CHANGES.txt

1.3.3 - 2015-11-12

  1. Added support for Apache Tajo via pull-request:
     https://github.com/GoogleCloudPlatform/bdutil/pull/67
  2. The YARN capacity scheduler can now be configured when deploying a
     cluster: https://github.com/GoogleCloudPlatform/bdutil/pull/64


1.3.2 - 2015-09-15

  1. Updated Spark configurations to make Cloud Bigtable work with Spark.
  2. Added wrappers bigtable-spark-shell and bigtable-spark-submit to use
     with bigtable plugin; only installed if bigtable_env.sh is used.
  3. Updated default Hadoop 2 version to 2.7.1.
  4. Added support for Apache Hama.
  5. Added support for setting up a standalone NFS cache server for GCS
     consistency using standalone_nfs_cache_env.sh, along with configurable
     GCS_CACHE_MASTER_HOSTNAME to point subsequent clusters at the shared
     NFS cache server. See standalone_nfs_cache_env.sh for usage.
  6. Added explicit check for ordering of imports spark_env.sh relative
     to bigtable_env.sh; Spark must come before Bigtable.
  7. Fixed spelling of "amount" in some documentation.
  8. Fixed directory resolution to bdutil when using symlinks.
  9. Added Dockerfile for bdutil.
  10. Updated default Spark version to 1.5.0; for Spark 1.5.0+, core-site.xml
      will also set 'fs.gs.reported.permissions' to 733, otherwise the Hive
      1.2.1 will error out when using SparkSQL. Hadoop MapReduce will print
      a harmless warning in this case, but otherwise works fine. Additionally,
      Spark auto-restart configuration now contains logic to use correct
      syntax for start-slave.sh depending on whether it's Spark 1.4+, and
      Spark auto-restarted daemons now correctly run under user 'hadoop'.


1.3.1 - 2015-07-09

  1. Added plugin for deploying MapR under platforms/mapr/mapr_env.sh; see
     platforms/mapr/README.md for details.
  2. Changed mapreduce.fileoutputcommitter.algorithm.version to "2"; this should
     only have an effect when running with Hadoop 2.7+, where it significantly
     speeds up job-commit time when using the GCS connector.
     See https://issues.apache.org/jira/browse/MAPEDUCE-4815 for more details.
  3. Added an option ENABLE_STORM_BIGTABLE to extensions/storm/storm_env.sh to
     set up using Google Cloud Bigtable from Apache Storm.
  4. Updated Flink version to 0.9.0.
  5. Switched from using SPARK_CLASSPATH to using SPARK_DIST_CLASSPATH pointed
     at the Hadoop classpath to inherit gcs-connector and other Hadoop libraries
     on the default Spark classpath. This gets rid of a warning message about
     SPARK_CLASSPATH deprecation when running Spark, and improves access to
     related Hadoop libraries from Spark jobs.
  6. Fixed reboot recovery for single-node clusters; this includes the ability
     for single-node clusters to recover from issuing "Stop" and then "Start"
     commands via the GCE API.
  7. Added explicit value for mapreduce.job.working.dir in Ambari config; this
     works around a bug in PigInputFormat where an exception is thrown with
     "Wrong FS scheme" when the default filesystem doesn't have the same scheme
     as the filesystem of the input file(s) (e.g. when reading GCS files and
     the default FS is HDFS). Pig reading from GCS should now work in ambari
     bdutil deployments.
  8. Fixed a bug where Hive deployed under ambari_env.sh is unable to
     LOAD DATA INPATH 'gs://<...>' due to Hive server needing to be restarted
     after GCS connector installation to pick it up on its classpath.


