Merge pull request #32833 from vespa-engine/hmusum/support-setting-zo…

…okeeper-pre-alloc-size

Support setting zookeeper preAllocSize with feature flag

config-model-api/abi-spec.json

@@ -758,10 +758,10 @@
       "public static final enum com.yahoo.config.application.api.ValidationId accessControl",
       "public static final enum com.yahoo.config.application.api.ValidationId globalEndpointChange",
       "public static final enum com.yahoo.config.application.api.ValidationId zoneEndpointChange",
-      "public static final enum com.yahoo.config.application.api.ValidationId redundancyIncrease",
       "public static final enum com.yahoo.config.application.api.ValidationId redundancyOne",
       "public static final enum com.yahoo.config.application.api.ValidationId pagedSettingRemoval",
-      "public static final enum com.yahoo.config.application.api.ValidationId certificateRemoval"
+      "public static final enum com.yahoo.config.application.api.ValidationId certificateRemoval",
+      "public static final enum com.yahoo.config.application.api.ValidationId redundancyIncrease"
     ]
   },
   "com.yahoo.config.application.api.ValidationOverrides$Allow" : {
@@ -1341,7 +1341,8 @@
       "public boolean enforceStrictlyIncreasingClusterStateVersions()",
       "public boolean distributionConfigFromClusterController()",
       "public boolean useLegacyWandQueryParsing()",
-      "public boolean forwardAllLogLevels()"
+      "public boolean forwardAllLogLevels()",
+      "public long zookeeperPreAllocSize()"
     ],
     "fields" : [ ]
   },

config-model-api/src/main/java/com/yahoo/config/model/api/ModelContext.java

@@ -123,6 +123,7 @@ interface FeatureFlags {
         @ModelFeatureFlag(owners = {"vekterli"}) default boolean distributionConfigFromClusterController() { return false; }
         @ModelFeatureFlag(owners = {"arnej"}) default boolean useLegacyWandQueryParsing() { return true; }
         @ModelFeatureFlag(owners = {"hmusum"}) default boolean forwardAllLogLevels() { return true; }
+        @ModelFeatureFlag(owners = {"hmusum"}) default long zookeeperPreAllocSize() { return 65536L; }
     }
 
     /** Warning: As elsewhere in this package, do not make backwards incompatible changes that will break old config models! */

config-model/src/main/java/com/yahoo/vespa/model/container/configserver/ConfigserverCluster.java

@@ -4,6 +4,7 @@
 import com.yahoo.cloud.config.ConfigserverConfig;
 import com.yahoo.cloud.config.CuratorConfig;
 import com.yahoo.cloud.config.ZookeeperServerConfig;
+
 import com.yahoo.config.model.producer.TreeConfigProducer;
 import com.yahoo.config.provision.Environment;
 import com.yahoo.config.provision.RegionName;
@@ -20,6 +21,8 @@
 import java.util.Optional;
 import java.util.stream.IntStream;
 
+import static com.yahoo.config.model.api.ModelContext.FeatureFlags;
+
 /**
  * Represents a config server cluster.
  *
@@ -34,11 +37,13 @@ public class ConfigserverCluster extends TreeConfigProducer
         ZookeeperServerConfig.Producer {
 
     private final ConfigOptions options;
+    private final FeatureFlags featureFlags;
     private ContainerCluster<?> containerCluster;
 
-    public ConfigserverCluster(TreeConfigProducer<?> parent, String subId, ConfigOptions options) {
+    public ConfigserverCluster(TreeConfigProducer<?> parent, String subId, ConfigOptions options, FeatureFlags featureFlags) {
         super(parent, subId);
         this.options = options;
+        this.featureFlags = featureFlags;
     }
 
     public void setContainerCluster(ContainerCluster<?> containerCluster) {
@@ -87,6 +92,7 @@ public void getConfig(ZookeeperServerConfig.Builder builder) {
         builder.reconfigureEnsemble(!isHostedVespa);
         builder.snapshotMethod(options.zooKeeperSnapshotMethod());
         builder.juteMaxBuffer(options.zookeeperJuteMaxBuffer());
+        builder.preAllocSizeKb(featureFlags.zookeeperPreAllocSize());
     }
 
