Merge branch 'slack-22.0' into tanjin-scatter-update-limit-optout

Author: tanjinx

go/flags/endtoend/vtcombo.txt

@@ -44,6 +44,7 @@ Flags:
       --config-path strings                                              Paths to search for config files in. (default [{{ .Workdir }}])
       --config-persistence-min-interval duration                         minimum interval between persisting dynamic config changes back to disk (if no change has occurred, nothing is done). (default 1s)
       --config-type string                                               Config file type (omit to infer config type from file extension).
+      --consolidator-cache-proto3-rows                                   If true, the consolidation leader pre-caches proto3-encoded rows so that waiters avoid redundant encoding work.
       --consolidator-query-waiter-cap int                                Configure the maximum number of clients allowed to wait on the consolidator.
       --consolidator-query-waiter-cap-method string                      Configure the method when consolidator waiter cap is exceeded. Options: fallthrough, reject. (default "fallthrough")
       --consolidator-stream-query-size int                               Configure the stream consolidator query size in bytes. Setting to 0 disables the stream consolidator. (default 2097152)
@@ -295,20 +296,23 @@ Flags:
       --queryserver-config-query-cache-memory int                        query server query cache size in bytes, maximum amount of memory to be used for caching. vttablet analyzes every incoming query and generate a query plan, these plans are being cached in a lru cache. This config controls the capacity of the lru cache. (default 33554432)
       --queryserver-config-query-pool-max-idle-count int                 query server query pool - maximum number of idle connections to retain in the pool. Use this to balance between faster response times during traffic bursts and resource efficiency during low-traffic periods.
       --queryserver-config-query-pool-timeout duration                   query server query pool timeout, it is how long vttablet waits for a connection from the query pool. If set to 0 (default) then the overall query timeout is used instead.
+      --queryserver-config-query-pool-waiter-cap uint                    query server query pool waiter cap is the maximum number of queries allowed to wait for a connection from the pool. If set to 0 (default) then there is no limit.
       --queryserver-config-query-timeout duration                        query server query timeout, this is the query timeout in vttablet side. If a query takes more than this timeout, it will be killed. (default 30s)
       --queryserver-config-schema-change-signal                          query server schema signal, will signal connected vtgates that schema has changed whenever this is detected. VTGates will need to have -schema_change_signal enabled for this to work (default true)
       --queryserver-config-schema-reload-time duration                   query server schema reload time, how often vttablet reloads schemas from underlying MySQL instance. vttablet keeps table schemas in its own memory and periodically refreshes it from MySQL. This config controls the reload time. (default 30m0s)
       --queryserver-config-stream-buffer-size int                        query server stream buffer size, the maximum number of bytes sent from vttablet for each stream call. It's recommended to keep this value in sync with vtgate's stream_buffer_size. (default 32768)
       --queryserver-config-stream-pool-max-idle-count int                query server stream pool - maximum number of idle connections to retain in the pool. Use this to balance between faster response times during traffic bursts and resource efficiency during low-traffic periods.
       --queryserver-config-stream-pool-size int                          query server stream connection pool size, stream pool is used by stream queries: queries that return results to client in a streaming fashion (default 200)
       --queryserver-config-stream-pool-timeout duration                  query server stream pool timeout, it is how long vttablet waits for a connection from the stream pool. If set to 0 (default) then there is no timeout.
+      --queryserver-config-stream-pool-waiter-cap uint                   query server stream pool waiter cap is the maximum number of streaming queries allowed to wait for a connection from the pool. If set to 0 (default) then there is no limit.
       --queryserver-config-strict-table-acl                              only allow queries that pass table acl checks
       --queryserver-config-terse-errors                                  prevent bind vars from escaping in client error messages
       --queryserver-config-transaction-cap int                           query server transaction cap is the maximum number of transactions allowed to happen at any given point of a time for a single vttablet. E.g. by setting transaction cap to 100, there are at most 100 transactions will be processed by a vttablet and the 101th transaction will be blocked (and fail if it cannot get connection within specified timeout) (default 20)
       --queryserver-config-transaction-timeout duration                  query server transaction timeout, a transaction will be killed if it takes longer than this value (default 30s)
       --queryserver-config-truncate-error-len int                        truncate errors sent to client if they are longer than this value (0 means do not truncate)
       --queryserver-config-txpool-max-idle-count int                     query server transaction pool - maximum number of idle connections to retain in the pool. Use this to balance between faster response times during traffic bursts and resource efficiency during low-traffic periods.
       --queryserver-config-txpool-timeout duration                       query server transaction pool timeout, it is how long vttablet waits if tx pool is full (default 1s)
+      --queryserver-config-txpool-waiter-cap uint                        query server transaction pool waiter cap is the maximum number of transactions allowed to wait for a connection from the pool. If set to 0 (default) then there is no limit.
       --queryserver-config-warn-result-size int                          query server result size warning threshold, warn if number of rows returned from vttablet for non-streaming queries exceeds this
       --queryserver-enable-views                                         Enable views support in vttablet.
       --queryserver_enable_online_ddl                                    Enable online DDL. (default true)

