. FTSearchQuery applies CountOnly and keeps the offset (#4268) by @RaphaelFakhri. FTHybridWithArgs always sends the COMBINE argument count (#4298) by @lazerg. NewDropIndexBuilder(...).DeleteDocs() does not panic (#4105) by @PragalvaXFREZ
  • nil options no longer panic: FTExplainWithArgs, FTSynUpdateWithArgs and CFReserveWithArgs (#4248) by @hktitof, HGetEXWithArgs and HSetEXWithArgs (#4214) by @LindseyZ1205. JSONGetWithArgs with nil options sends the paths (#4185) by @LindseyZ1205
  • Argument order and count: IGNORE is sent before LABELS in the five time-series builders that support it (for example TSAddWithArgs and TSCreateWithArgs), because LABELS takes the rest of the arguments (#4160) by @yuefdev. ModuleLoadex sends ARGS once (#4191) by @LindseyZ1205. HSet and HMSet flatten typed map[string]T arguments such as map[string]int (#4235) by @ump45nose
  • Sub-millisecond timeouts: XRead, XReadGroup, XClaim, Wait and WaitAOF no longer send a positive timeout shorter than 1 ms as 0 (wait forever) (#4292) by @Dv2Corn
  • Geo queries: GeoRadius and GeoRadiusByMember no longer change the caller's *GeoRadiusQuery, so the second call with the same query works (#4173) by @LindseyZ1205
  • Scripts: a queued SCRIPT LOAD (in a pipeline) no longer gives an empty digest to NewScriptServerSHA (#4226) by @ump45nose
  • HOTKEYS: the CPU time fields use the names that the server sends (#4206) by @ump45nose
  • Struct scan: the value passed to a TextUnmarshaler or BinaryUnmarshaler is a copy, so an unmarshaler that decodes in place cannot change the reply (#4078) by @saddamr3e

Observability

  • redisotel-native: server.port no longer contains the pool id (for example 6379_1) (#4086) by @zhechen-cs. The connection count includes connections that existed before Init (#4089) by @ninadsutrave. Pending requests and state no longer drift across Init and Shutdown (#4142) by @ninadsutrave. The pool usage metric does not underflow when pool statistics disagree (#4031) by @andystaples

⚡ Performance

  • Short bulk strings in one write: Writer writes the header of a bulk string with one call, and writes the value and CRLF together when the value is 32 bytes or less. Benchmarks of WriteArgs into a bufio.Writer showed 15% to 31% less time for each command on three CPUs, for example -24% to -26% for GET key and -27% to -31% for an 8-pair MSET. The bytes on the wire do not change (#4281) by @g4lb
  • OTel metrics and full duplex: see the full-duplex highlight above (#4269)

🧪 Testing & Infrastructure

  • CI images: the test matrix uses Redis 8.10.3 and the 8.12 release candidate (8.12-rc3) (#4231, #4241, #4264, #4288) by @ndyakov
  • 32-bit build: a 386 CI job runs go vet and the tests, and the root tests compile for 32-bit (#4139) by @lazerg
  • Dependencies: the OpenTelemetry SDK and golang.org/x/net are upgraded in the OTel adapters and examples, to fix govulncheck findings (#4087) by @ofekshenawa
  • Regression tests: a cold-start full-duplex cluster test bounds the number of connections (#4223) by @ndyakov, and a test locks the MSETEX multi-shard split (#4314) by @Zoomish
  • Test structure and flaky specs: the autopipeline tests are consolidated from 28 files into 5 (#4091), and some tests move to the standard testing package (#4085, #4187). Fixes for flaky or Redis Enterprise-incompatible specs: #4172, #4180, #4210, #4211, #4215, #4218, #4255, #4260, #4270, #4307

👥 Contributors

We'd like to thank all the contributors who worked on this release!

@AdamMagued, @aditya-rai-5, @andystaples, @aozmekik, @Avv123, @b84330808, @CaliLuke, @cpoole, @dunglas, @Dv2Corn, @g4lb, @Grivn, @harshitdeora, @hktitof, @jayhemnani9910, @jjj-n, @khasanovbi, @lazerg, @LindseyZ1205, @meruiden, @MokiMeow, @mtnaskari, @Nacl-HCl, @ndyakov, @ninadsutrave, @nk-batman21, @ofekshenawa, @othoathydrahost, @pcbeingused333, @PerryLink, @PragalvaXFREZ, @Purv-Kabaria, @RaphaelFakhri, @rleungx, @roganartu, @saddamr3e, @Samuel-Sorial, @SergeiSadov, @soujasK, @spicysharkco, @SulimanAbdulrazzaq, @TastyHeadphones, @ump45nose, @VedantMadane, @vladvildanov, @vlady-kotsev, @waw666waw666, @xh123637, @xixi-box, @yamaankhan20, @yuefdev, @zhechen-cs, @Zoomish


Full Changelog: v9.23.0...v9.24.0-beta.1

Don't miss a new go-redis release

NewReleases is sending notifications on new releases.