[dependabot]

Bumps github.com/redis/go-redis/extra/redisotel/v9 from 9.18.0 to 9.23.0.

Release notes

Sourced from github.com/redis/go-redis/extra/redisotel/v9's releases.

9.23.0

This is a minor release. It contains everything from 9.23.0-beta.1, so these notes cover the full 9.22.0 → 9.23.0 upgrade.

⚠️ Changes to check when you upgrade from 9.22.0:

• Go 1.26 is now the minimum Go version (#4060). The go directive of the root module and of the submodules moved from 1.24 to 1.26, as part of the dependency bumps for govulncheck findings. Projects that build with Go 1.24 or 1.25 must upgrade the toolchain.
• Each Client has a dedicated pipeline connection pool (#4002). This includes failover clients and cluster node clients. Pipeline, TxPipeline, and autopipeline operations use this pool, so they do not compete with regular commands for main-pool connections. The pool is for burst capacity only: it dials only when necessary, an unused pool holds zero connections, and if the pool is full an operation uses the main pool. Set PipelinePoolSize: -1 to get the previous single-pool behavior.
• AutoPipelineOptions.MaxBatchBytes has a default of 128 KiB (before, there was no limit). The default prevents a write/reply deadlock; it is not a throughput control. The buffers of the pipeline pool also have a default of 128 KiB.

⚠️ Changes to experimental APIs since 9.23.0-beta.1 (no effect if you upgrade from 9.22.0):

• AutoPipelineOptions.FullDuplexFastSubmit is removed (#4018). The submit channel is replaced by a queue that takes a full wave of commands in one lock, so the separate fast path is not necessary. Remove the field from your options.
• Options.ClientSideCacheRefreshRecencyWindow is removed (#4034). Refresh-on-invalidate now reads again every cached entry of an invalidated key and does not look at how recently the entry was read. On a write-heavy keyspace this means refresh traffic across the full resident cache. Remove the field from your options.
• LocalCache.LRUClock() is removed (#4034). It returned the recency token that the removed recency window used, and has no replacement. Remove calls to it.
• ClientSideCacheInvalidationBatchWindow served stale values in beta.1 (#4033). Under load the batcher applied only ~15% of invalidations, and the cache served invalidated values with a ~100% hit rate. 9.23.0 applies all of them. If you used this option on beta.1, upgrade.

🚀 Highlights

Full-Duplex Auto-Pipelining (Experimental)

Set AutoPipelineOptions.FullDuplex to enable the full-duplex mode of the automatic pipeliner. The default mode sends one batch per round trip. In full-duplex mode the engine holds one pipeline-pool connection, and a writer goroutine and a reader goroutine move the ordered command stream in both directions at the same time. On a high-latency link each command completes in approximately one round-trip time (RTT) on a single connection. On a 50 ms WAN profile: ~389k ops/s at 52 ms p50, against ~207k ops/s at 116 ms on the half-duplex ordered path.

The mode works on both faces of a standalone Client (AutoPipeline() and AsyncAutoPipeline()), and natively on a ClusterClient when routing goes to masters only. On a cluster, one child engine per master sends each command to the node that owns its slot, and follows MOVED / ASK redirects and retryable replies (LOADING, READONLY, ...) through the usual cluster redirect path. With replica routing (ReadOnly, RouteByLatency, or RouteRandomly), the autopipeliner uses the half-duplex shard flushers, which obey the shard picker. Config().FullDuplex reports the mode in effect.

Options:

• FullDuplexWindow — the maximum number of commands in flight (backpressure).
• FullDuplexIdleTimeout and FullDuplexMaxHold — when the engine returns the held connection to the pool. The pool hooks then do re-authentication and maintenance-notification handoffs.
• MaxFlushDelay — now also used in full-duplex mode (#4014). The writer waits only when enough commands are in flight, so low-concurrency callers keep the 1×RTT behavior. With MaxFlushDelay=250µs at 1024 callers: +22% ops/s and −26% p99. The default is 0, so this is opt-in.
• NumShards — on a standalone Client, the number of full-duplex engines (#4022). See below.

Pipeline() on the full-duplex connection (#4021). Standalone Client only. Before, ap.Pipeline() used a pooled connection for each Exec. It now submits the batch to the engine as one contiguous run (FDPipelined), so replies come back in submit order, as on a dedicated connection. At 256–1024 callers with 10 commands per Exec: +49% to +52% throughput on one socket instead of ~150. Errors, retries, hooks, and metrics behave like an ordinary pipeline. A batch that cannot use the stream (a blocking command, a per-command read timeout, a diverted command, or a batch larger than the queue) falls back to the ordinary pipeline. TxPipeline stays on a pooled connection, because MULTI/EXEC needs connection affinity. On a ClusterClient, ap.Pipeline() is still the ordinary cluster pipeline. A batch that the queue cannot admit now reserves its slots and waits in a first-come line, so single commands cannot starve a long pipeline (#4057, merged as part of #4021) by @​vlady-kotsev.

