Skip to content

ExpediaGroup/circus-train

Repository files navigation

Circus Train.

Start using

You can obtain Circus Train from Maven Central:

Maven Central build Coverage Status GitHub license

Overview

Circus Train replicates Hive tables between clusters on request. It replicates both the table's data and metadata. Unlike many other solutions it has a light touch, requiring no direct integration with Hive's core services. It can copy either entire unpartitioned tables or user defined sets of partitions on partitioned tables. A more detailed overview and the background of this project can be found in this blog post: Replicating big datasets in the cloud.

Other features include:

  • Replication of any of the following:
    • an entire table and all its data on each run.
    • a subset of the data represented by a filter query (e.g. for a data set partitioned by load date you could run Circus Train once a day and replicate only the previous day's data).
    • a subset of the data detected by looking for any changes in the data.
    • views.
    • table metadata (i.e. automatically replicates any DDL changes).
  • Replications between any Hadoop compatible file systems (HDFS, S3, GCS).
  • A snapshot can be taken of the source data so any changes to it while replication occurs don't result in errors or incomplete data being copied.
  • Data is replicated to a unique location at the destination so any overwriting of existing data is transparent to anyone querying the data at the same time (i.e. end users won't see inconsistent or incomplete data or receive errors). Deletion of orphaned overwritten data can be handled by a configurable housekeeping process which removes this data after an interval (to protect any "in flight" queries on the older data from errors).
  • Hive database and tables can optionally be renamed to have different names in the destination.
  • Control of how the data is copied - what file attributes are copied, bandwidth throttling etc.
  • Data consistency checking is performed as part of the replication.
  • Numerous retry mechanisms are used in order to handle the many connectivity and reliability issues that can occur when replicating across locations.
  • Transport encryption for both metadata and data.
  • Use of a Java Keystore Hadoop Credentials Provider to keep authentication keys secure.
  • Connectivity with private networks through a bastion host and direct connections.
  • Table, partition, and column statistics are also copied to preserve performance at the replica.
  • Metrics about replication (success/failure, bandwidth, bytes copied etc.) can be sent to Graphite for building dashboards, alerting etc.
  • Pluggable replication event listeners so users can implement their own functionality based on events occurring during a replication (e.g. triggering a downstream job when a replication completes). A sample listener that uses Amazon's Simple Notification Service (SNS) is provided as a usable reference implementation of this.
  • Pluggable transformations so users can implement their own functionality to transform table, partition and column statistics metadata during a replication.
  • Configuration can be provided using YAML, property files, environment variables and command line arguments.

Related projects

Circus Train has been architected in a modular fashion to allow it to be more easily integrated with existing systems and to have additional functionality added to it. Here are some projects which are outside of the core Circus Train but provide useful additional features:

General operation

Below is a high level summary of the steps that Circus Train performs during the course of a typical run (different configuration might change this).

  1. Fetch the configured entities to replicate from the source Hive metastore.
  2. Extract the data locations for these entities.
  3. Create an HDFS snapshot of the source data if possible - if the source data is in HDFS.
  4. Create a new and unique destination folder on the replica.
  5. Run a distributed copy operation on the local cluster to copy the data from the source snapshot to the new replica folder. For S3 to S3 operation the copy is delegated to the AWS backend.
  6. Delete the snapshot on the source - if a snapshot was created.
  7. Transform the source Hive entities to make them relevant to the destination:
    1. Set various table/partition parameters with replication information.
    2. Modify the data storage locations to those in the new replica folder.
  8. Create any entities in the replica metastore if they don't already exist, otherwise alter the existing entities.
  9. Schedule the deletion of any replica data that has now been replaced with new data.

Installation

Download the TGZ from Maven central and then uncompress the file by executing:

tar -xzf circus-train-<version>-bin.tgz

Although it's not necessary, we recommend exporting the environment variable CIRCUS_TRAIN_HOME by setting its value to wherever you extracted it to:

export CIRCUS_TRAIN_HOME=/<foo>/<bar>/circus-train-<version>

Refer to the configuration section below on what is needed to customise the configuration files before continuing.

Usage

To run Circus Train you just need to execute the bin/circus-train.sh script in the installation directory and pass the configuration file which includes the replication configurations:

    $CIRCUS_TRAIN_HOME/bin/circus-train.sh --config=/path/to/config/file.yml

EMR

If you are planning to run Circus Train on EMR you will need to set up the EMR classpath by exporting the following environment variables before calling the bin/circus-train.sh script:

    export HCAT_LIB=/usr/lib/hive-hcatalog/share/hcatalog/
    export HIVE_LIB=/usr/lib/hive/lib/

Note that the paths above are correct as of when this document was last updated but may differ across EMR versions, refer to the EMR release guide for more up to date information if necessary.

Note: When replicating data to a Hadoop cluster you can set the replication factor. Circus Train uses the dfs.replication setting from the cluster that is running Circus Train. On EMR this replication factor is set to 1 by AWS defaults. The setting can be overridden by changing the dfs.replication value on the cluster that you run on. The setting can be found in /etc/hadoop/conf/hdfs-site.xml. Alternatively it can be set in the Circus Train configuration file like so:

source-catalog:
  configuration-properties:
    dfs.replication: 3

So if you are running on EMR and replicating into an on-premises cluster this is something you want to set.

Exit codes

Circus Train returns the following exit codes:

  • 0 if all configured replications succeeded.
  • -1 if all the configured replications failed.
  • -2 if at least one of the configured replications succeeded but one or more of the other replications failed.