1.3.0 - 2015-05-27

  1. Upgraded default Hadoop 2 version to 2.6.0.
  2. Added support for making a portion of worker VMs run as "preemptible vms"
     by setting --preemptible or PREEMPTIBLE_FRACTION to a value between
     0.0 and 1.0 to specify the fraction of workers to run as preemptible.
  3. Added support for deploying onto ubuntu-12-04 or ubuntu-14-04 images.
  4. Added support for specifying boot disk sizes via --master_boot_disk_size_gb
     and --worker_boot_disk_size_gb or MASTER_BOOT_DISK_SIZE_GB and
     WORKER_BOOT_DISK_SIZE_GB; uses default derived from base image if unset.
  5. Upgraded default Spark version to 1.3.1.
  6. Removed datastore-connector installation options and samples; the connector
     has been deprecated since February 17th, 2015. For alternatives see:
     https://groups.google.com/forum/#!topic/gcp-hadoop-announce/D3_OZuqn4_o
  7. Added workaround for a bug where ambari_env.sh and ambari_manual_env.sh
     would fail to copy mapreduce.tar.gz, pig.tar.gz, etc., into
     hdfs:///hdp/apps/... during setup. Ambari should now work out-of-the-box.


1.2.1 - 2015-05-05

  1. Install Java JDK with Spark; this allows spark-sql to correctly run
     out-of-the-box.
  2. New --master_machine_type/-M flag for setting a different master
     machine type vs worker machine type.
  3. Updated default Spark version to 1.3.0; SparkSQL scripts may need
     modifications to use the new DataFrames; see Spark's migration guide:
     http://spark.apache.org/docs/1.3.0/sql-programming-guide.html#migration-guide
  4. Fixed CentOS 7 support.
  5. Added basic support for using local SSDs via --master_local_ssd_count
     and --worker_local_ssd_count.
  6. Removed default zone, trying to get default zone from gcloud instead or
     otherwise requiring an explicit zone setting.
  7. Fixed JobHistory permissions on HDFS.
  8. Added new extension: extensions/bigtable/bigtable_env.sh for deploying
     a cluster with the HBase-compatible connector for Google Cloud Bigtable
     installed.


1.2.0 - 2015-02-26

  1. Fixed reboots on CentOS 7.
  2. Fixed Ambari-plugin support of reusing persistent disks across deployments.
  3. Added support for Apache Flink.
  4. Made all UPLOAD_FILES be relative to the bdutil directory.
  5. Added basic HBase, CDH, and Storm plugins for bdutil.
         ./bdutil -e hbase
         ./bdutil -e cdh
         ./bdutil -e storm
  6. Only symlink the GCS connector for Client nodes.
  7. Fixed memory allocation for Spark executors running on YARN; Created
     extension extensions/spark/spark_on_yarn_env.sh to support Spark
     on YARN without Spark daemons. the combination of spark_env.sh and
     hadoop2_env.sh will allow the user to submit Spark Jobs to either the
     Spark Master or YARN.
  8. Enabled restart spark processes on reboot.
  9. Added support for the GCS connector with ambari_manual_env.sh. See
     the "Can I deploy HDP manually using Ambari" section in
     platforms/hdp/README.md
  10. Added an experimental env file to enable cluster resizing.
        See extensions/google/experimental/resize_env.sh.
  11. Updated default Spark version to 1.2.1.
  12. Updated Spark driver memory settings to scale with VM size.
  13. Added import_env support to generate_config.  For example:
        "./bdutil -e spark -b my-bucket -p my-project generate_config my_env.sh"
        makes my_env.sh contain "import_env /path/to/spark_env.sh".
  14. Use default mount options to avoid SELinux issues on reboot.