go/flags/endtoend/vttablet.txt

@@ -78,6 +78,7 @@ Flags:
       --config-path strings                                              Paths to search for config files in. (default [{{ .Workdir }}])
       --config-persistence-min-interval duration                         minimum interval between persisting dynamic config changes back to disk (if no change has occurred, nothing is done). (default 1s)
       --config-type string                                               Config file type (omit to infer config type from file extension).
+      --consolidator-cache-proto3-rows                                   If true, the consolidation leader pre-caches proto3-encoded rows so that waiters avoid redundant encoding work.
       --consolidator-query-waiter-cap int                                Configure the maximum number of clients allowed to wait on the consolidator.
       --consolidator-query-waiter-cap-method string                      Configure the method when consolidator waiter cap is exceeded. Options: fallthrough, reject. (default "fallthrough")
       --consolidator-stream-query-size int                               Configure the stream consolidator query size in bytes. Setting to 0 disables the stream consolidator. (default 2097152)
@@ -287,20 +288,23 @@ Flags:
       --queryserver-config-query-cache-memory int                        query server query cache size in bytes, maximum amount of memory to be used for caching. vttablet analyzes every incoming query and generate a query plan, these plans are being cached in a lru cache. This config controls the capacity of the lru cache. (default 33554432)
       --queryserver-config-query-pool-max-idle-count int                 query server query pool - maximum number of idle connections to retain in the pool. Use this to balance between faster response times during traffic bursts and resource efficiency during low-traffic periods.
       --queryserver-config-query-pool-timeout duration                   query server query pool timeout, it is how long vttablet waits for a connection from the query pool. If set to 0 (default) then the overall query timeout is used instead.
+      --queryserver-config-query-pool-waiter-cap uint                    query server query pool waiter cap is the maximum number of queries allowed to wait for a connection from the pool. If set to 0 (default) then there is no limit.
       --queryserver-config-query-timeout duration                        query server query timeout, this is the query timeout in vttablet side. If a query takes more than this timeout, it will be killed. (default 30s)
       --queryserver-config-schema-change-signal                          query server schema signal, will signal connected vtgates that schema has changed whenever this is detected. VTGates will need to have -schema_change_signal enabled for this to work (default true)
       --queryserver-config-schema-reload-time duration                   query server schema reload time, how often vttablet reloads schemas from underlying MySQL instance. vttablet keeps table schemas in its own memory and periodically refreshes it from MySQL. This config controls the reload time. (default 30m0s)
       --queryserver-config-stream-buffer-size int                        query server stream buffer size, the maximum number of bytes sent from vttablet for each stream call. It's recommended to keep this value in sync with vtgate's stream_buffer_size. (default 32768)
       --queryserver-config-stream-pool-max-idle-count int                query server stream pool - maximum number of idle connections to retain in the pool. Use this to balance between faster response times during traffic bursts and resource efficiency during low-traffic periods.
       --queryserver-config-stream-pool-size int                          query server stream connection pool size, stream pool is used by stream queries: queries that return results to client in a streaming fashion (default 200)
       --queryserver-config-stream-pool-timeout duration                  query server stream pool timeout, it is how long vttablet waits for a connection from the stream pool. If set to 0 (default) then there is no timeout.
+      --queryserver-config-stream-pool-waiter-cap uint                   query server stream pool waiter cap is the maximum number of streaming queries allowed to wait for a connection from the pool. If set to 0 (default) then there is no limit.
       --queryserver-config-strict-table-acl                              only allow queries that pass table acl checks
       --queryserver-config-terse-errors                                  prevent bind vars from escaping in client error messages
       --queryserver-config-transaction-cap int                           query server transaction cap is the maximum number of transactions allowed to happen at any given point of a time for a single vttablet. E.g. by setting transaction cap to 100, there are at most 100 transactions will be processed by a vttablet and the 101th transaction will be blocked (and fail if it cannot get connection within specified timeout) (default 20)
       --queryserver-config-transaction-timeout duration                  query server transaction timeout, a transaction will be killed if it takes longer than this value (default 30s)
       --queryserver-config-truncate-error-len int                        truncate errors sent to client if they are longer than this value (0 means do not truncate)
       --queryserver-config-txpool-max-idle-count int                     query server transaction pool - maximum number of idle connections to retain in the pool. Use this to balance between faster response times during traffic bursts and resource efficiency during low-traffic periods.
       --queryserver-config-txpool-timeout duration                       query server transaction pool timeout, it is how long vttablet waits if tx pool is full (default 1s)
+      --queryserver-config-txpool-waiter-cap uint                        query server transaction pool waiter cap is the maximum number of transactions allowed to wait for a connection from the pool. If set to 0 (default) then there is no limit.
       --queryserver-config-warn-result-size int                          query server result size warning threshold, warn if number of rows returned from vttablet for non-streaming queries exceeds this
       --queryserver-enable-views                                         Enable views support in vttablet.
       --queryserver_enable_online_ddl                                    Enable online DDL. (default true)