Several engines on one client (#4022). With FullDuplex, NumShards > 1 now means N engines on one client: one pool, one hook chain, N held connections. Commands route by a hash of their first key, so commands on the same key stay on one connection and keep their order without Unordered. Keyless commands are distributed round-robin. An FDPipelined batch stays on one engine. The pipeline pool must hold at least NumShards connections for each full-duplex autopipeliner. More engines help only under high concurrency: with 10-command pipelines, 8 engines were 30% slower than 1 at 8 callers, equal at about 128 callers, and 2.2× faster at 1024 callers.

The full-duplex path supports blocking commands, Options.Limiter (one admission per written batch), per-command and pipeline hooks, OTel metrics, retry budgets (a NoRetry command is never sent again after it is on the wire), and seamless maintenance handoffs. Limits of the stream model:

• A hook can observe a command, but cannot stop it. A ProcessHook that returns without calling next does not cancel the command on the full-duplex path; the command is already queued on the held connection. Run policy or kill-switch hooks on a plain client or on the half-duplex autopipeliner.
• Diverted commands are not ordered with the stream. The engine diverts blocking commands, connection-hostile commands, and managed HIMPORT commands. On the async face, wait for the result of a diverted command before you submit a command that depends on it.
• Single commands on the full-duplex stream do not use the client-side cache. If CSC is enabled, a cacheable command on the stream does not read or write the cache.
• Duration metrics are recorded per reply. A command that fails before it reaches the reader (lease failure, limiter denial, retry exhaustion, close) calls the error callback but records no operation-duration sample.

rdb := redis.NewClient(&redis.Options{
	Addr: "localhost:6379",
	AutoPipelineOptions: &redis.AutoPipelineOptions{
		FullDuplex: true, // stream commands on one held connection, ~1 RTT each
	},
})
defer rdb.Close()
</tr></table> 

... (truncated)

Changelog

Sourced from github.com/redis/go-redis/extra/redisotel/v9's changelog.

9.23.0 (2026-10-05)

This is a minor release. It contains everything from 9.23.0-beta.1, so these notes cover the full 9.22.0 → 9.23.0 upgrade. The release adds:

• An experimental full-duplex mode for the automatic pipeliner, now with Pipeline() on the full-duplex connection and several engines on one client.
• Refresh-on-invalidate and miss coalescing for client-side caching, and a faster cache read path.
• Better distribution of cluster reads under latency-based routing.
• Client.EphemeralConn, an opt-in limit on queued autopipeline commands, and many stability fixes.

⚠️ Changes to check when you upgrade from 9.22.0:

• Go 1.26 is now the minimum Go version (#4060). The go directive of the root module and of the submodules moved from 1.24 to 1.26, as part of the dependency bumps for govulncheck findings. Projects that build with Go 1.24 or 1.25 must upgrade the toolchain.
• Each Client has a dedicated pipeline connection pool (#4002). This includes failover clients and cluster node clients. Pipeline, TxPipeline, and autopipeline operations use this pool, so they do not compete with regular commands for main-pool connections. The pool is for burst capacity only: it dials only when necessary, an unused pool holds zero connections, and if the pool is full an operation uses the main pool. Set PipelinePoolSize: -1 to get the previous single-pool behavior.
• AutoPipelineOptions.MaxBatchBytes has a default of 128 KiB (before, there was no limit). The default prevents a write/reply deadlock; it is not a throughput control. The buffers of the pipeline pool also have a default of 128 KiB.

⚠️ Changes to experimental APIs since 9.23.0-beta.1 (no effect if you upgrade from 9.22.0):

• AutoPipelineOptions.FullDuplexFastSubmit is removed (#4018). The submit channel is replaced by a queue that takes a full wave of commands in one lock, so the separate fast path is not necessary. Remove the field from your options.
• Options.ClientSideCacheRefreshRecencyWindow is removed (#4034). Refresh-on-invalidate now reads again every cached entry of an invalidated key and does not look at how recently the entry was read. On a write-heavy keyspace this means refresh traffic across the full resident cache. Remove the field from your options.
• LocalCache.LRUClock() is removed (#4034). It returned the recency token that the removed recency window used, and has no replacement. Remove calls to it.
• ClientSideCacheInvalidationBatchWindow served stale values in beta.1 (#4033). Under load the batcher applied only ~15% of invalidations, and the cache served invalidated values with a ~100% hit rate. 9.23.0 applies all of them. If you used this option on beta.1, upgrade.

🚀 Highlights

Full-Duplex Auto-Pipelining (Experimental)

Set AutoPipelineOptions.FullDuplex to enable the full-duplex mode of the automatic pipeliner. The default mode sends one batch per round trip. In full-duplex mode the engine holds one pipeline-pool connection, and a writer goroutine and a reader goroutine move the ordered command stream in both directions at the same time. On a high-latency link each command completes in approximately one round-trip time (RTT) on a single connection. On a 50 ms WAN profile: ~389k ops/s at 52 ms p50, against ~207k ops/s at 116 ms on the half-duplex ordered path.

