Migration Guide
Upgrading from 0.4.0 to 0.4.1
-
Since 0.4.1, Celeborn master adds a limit to the estimated partition size used for computing worker slots. This size is now constrained within the range specified by
celeborn.master.estimatedPartitionSize.minSizeandceleborn.master.estimatedPartitionSize.maxSize. -
Since 0.4.1, Celeborn changed the fallback configuration of
celeborn.client.rpc.getReducerFileGroup.askTimeout,celeborn.client.rpc.registerShuffle.askTimeoutandceleborn.client.rpc.requestPartition.askTimeoutfromceleborn.<module>.io.connectionTimeouttoceleborn.rpc.askTimeout.
Upgrading from 0.3 to 0.4
-
Since 0.4.0, Celeborn won't be compatible with Celeborn client that versions below 0.3.0. Note that: It's strongly recommended to use the same version of Client and Celeborn Master/Worker in production.
-
Since 0.4.0, Celeborn won't support
org.apache.spark.shuffle.celeborn.RssShuffleManager. -
Since 0.4.0, Celeborn changed the default value of
celeborn.<module>.io.numConnectionsPerPeerfrom2to1. -
Since 0.4.0, Celeborn has changed the names of the prometheus master and worker configuration as shown in the table below:
Key Before v0.4.0 Key After v0.4.0 celeborn.metrics.master.prometheus.hostceleborn.master.http.hostceleborn.metrics.master.prometheus.portceleborn.master.http.portceleborn.metrics.worker.prometheus.hostceleborn.worker.http.hostceleborn.metrics.worker.prometheus.portceleborn.worker.http.port -
Since 0.4.0, Celeborn deprecate
celeborn.worker.storage.baseDir.prefixandceleborn.worker.storage.baseDir.number. Please useceleborn.worker.storage.dirsinstead. -
Since 0.4.0, Celeborn deprecate
celeborn.storage.activeTypes. Please useceleborn.storage.availableTypesinstead. -
Since 0.4.0, Celeborn worker removes configuration
celeborn.worker.userResourceConsumption.update.interval. -
Since 0.4.0, Celeborn master metrics
PartitionWrittenis renamed asActiveShuffleSize. -
Since 0.4.0, Celeborn master metrics
PartitionFileCountis renamed asActiveShuffleFileCount.
Upgrading from 0.3.1 to 0.3.2
-
Since 0.3.1, Celeborn changed the default value of
raft.client.rpc.request.timeoutfrom3sto10s. -
Since 0.3.1, Celeborn changed the default value of
raft.client.rpc.watch.request.timeoutfrom10sto20s.
Upgrading from 0.3.0 to 0.3.1
-
Since 0.3.1, Celeborn changed the default value of
celeborn.worker.directMemoryRatioToResumefrom0.5to0.7. -
Since 0.3.1, Celeborn changed the default value of
celeborn.worker.monitor.disk.check.intervalfrom60to30. -
Since 0.3.1, name of JVM metrics changed, see details at CELEBORN-1007.
Upgrading from 0.2 to 0.3
-
Celeborn 0.2 Client is compatible with 0.3 Master/Server, it allows to upgrade Master/Worker first then Client. Note that: It's strongly recommended to use the same version of Client and Celeborn Master/Worker in production.
-
Since 0.3.0, the support of deprecated configurations
rss.*is removed. All configurations listed in 0.2.1 docs still take effect, but some of those are deprecated too, please read the bootstrap logs and follow the suggestion to migrate to the new configuration. -
From 0.3.0 on the default value for
celeborn.client.push.replicate.enabledis changed fromtruetofalse, users who want replication on should explicitly enable replication. For example, to enable replication for Spark users should add the spark config when submitting job:spark.celeborn.client.push.replicate.enabled=true -
From 0.3.0 on the default value for
celeborn.worker.storage.workingDiris changed fromhadoop/rss-worker/shuffle_datatoceleborn-worker/shuffle_data, users who want to use origin working dir path should set this configuration. -
Since 0.3.0, configuration namespace
celeborn.ha.masteris deprecated, and will be removed in the future versions. All configurationsceleborn.ha.master.*should migrate toceleborn.master.ha.*. -
Since 0.3.0, environment variables
CELEBORN_MASTER_HOSTandCELEBORN_MASTER_PORTare removed. InsteadCELEBORN_LOCAL_HOSTNAMEworks on both master and worker, which takes high priority than configurations defined in properties file. -
Since 0.3.0, the Celeborn Master URL schema is changed from
rss://toceleborn://, for users who start Worker bysbin/start-worker.sh rss://<master-host>:<master-port>, should migrate tosbin/start-worker.sh celeborn://<master-host>:<master-port>. -
Since 0.3.0, Celeborn supports overriding Hadoop configuration(
core-site.xml,hdfs-site.xml, etc.) from Celeborn configuration with the additional prefixceleborn.hadoop.. On Spark client side, user should set Hadoop configuration likespark.celeborn.hadoop.foo=bar, note thatspark.hadoop.foo=bardoes not take effect; on Flink client and Celeborn Master/Worker side, user should set likeceleborn.hadoop.foo=bar. -
Since 0.3.0, Celeborn master metrics
BlacklistedWorkerCountis renamed asExcludedWorkerCount. -
Since 0.3.0, Celeborn master http request url
/blacklistedWorkersis renamed as/excludedWorkers. -
Since 0.3.0, introduces a terminology update for Celeborn worker data replication, replacing the previous
master/slaveterminology withprimary/replica. In alignment with this change, corresponding metrics keywords have been adjusted. The following table presents a comprehensive overview of the changes:Key Before v0.3.0 Key After v0.3.0 MasterPushDataTimePrimaryPushDataTimeMasterPushDataHandshakeTimePrimaryPushDataHandshakeTimeMasterRegionStartTimePrimaryRegionStartTimeMasterRegionFinishTimePrimaryRegionFinishTimeSlavePushDataTimeReplicaPushDataTimeSlavePushDataHandshakeTimeReplicaPushDataHandshakeTimeSlaveRegionStartTimeReplicaRegionStartTimeSlaveRegionFinishTimeReplicaRegionFinishTime -
Since 0.3.0, Celeborn's spark shuffle manager change from
org.apache.spark.shuffle.celeborn.RssShuffleManagertoorg.apache.spark.shuffle.celeborn.SparkShuffleManager. User can set spark propertyspark.shuffle.managertoorg.apache.spark.shuffle.celeborn.SparkShuffleManagerto use Celeborn remote shuffle service. In 0.3.0, Celeborn still supportorg.apache.spark.shuffle.celeborn.RssShuffleManager, it will be removed in 0.4.0.