1.1.0 - 2015-01-22

  1. Added plugin for deploying Ambari/HDP with:
         ./bdutil -e platforms/hdp/ambari_env.sh deploy
  2. Set dfs.replication to 2 under conf/hadoop*/hdfs-template.xml; this suits
     PD deployments better than r=3, but if deploying with HDFS residing on
     non-PD storage, the value should be reverted to 3.
  3. Enabled Spark EventLog for Spark deployments, logging to
     gs://${CONFIGBUCKET}/spark-eventlog-base/${MASTER_HOSTNAME}
  4. Migrated off of misc deprecated fields in favor of using
     spark-defaults.conf for Spark 1.0+; cleans up warnings on spark-submit.
  5. Moved SPARK_LOG_DIR from default of ${SPARK_HOME}/logs into
     /hadoop/spark/logs so that they reside on the large PD if it exists.
  6. Upgraded default Spark version to 1.2.0.
  7. Added bdutil_env option INSTALL_JDK_DEVEL to optionally install full JDK
     with compiler/tools instead of just the minimal JRE; set to 'true' in
     single_node_env.sh and ambari_env.sh.
  8. Added python script to allocate memory more intelligently in Hadoop 2.
  9. Upgraded default Hadoop 2 version to 2.5.2.


1.0.1 - 2014-12-16

  1. Replaced usage of deprecated gcutil with gcloud compute.
  2. Changed GCE_SERVICE_ACCCOUNT_SCOPES from a comma separated list to a bash
     array.
  3. Fixed cleanup of pig-validate-setup.sh, hive-validate-setup.sh and
     spark-validate-setup.sh.
  4. Upgraded default Spark version to 1.1.1.
  5. The default zone for instances is now us-central1-a.


0.36.4 - 2014-10-17

  1. Added bdutil flags --worker_attached_pds_size_gb and
     --master_attached_pd_size_gb corresponding to the bdutil_env variables of
     the same names.
  2. Added bdutil_env.sh variables and corresponding flags:
     --worker_attached_pds_type and --master_attached_pd_type to specify
     the type or PD to create, 'pd-standard' or 'pd-ssd'. Default: pd-standard.
  3. Fixed a bug where we forgot to actually add
     extensions/querytools/setup_profiles.sh to the COMMAND_GROUPS under
     extensions/querytools/querytools_env.sh; now it's actually possible to
     run 'pig' or 'hive' directly with querytools_env.sh installed.
  4. Fixed a bug affecting Hadoop 1.2.1 HDFS persistence across deployments
     where dfs.data.dir directories inadvertently had their permissions
     modified to 775 from the correct 755, and thus caused datanodes to fail to
     recover the data. Only applies in the use case of setting:

         CREATE_ATTACHED_PDS_ON_DEPLOY=false
         DELETE_ATTACHED_PDS_ON_DELETE=false

     after an initial deployment to persist HDFS across a delete/deploy command.
     The explicit directory configuration is now set in bdutil_env.sh with
     the variable HDFS_DATA_DIRS_PERM, which is in turn wired into
     hdfs-site.xml.
  5. Added mounted disks to /etc/fstab to re-mount them on boot.
  6. bdutil now uses a search path mechanism to look for env files to reduce
     the amount of typing necessary to specify env files. For each argument
     to the -e (or --env_var_files) command line option, if the argument
     specifies just a base filename without a directory, bdutil will use
     the first file of that name that it finds in the following directories:
       1. The current working directory (.).
       2. Directories specified as a colon-separated list of directories in
          the environment variable BDUTIL_EXTENSIONS_PATH.
       3. The bdutil directory (where the bdutil script is located).
       4. Each of the extensions directories within the bdutil directory.
     If the base filename is not found, it will try appending "_env.sh" to
     the filename and look again in the same set of directories.
     This change allows the following:
       1. You can specify standard extensions succinctly, such as
          "-e spark" for the spark extension, or "-e hadoop2" to use Hadoop 2.
       2. You can put the bdutil directory in your PATH and run bdutil
          from anywhere, and it will still find all its own files.
       3. You can run bdutil from a directory containing your custom env
          files and use filename completion to add them to a bdutil command.
       4. You can collect your custom env files into one directory, set
          BDUTIL_EXTENSIONS_PATH to point to that directory, run bdutil
          from anywhere, and specify your custom env files by name only.
  7. Added new boolean setting to bdutil_env.sh, ENABLE_NFS_GCS_FILE_CACHE,
     which defaults to 'true'. When true, the GCS connector will be configured
     to use its new "FILESYSTEM_BACKED" DirectoryListCache for immediate
     cluster-wide list consistency, allowing multi-stage pipelines in e.g. Pig
     and Hive to safely operate with DEFAULT_FS=gs. With this setting, bdutil
     will install and configure an NFS export point on the master node, to
     be mounted as the shared metadata cache directory for all cluster nodes.
  8. Fixed a bug where the datastore-to-bigquery sample neglected to set a
     'filter' in its query based on its ancestor entities.
  9. YARN local directories are now set to spread IO across all directories
     under /mnt.
  10. YARN container logs will be written to /hadoop/logs/.
  11. The Hadoop 2 MR Job History Server will now be started on the master node.
  12. Added /etc/init.d entries for Hadoop daemons to restart them after
      VM restarts.
  13. Moved "hadoop fs -test" of gcs-connector to end of Hadoop setup, after
      starting Hadoop daemons.
  14. The spark_env.sh extension will now install numpy.