go/pools/smartconnpool/pool.go

@@ -18,6 +18,7 @@ package smartconnpool
 
 import (
 	"context"
+	"errors"
 	"math/rand/v2"
 	"sync"
 	"sync/atomic"
@@ -39,6 +40,9 @@ var (
 	// ErrConnPoolClosed is returned when trying to get a connection from a closed conn pool
 	ErrConnPoolClosed = vterrors.New(vtrpcpb.Code_INTERNAL, "connection pool is closed")
 
+	// ErrPoolWaiterCapReached is returned when the waiter cap has been reached
+	ErrPoolWaiterCapReached = vterrors.New(vtrpcpb.Code_RESOURCE_EXHAUSTED, "connection pool waiter cap reached")
+
 	// PoolCloseTimeout is how long to wait for all connections to be returned to the pool during close
 	PoolCloseTimeout = 10 * time.Second
 )
@@ -52,6 +56,7 @@ type Metrics struct {
 	idleClosed           atomic.Int64
 	diffSetting          atomic.Int64
 	resetSetting         atomic.Int64
+	waiterCapRejected    atomic.Int64
 }
 
 func (m *Metrics) MaxLifetimeClosed() int64 {
@@ -86,6 +91,10 @@ func (m *Metrics) ResetSettingCount() int64 {
 	return m.resetSetting.Load()
 }
 
+func (m *Metrics) WaiterCapRejected() int64 {
+	return m.waiterCapRejected.Load()
+}
+
 type Connector[C Connection] func(ctx context.Context) (C, error)
 type RefreshCheck func() (bool, error)
 