     @Override

config-model/src/main/java/com/yahoo/vespa/model/container/xml/ConfigServerContainerModelBuilder.java

@@ -28,7 +28,7 @@ public ConfigServerContainerModelBuilder(ConfigOptions options) {
     @Override
     public void doBuild(ContainerModel model, Element spec, ConfigModelContext modelContext) {
         ConfigserverCluster cluster = new ConfigserverCluster(modelContext.getParentProducer(), "configserver",
-                                                              options);
+                                                              options, modelContext.featureFlags());
         super.doBuild(model, spec, modelContext.withParent(cluster));
         cluster.setContainerCluster(model.getCluster());
     }

configdefinitions/src/vespa/zookeeper-server.def

@@ -61,3 +61,7 @@ learnerAsyncSending bool default=false
 # Whether the ZooKeeper ensemble should be reconfigured automatically if servers change. This only has an effect if
 # dynamicReconfiguration=true.
 reconfigureEnsemble bool default=true
+
+# To avoid seeks ZooKeeper allocates space in the transaction log file in blocks of preAllocSize kilobytes. The default block size is 64M
+# Can/should be reduced for tests
+preAllocSizeKb long default=65536

configserver/src/main/java/com/yahoo/vespa/config/server/deploy/ModelContextImpl.java

@@ -215,6 +215,7 @@ public static class FeatureFlags implements ModelContext.FeatureFlags {
         private final boolean distributionConfigFromClusterController;
         private final boolean useLegacyWandQueryParsing;
         private final boolean forwardAllLogLevels;
+        private final long zookeeperPreAllocSize;
 
         public FeatureFlags(FlagSource source, ApplicationId appId, Version version) {
             this.defaultTermwiseLimit = Flags.DEFAULT_TERM_WISE_LIMIT.bindTo(source).with(appId).with(version).value();
@@ -258,7 +259,7 @@ public FeatureFlags(FlagSource source, ApplicationId appId, Version version) {
             this.sortBlueprintsByCost = Flags.SORT_BLUEPRINTS_BY_COST.bindTo(source).with(appId).with(version).value();
             this.persistenceThreadMaxFeedOpBatchSize = Flags.PERSISTENCE_THREAD_MAX_FEED_OP_BATCH_SIZE.bindTo(source).with(appId).with(version).value();
             this.logserverOtelCol = Flags.LOGSERVER_OTELCOL_AGENT.bindTo(source).with(appId).with(version).value();
-            this.sharedHosts = PermanentFlags.SHARED_HOST.bindTo(source).with( appId).with(version).value();
+            this.sharedHosts = PermanentFlags.SHARED_HOST.bindTo(source).with(appId).with(version).value();
             this.adminClusterArchitecture = Architecture.valueOf(PermanentFlags.ADMIN_CLUSTER_NODE_ARCHITECTURE.bindTo(source).with(appId).with(version).value());
             this.logserverNodeMemory = PermanentFlags.LOGSERVER_NODE_MEMORY.bindTo(source).with(appId).with(version).value();
             this.symmetricPutAndActivateReplicaSelection = Flags.SYMMETRIC_PUT_AND_ACTIVATE_REPLICA_SELECTION.bindTo(source).with(appId).with(version).value();
@@ -267,6 +268,7 @@ public FeatureFlags(FlagSource source, ApplicationId appId, Version version) {
             this.distributionConfigFromClusterController = Flags.DISTRIBUTION_CONFIG_FROM_CLUSTER_CONTROLLER.bindTo(source).with(appId).with(version).value();
             this.useLegacyWandQueryParsing = Flags.USE_LEGACY_WAND_QUERY_PARSING.bindTo(source).with(appId).with(version).value();
             this.forwardAllLogLevels = PermanentFlags.FORWARD_ALL_LOG_LEVELS.bindTo(source).with(appId).with(version).value();
+            this.zookeeperPreAllocSize = Flags.ZOOKEEPER_PRE_ALLOC_SIZE_KIB.bindTo(source).value();
         }
 
