Revision a7061474f905fe60a9581dbf4e14431d00b130a2 authored by Felix GV on 12 December 2015, 01:26:55 UTC, committed by Felix GV on 15 December 2015, 02:45:06 UTC
Summary: This new mode provides the capability of pushing to multiple clusters with different number of nodes and different partition assignments. Compatibility: Although this new mode only works if both the BnP job and the Voldemort servers are upgraded, the change can be rolled out gradually without breaking anything. There is a negotiation phase at the beginning of the BnP job which determines if all servers of all clusters are capable and willing (i.e.: configured) of using the new mode. If not all servers are upgraded and enabled, then the BnP job falls back to its old behavior. Likewise, if a server gets a fetch request from a non-upgraded BnP job, it will work just like before. By default, servers answer the negotiation by saying they support the new mode. The old behavior can be forced with the following server-side configuration: readonly.build.primary.replicas.only=false Running in this new mode has several implications: 1. When running in the new mode, store files are stored in the BnP output directory under nested partition directories, rather than in nested node directories. 2. The MR job uses half as many reducers and half as much shuffle bandwidth compared to before. 3. The meta checksum is now done per partition, rather than per node. 4. Instead of having one .metadata file per partition, there is now only a single full-store.metadata file at the root of the output directory. 5. The server-side HdfsFetcher code inspects the metadata file and determines if it should operate in 'build.primary.replicas.only' mode or not. If yes, then the server determines which partitions it needs to fetch on its own, rather than relying on what the BnP job placed in a node-specific directory. 6. The replica type number contained in Read-Only V2 file names is now useless, but we are keeping it in there just to avoid unnecessary changes. 7. When initializing a Read-Only V2 store directory, the server now looks for files named with the incorrect replica type, and if it finds any, it renames them to the replica type expected by this server. Other changes: 1. Added socket port to Node's toString functions. Also made the output of the Node's toString(), briefToString() and getStateString() functions more consistent. 2. Introduced new Protobuf message for the GetConfig admin request. This new message is intended to be a generic way to retrieve any of server config. 3. Refactored VoldemortConfig to provide access to any config by its string key. Also cleaned up a lot of hard-coded strings, which are constants now. 4. Various minor refactorings in BnP code.
1 parent 69fcd3f
run-bnp.sh
#!/bin/bash
base_dir=$(cd $(dirname $0)/.. && pwd)
HADOOP_CONFIG_PATH_DEFAULT=/etc/hadoop/conf/
if [[ $# < 1 ]]; then
echo "Usage: $0 config_file [hadoop_config_path]"
echo ""
echo "config_file : The file containing necessary config values, in key=value format."
echo "hadoop_config_path : The path for *-site.xml files. Defaults to $HADOOP_CONFIG_PATH_DEFAULT"
exit 1
fi
CONFIG_FILE=$1
if [[ $# > 1 ]]; then
HADOOP_CONFIG_PATH=$2
else
HADOOP_CONFIG_PATH=$HADOOP_CONFIG_PATH_DEFAULT
fi
VERSION=`grep 'curr.release' $base_dir/gradle.properties | sed -e 's/curr.release=\(.*\)/\1/'`
echo "Voldemort version detected: $VERSION"
echo "Executing BnP with:"
echo "config_file : $CONFIG_FILE"
echo "hadoop_config_path : $HADOOP_CONFIG_PATH"
# The jar file's name depends on the directory name...
export VOLDEMORT_JAR="$(echo $base_dir/build/libs/*-$VERSION-bnp.jar)"
export HADOOP_CLASSPATH=$HADOOP_CONFIG_PATH
hadoop jar $VOLDEMORT_JAR voldemort.store.readonly.mr.azkaban.VoldemortBuildAndPushJobRunner $CONFIG_FILE
echo "BnP run script finished!"
Computing file changes ...