0.35.2 - 2014-09-18

  1. When installing Hadoop 1 and 2, snappy will now be installed and symbolic
     links will be created from the /usr/lib or /usr/lib64 tree to the Hadoop
     native library directory.
  2. When installing Hadoop 2, bdutil will attempt to download and install
     precompiled native libraries for the installed version of Hadoop.
  3. Modified default hadoop-validate-setup.sh to use 10MB of random data
     instead of the old 1MB, otherwise it doesn't work for larger clusters.
  4. Added a health check script in Hadoop 1 to check if Jetty failed to load
     for the TaskTracker as in [MAPREDUCE-4668].
  5. Added ServerAliveInterval and ServerAliveCountMax SSH options to SSH
     invocations to detect dropped connections.
  6. Pig and Hive installation (extensions/querytools/querytools_env.sh) now
     sets DEFAULT_FS='hdfs'; reading from GCS using explicit gs:// URIs will
     still work normally, but intermediate data for multi-stage pipelines will
     now reside on HDFS. This is because Pig and Hive more commonly rely on
     immediate "list consistency" across clients, and thus are more susceptible
     to GCS "eventual list consistency" semantics even if the majority case
     works fine.
  7. Changed occurrences of 'hdpuser' to 'hadoop' in querytools_env.sh, such
     that Pig and Hive will be installed under /home/hadoop instead of
     /home/hdpuser, and the files will be owned by 'hadoop' instead of
     'hdpuser'; this is more consistent with how other extensions have been
     handled.
  8. Modified extensions/querytools/querytools_env.sh to additionally insert
     the Pig and Hive 'bin' directories into the PATH environment variable
     for all users, such that SSH'ing into the master provides immediate
     access to launching 'pig' or 'hive' without requiring
     "sudo sudo -i -u hdpuser"; removed 'chmod 600 hive-site.xml' so that any
     user can successfully run 'hive' directly.
  9. Added extensions/querytools/{hive, pig}-validate-setup.sh which can be
     used as a quick test of Pig/Hive functionality:
     ./bdutil shell < extensions/querytools/pig-validate-setup.sh
  10. Updated extensions/spark/spark_env.sh to now use spark-1.1.0 by default.
  11. Added new BigQuery connector sample under bdutil-0.35.2/samples as file
      streaming_word_count.sh which demonstrates using the new support for
      the older "hadoop.mapred.*" interfaces via hadoop-streaming.jar.