@@ -95,6 +104,7 @@ type Config[C Connection] struct {
 	IdleTimeout     time.Duration
 	MaxLifetime     time.Duration
 	RefreshInterval time.Duration
+	MaxWaiters      uint
 	LogWait         func(time.Time)
 }
 
@@ -149,6 +159,9 @@ type ConnPool[C Connection] struct {
 		refreshInterval atomic.Int64
 		// logWait is called every time a client must block waiting for a connection
 		logWait func(time.Time)
+		// maxWaiters is the maximum number of clients that can be waiting for a connection;
+		// 0 means unlimited
+		maxWaiters uint
 	}
 
 	Metrics Metrics
@@ -165,7 +178,14 @@ func NewPool[C Connection](config *Config[C]) *ConnPool[C] {
 	pool.config.idleTimeout.Store(config.IdleTimeout.Nanoseconds())
 	pool.config.refreshInterval.Store(config.RefreshInterval.Nanoseconds())
 	pool.config.logWait = config.LogWait
+	pool.config.maxWaiters = config.MaxWaiters
 	pool.wait.init()
+	pool.wait.onWait = func() {
+		pool.Metrics.waitCount.Add(1)
+	}
+	pool.wait.onWaiterCapReached = func() {
+		pool.Metrics.waiterCapRejected.Add(1)
+	}
 
 	return pool
 }
@@ -374,8 +394,7 @@ func (pool *ConnPool[D]) RefreshInterval() time.Duration {
 	return time.Duration(pool.config.refreshInterval.Load())
 }
 
-func (pool *ConnPool[C]) recordWait(start time.Time) {
-	pool.Metrics.waitCount.Add(1)
+func (pool *ConnPool[C]) recordWaitDuration(start time.Time) {
 	pool.Metrics.waitTime.Add(time.Since(start).Nanoseconds())
 	if pool.config.logWait != nil {
 		pool.config.logWait(start)
@@ -612,11 +631,14 @@ func (pool *ConnPool[C]) get(ctx context.Context) (*Pooled[C], error) {
 			return nil, ErrConnPoolClosed
 		}
 
-		conn, err = pool.wait.waitForConn(ctx, nil, *closeChan)
+		conn, err = pool.wait.waitForConn(ctx, nil, *closeChan, pool.config.maxWaiters)
 		if err != nil {
+			if errors.Is(err, ErrPoolWaiterCapReached) {
+				return nil, err
+			}
 			return nil, ErrTimeout
 		}
-		pool.recordWait(start)
+		pool.recordWaitDuration(start)
 	}
 	// no connections available and no connections to wait for (pool is closed)
 	if conn == nil {
@@ -675,11 +697,14 @@ func (pool *ConnPool[C]) getWithSetting(ctx context.Context, setting *Setting) (
 			return nil, ErrConnPoolClosed
 		}
 
-		conn, err = pool.wait.waitForConn(ctx, setting, *closeChan)
+		conn, err = pool.wait.waitForConn(ctx, setting, *closeChan, pool.config.maxWaiters)
 		if err != nil {
+			if errors.Is(err, ErrPoolWaiterCapReached) {
+				return nil, err
+			}
 			return nil, ErrTimeout
 		}
-		pool.recordWait(start)
+		pool.recordWaitDuration(start)
 	}
 	// no connections available and no connections to wait for (pool is closed)
 	if conn == nil {
@@ -920,6 +945,9 @@ func (pool *ConnPool[C]) RegisterStats(stats *servenv.Exporter, name string) {
 	stats.NewCounterFunc(name+"GetSetting", "Tablet server conn pool get with setting count", func() int64 {
 		return pool.Metrics.GetSettingCount()
 	})
+	stats.NewCounterFunc(name+"WaiterCapRejected", "Tablet server conn pool waiter cap rejected", func() int64 {
+		return pool.Metrics.WaiterCapRejected()
+	})
 	stats.NewCounterFunc(name+"DiffSetting", "Number of times pool applied different setting", func() int64 {
 		return pool.Metrics.DiffSettingCount()
 	})