         @Override public int heapSizePercentage() { return heapPercentage; }
@@ -324,6 +326,7 @@ public FeatureFlags(FlagSource source, ApplicationId appId, Version version) {
         @Override public boolean distributionConfigFromClusterController() { return distributionConfigFromClusterController; }
         @Override public boolean useLegacyWandQueryParsing() { return useLegacyWandQueryParsing; }
         @Override public boolean forwardAllLogLevels() { return forwardAllLogLevels; }
+        @Override public long zookeeperPreAllocSize() { return zookeeperPreAllocSize; }
     }
 
     public static class Properties implements ModelContext.Properties {

flags/src/main/java/com/yahoo/vespa/flags/Flags.java

@@ -484,6 +484,13 @@ public class Flags {
             "Whether node snapshots should be created when host storage is discarded",
             "Takes effect immediately");
 
+    public static final UnboundLongFlag ZOOKEEPER_PRE_ALLOC_SIZE_KIB = defineLongFlag(
+            "zookeeper-pre-alloc-size", 65536,
+            List.of("hmusum"), "2024-11-11", "2025-01-11",
+            "Setting for zookeeper.preAllocSize flag in KiB, can be reduced from default value "
+            + "e.g. when running tests to avoid writing a large, sparse, mostly unused file",
+            "Takes effect on restart of Docker container");
+
     /** WARNING: public for testing: All flags should be defined in {@link Flags}. */
     public static UnboundBooleanFlag defineFeatureFlag(String flagId, boolean defaultValue, List<String> owners,
                                                        String createdAt, String expiresAt, String description,

zookeeper-server/zookeeper-server-common/src/main/java/com/yahoo/vespa/zookeeper/Configurator.java

@@ -56,6 +56,8 @@ public Configurator(ZookeeperServerConfig zookeeperServerConfig) {
         System.setProperty("zookeeper.learner.asyncSending", String.valueOf(zookeeperServerConfig.learnerAsyncSending()));
         // Enable creation of TTL Nodes.
         System.setProperty("zookeeper.extendedTypesEnabled", "true");
+        // Space in the transaction log file in kilobytes.
+        System.setProperty("zookeeper.preAllocSize", String.valueOf(zookeeperServerConfig.preAllocSizeKb()));
     }
 
     void writeConfigToDisk() {

zookeeper-server/zookeeper-server-common/src/test/java/com/yahoo/vespa/zookeeper/ConfiguratorTest.java

@@ -64,12 +64,12 @@ public void config_is_written_correctly_with_one_server() {
     }
 
     @Test
-    public void config_is_written_correctly_with_multiple_servers() throws IOException {
+    public void config_is_written_correctly_with_multiple_servers() {
         three_config_servers(false);
     }
 
     @Test
-    public void config_is_written_correctly_with_multiple_servers_on_hosted_vespa() throws IOException {
+    public void config_is_written_correctly_with_multiple_servers_on_hosted_vespa() {
         three_config_servers(true);
     }
 
@@ -155,7 +155,7 @@ public void test_parsing_config() throws IOException {
                      Files.readString(cfgFile.toPath()));
     }
 
-    private void three_config_servers(boolean hosted) throws IOException {
+    private void three_config_servers(boolean hosted) {
         ZookeeperServerConfig.Builder builder = new ZookeeperServerConfig.Builder();
         builder.zooKeeperConfigFile(cfgFile.getAbsolutePath());
         builder.server(newServer(0, "foo", 123, 321, false));