0.35.1 - 2014-08-07

  1. Added a boolean bdutil option DEBUG_MODE with corresponding flags
     -D/--debug which turns on high-verbosity modes for gcutil and gsutil
     calls during the deployment, including on the remote VMs.
  2. Added the ability for the Google connectors, bdconfig, and Hadoop
     distributions to be stored and fetched from gs:// style URLs in addition
     to http:// URLs.
  3. In VERBOSE_MODE, on failure the detailed debuginfo.txt is now also printed
     to the console in addition to being available in the /tmp directory.
  4. Moved all configuration code in conf/.
  5. Changed the default PREFIX to 'hadoop' instead of 'hs-ghfs', and the
     naming convention for masters/workers to follow $PREFIX-m and $PREFIX-w-$i
     instead of $PREFIX-nn and $PREFIX-dn-$i.  IMPORTANT: This breaks
     compatibility with existing clusters deployed with bdutil 0.34.x and older,
     but there is a new flag "--old_hostname_suffixes" to continue using the old
     -nn/-dn naming convention. For example, to turn
     down an old cluster if you've been using the default prefix:
     ./bdutil --prefix=hs-ghfs --old_hostname_suffixes delete
  6. Fixed a bug in VM environments where run_command could not find commands
     such as 'hadoop' in their PATH.
  7. Update BigQuery / Datastore sample scripts to be used with
     "./bdutil run_command." rather than locally.
  8. Added a test to guarantee VMs had no more than 64 characters in their fully
     qualified domain names.
  9. Added the import_env helper to allow "_env.sh" files to depend on each
     other.
  10. Renamed spark1_env.sh to spark_env.sh.
  11. Added a gsutil update check upon first entering a VM.


0.34.4 - 2014-06-23

  1. Switched default gcs-connector version to 1.2.7 for patch fixing a bug
     where globs wrongly reported "not found" in some cases in Hadoop 2.2.0.


0.34.3 - 2014-06-13

  1. Jobtracker / Resource manager recovery has been enabled by default to
     preserve job queues if the daemon dies.
  2. Fixed single_node_env.sh to work with hadoop2_env.sh
  3. Two new commands were added to bdutil: socksproxy and shell; socksproxy
     will establish a SOCKS proxy to the cluster and shell will start an SSH
     session to the namenode.
  4. A new variable, GCE_NETWORK, was added to bdutil_env.sh and can be set
     from the command line via the --network flag when deploying a cluster or
     generating a configuration file. The network specified by GCE_NETWORK
     must exist and must allow SSH connections from the host running bdutil
     and must allow intra-cluster communication.
  5. Increased configured heap sizes of the master daemons (JobTracker,
     NameNode, SecondaryNameNode, and ResourceManager).
  6. The HADOOP_LOG_DIR is now /hadoop/logs instead of the default
     /home/hadoop/hadoop-install/logs; if using attached PDs for larger disk
     storage, this directory resides on that attached PD rather than the
     boot volume, so that Hadoop logs will no longer fill up the boot disk.
  7. Added new extensions under bdutil-<version>/extensions/spark; includes
     spark_shark_env.sh and spark1_env.sh, both compatible for mixing with
     Hadoop2 as well. For now, doesn't use Mesos or YARN in either case,
     but suitable for single-user or Spark-only setups. The spark_shark_env.sh
     extension installs Spark + Shark 0.9.1, while spark1_env.sh only installs
     Spark 1.0.0, in which case Spark SQL serves as the alternative to Shark.
  8. Cleaned updating of login scripts and $PATHs. Added safety check around
     sourcing of hadoop-config.sh, because it can kill shells by calling exit in
     Hadoop 2.


0.34.2 - 2014-06-05

  1. When using Hadoop 2 / YARN, and the default filesystem is set to 'gs', YARN
     log aggregation will be enabled and YARN application logs, including
     map-reduce task logs will be persisted to gs://<CONFIGBUCKET>/yarn-logs/.


0.34.1 - 2014-05-12

  1. Fixed a bug in the USE_ATTACHED_PDS feature (also enabled with
     -d/--use_attached_pds) where disks didn't get attached properly.