The mode works on both faces of a standalone Client (AutoPipeline() and AsyncAutoPipeline()), and natively on a ClusterClient when routing goes to masters only. On a cluster, one child engine per master sends each command to the node that owns its slot, and follows MOVED / ASK redirects and retryable replies (LOADING, READONLY, ...) through the usual cluster redirect path. With replica routing (ReadOnly, RouteByLatency, or RouteRandomly), the autopipeliner uses the half-duplex shard flushers, which obey the shard picker. Config().FullDuplex reports the mode in effect.

Options:

• FullDuplexWindow — the maximum number of commands in flight (backpressure).
• FullDuplexIdleTimeout and FullDuplexMaxHold — when the engine returns the held connection to the pool. The pool hooks then do re-authentication and maintenance-notification handoffs.
• MaxFlushDelay — now also used in full-duplex mode (#4014). The writer waits only when enough commands are in flight, so low-concurrency callers keep the 1×RTT behavior. With MaxFlushDelay=250µs at 1024 callers: +22% ops/s and −26% p99. The default is 0, so this is opt-in.
• NumShards — on a standalone Client, the number of full-duplex engines (#4022). See below.

Pipeline() on the full-duplex connection (#4021). Standalone Client only. Before, ap.Pipeline() used a pooled connection for each Exec. It now submits the batch to the engine as one contiguous run (FDPipelined), so replies come back in submit order, as on a dedicated connection. At 256–1024 callers with 10 commands per Exec: +49% to +52% throughput on one socket instead of ~150. Errors, retries, hooks, and metrics behave like an ordinary pipeline. A batch that cannot use the stream (a blocking command, a per-command read timeout, a diverted command, or a batch larger than the queue) falls back to the ordinary pipeline. TxPipeline stays on a pooled connection, because MULTI/EXEC needs connection affinity. On a ClusterClient, ap.Pipeline() is still the ordinary cluster pipeline. A batch that the queue cannot admit now reserves its slots and waits in a first-come line, so single commands cannot starve a long pipeline (#4057, merged as part of #4021) by @​vlady-kotsev.

Several engines on one client (#4022). With FullDuplex, NumShards > 1 now means N engines on one client: one pool, one hook chain, N held connections. Commands route by a hash of their first key, so commands on the same key stay on one connection and keep their order without Unordered. Keyless commands are distributed round-robin. An FDPipelined batch stays on one engine. The pipeline pool must hold at least NumShards connections for each full-duplex autopipeliner. More engines help only under high concurrency: with 10-command pipelines, 8 engines were 30% slower than 1 at 8 callers, equal at about 128 callers, and 2.2× faster at 1024 callers.

The full-duplex path supports blocking commands, Options.Limiter (one admission per written batch), per-command and pipeline hooks, OTel metrics, retry budgets (a NoRetry command is never sent again after it is on the wire), and seamless maintenance handoffs. Limits of the stream model:

• A hook can observe a command, but cannot stop it. A ProcessHook that returns without calling next does not cancel the command on the full-duplex path; the command is already queued on the held connection. Run policy or kill-switch hooks on a plain client or on the half-duplex autopipeliner.
• Diverted commands are not ordered with the stream. The engine diverts blocking commands, connection-hostile commands, and managed HIMPORT commands. On the async face, wait for the result of a diverted command before you submit a command that depends on it.
• Single commands on the full-duplex stream do not use the client-side cache. If CSC is enabled, a cacheable command on the stream does not read or write the cache.
• Duration metrics are recorded per reply. A command that fails before it reaches the reader (lease failure, limiter denial, retry exhaustion, close) calls the error callback but records no operation-duration sample.

rdb := redis.NewClient(&redis.Options{
</tr></table> 

... (truncated)

Commits

• 6f2b263 chore(release): v9.23.0 (#4074)
• ec50c69 feat(conn): add DisposableConn to discard on Close (#4039)
• e3d2548 fix(pool): fix ConnStateMachine waiter races (#4072)
• 2ee6d86 fix(pool): wake waiters when a connection is released to the pool (#4028)
• 83d45ed fix:(proto): support uintptr in WriteArg (#4054)
• c2002af docs(example): add full-duplex examples (#4061)
• b5f8864 feat(autopipeline): add MaxQueuedCommands hard limit (#4070)
• 89ecbc5 fix(options): reject invalid skip_verify values in ParseURL (#4064)
• f813adc feat(autopipeline): run N full-duplex engines on one client (#4022)
• 349bcf4 feat(autopipeline): run Pipeline on the full-duplex connection (#4021)
• Additional commits viewable in compare view