0.34.0 - 2014-05-08

  1. Changed sample applications and tools to use GenericOptionsParser instead
     of creating a new Configuration object directly.
  2. Added printout of bdutil version number alongside "usage" message.
  3. Added sleeps between async invocations of GCE API calls during deployment,
     configurable with: GCUTIL_SLEEP_TIME_BETWEEN_ASYNC_CALLS_SECONDS
  4. Added tee'ing of client-side console output into debuginfo.txt with better
     delineation of where the error is likely to have occurred.
  5. Just for extensions/querytools/querytools_env.sh, added an explicit
     mapred.working.dir to fix a bug where PigInputFormat crashes whenever the
     default FileSystem is different from the input FileSystem. This fix allows
     using GCS input paths in Pig with DEFAULT_FS='hdfs'.
  6. Added a retry-loop around "apt-get -y -qq update" since it may flake under
     high load.
  7. Significantly refactored bdutil into better-isolated helper functions, and
     added basic support for command-line flags and several new commnds. The old
     command "./bdutil env1.sh env2.sh" is now "./bdutil -e env1.sh,env2.sh".
     Type ./bdutil --help for an overview of all the new functionality.
  8. Added better checking of env and upload files before starting deployment.
  9. Reorganized bdutil_env.sh into logical sections with better descriptions.
  10. Significantly reduced amount of console output; printed dots indicate
     progress of async subprocesses. Controllable with VERBOSE_MODE or '-v'.
  11. Script and file dependencies are now staged through GCS rather than using
     gcutil push; drastically decreases bandwidth and improves scalability.
  12. Added MAX_CONCURRENT_ASYNC_PROCESSES to splitting the async loops into
     multiple smaller batches, to avoid OOMing.
  13. Made delete_cluster continue on error, still reporting a warning at the
     end if errors were encountered. This way, previously-failed cluster
     creations or deletions with partial resources still present can be
     cleaned up by retrying the "delete" command.


0.33.1 - 2014-04-09

  1. Added deployment scripts for the BigQuery and Datastore connectors.
  2. Added sample jarfiles for the BigQuery and Datastore connectors under
     a new /samples/ subdirectory along with scripts for running the samples.
  3. Set the default image type to backports-debian-7 for improved networking.


0.33.0 - 2014-03-21

  1. Renamed 'ghadoop' to 'bdutil' and ghadoop_env.sh to bdutil_env.sh.
  2. Bundled a collection of *-site.xml.template files in conf/ subdirectory
     which are integrated into the hadoop conf/ files in the remote scripts.
  3. Switched core-site template to new 'fs.gs.auth.*' syntax for
     enabling service-account auth.


0.32.0 - 2014-02-12

  1. ghadoop now always includes ghadoop_env.sh; only the overrides file needs
     to be specified, e.g. ghadoop deploy single_node_env.sh.
  2. Files in COMMAND_GROUPS are now relative to the directory in which ghadoop
     resides, rather than having to be inside libexec/. Absolute paths are
     also supported now.
  3. Added UPLOAD_FILES to ghadoop_env.sh which ghadoop will use to upload
     a list of relative or absolute file paths to every VM before starting
     execution of COMMAND_STEPS.
  4. Include full Hive and Pig sampleapp from Cloud Solutions with ghadoop;
     added extensions/querytools/querytools_env.sh to auto-install Hive and
     Pig as part of deployment. Usage:
         ./ghadoop deploy extensions/querytools/querytools_env.sh


0.31.1 - 2014-01-23

  1. Added CHANGES.txt for release notes.
  2. Switched from /hadoop/temp to /hadoop/tmp.
  3. Added support for running ghadoop as root; will display additional
     confirmation prompt before starting.
  4. run_gcutil_cmd() now displays the full copy/paste-able gcutil command.
  5. Now, only the public key of the master-generated ssh keypair is copied
     into GCS during setup and onto the datanodes. This fixes the occasional
     failed deployment due to GCS list-consistency, and is cleaner anyways.
     The ssh keypair is now more descriptively named: 'hadoop_master_id_rsa'.
  6. Added check for sshability from master to workers in start_hadoop.sh.
  7. Cleaned up internal gcutil commands, added printout of full command
     to ssh into the namenode at the end of the deployment.
  8. Removed indirect config references from *-site.xml.
  9. Stopped explicitly setting mapred.*.dir.


0.31.0 - 2014-01-14

  1. Preview release of ghadoop.