Logging

Circus Train uses Log4J for logging, to configure it to use a custom Log4J file add the path to the logging configuration file to the Circus Train configuration:

logging:
    config: file:/home/foo/circus-train/log4j.xml

Configuring replication and housekeeping

By default when Circus Train runs it will first perform all configured replications and then it will perform an additional "housekeeping" step where it removes any dereferenced data in the replica file system. Depending on the nature of your replications this housekeeping might take a while to run. It is possible to schedule a separate (e.g. daily) run to perform this and configure your other Circus Train runs to perform replication only. The above two scripts support this via a --modules parameter that can be set to one of the following values:

  • replication, housekeeping: This is the default value where Circus Train will first perform any configured table replications followed by the housekeeping process.
  • replication: Circus Train will perform table replications only. The housekeeping process will not be executed.
  • housekeeping: Circus Train will perform housekeeping only. Table replications will not be executed.

For example, the following command sequence will execute Circus Train twice, first performing replication and then housekeeping:

    $CIRCUS_TRAIN_HOME/bin/circus-train.sh --config=/path/to/config/file.yml --modules=replication
    $CIRCUS_TRAIN_HOME/bin/circus-train.sh --config=/path/to/config/file.yml --modules=housekeeping

The preferred way to execute the housekeeping process is by using the housekeeping.sh script which is described below.

Run housekeeping only

If you want to schedule housekeeping as a separate process then you should use the following script:

    $CIRCUS_TRAIN_HOME/bin/housekeeping.sh --config=/path/to/config/file.yml

Orphaned data strategy

To switch off housekeeping completely, and not put any paths to dereferenced data in the housekeeping database, orphaned-data-strategy can be set to NONE. This property is set to HOUSEKEEPING by default.

Naturally, to substitute housekeeping, Circus Train supports the addition of properties to newly created tables so that you can hook into Hive events to monitor and delete dereferenced data as appropriate. One such tool, Beekeeper, does exactly this. All that is required if Beekeeper is monitoring your Hive Metastore's events is the addition of Beekeeper-specific properties to your configuration (see Metadata transformations), and Beekeeper will handle the rest.

If your table already exists, you will need to add these properties to your replica table manually, as well as your Circus Train configuration.

Configuration Reference

Circus Train uses Spring Boot for configuration so you are free to use any of the many configuration strategies supported by this framework to configure your Circus Train instance.

Example configurations

The examples below all demonstrate configuration using YAML and provide fragments covering the most common use cases that should be useful as a basis for building your own configuration. A full configuration reference is provided in the following sections.

Configuring source and replica

The YAML fragment below shows some common options for setting up the base source (where data is coming from) and replica (where data is going to).

    source-catalog:
      name: on-premises-cluster
      disable-snapshots: false
    replica-catalog:
      name: aws-data-warehouse
      hive-metastore-uris: thrift://emr-master.compute.amazonaws.com:9083
      metastore-tunnel:
        route: ec2-user@bastion-host -> hadoop@emr-master
        private-keys: /home/user/.ssh/bastion-key-pair.pem,/home/user/.ssh/emr-key-pair.pem
        known-hosts: /home/user/.ssh/known_hosts
    security:
      credential-provider: jceks://hdfs[@namenode:port]/hcom/etl/<team>/conf/aws-<account>.jceks
    copier-options:
    table-replications:
      - ...

Replicate an unpartitioned table

The YAML fragment below shows some common options for setting up the replication of an unpartitioned table where the entire table will be replicated on each run.

    table-replications:
      - source-table:
          database-name: test_database
          table-name: test_unpartitioned_table
        replica-table:
          database-name: test_database
          table-name: test_unpartitioned_table
          table-location: s3://bucket/table/path

Replicate a partitioned table (with an errant table.location)

The YAML fragment below shows some common options for setting up the replication of a partitioned table where the table location in Hive is incorrect (and is thus being overridden to be the actual base location on HDFS) and also only the last 7 days worth of data is replicated on each run.

    table-replications:
      - source-table:
          database-name: test_database
          table-name: test_partitioned_table
          table-location: /source/my_folder/test_database/test_table_partitioned
          partition-filter: "local_date >= #{#nowEuropeLondon().minusDays(7).toString('yyyy-MM-dd')}"
          partition-limit: 3
        replica-table:
          table-location: s3://bucket/table/path

Replicate a view

The YAML fragment below shows some common options for setting up the replication of a view which uses a table that has already been replicated to the target metastore.

    table-replications:
      - replication-mode: METADATA_MIRROR
        table-mappings:
          test_database.test_unpartitioned_table: replica_db.replica_test_unpartitioned_table
        source-table:
          database-name: test_database
          table-name: test_view
        replica-table:
          database-name: replica_db
          table-name: replica_test_view

Replication configuration reference

The table below describes all the available configuration values for Circus Train.

Property Required Description
source-catalog.name Yes A name for the source catalog for events and logging.
source-catalog.disable-snapshots No Controls whether HDFS snapshots will be used on the source data locations when replicating data. If the source location is HDFS and has been made "snapshottable" we recommend setting this to false. Default is false it will log a warning if the folder is not "snapshottable" and continue.
source-catalog.hive-metastore-uris No Fully qualified URI of the source cluster's Hive metastore Thrift service. If not specified values are taken from the hive-site.xml on the Hadoop classpath of the machine that's running Circus Train. This property mimics the Hive property "hive.metastore.uris" and allows multiple comma separated URIs.
source-catalog.site-xml No A list of Hadoop configuration XML files to add to the configuration for the source.
source-catalog.configuration-properties No A list of key: value pairs to add to the Hadoop configuration for the source.
source-catalog.metastore-tunnel.* No See metastore tunnel configuration values below.
replica-catalog.name Yes A name for the replica catalog for events and logging.
replica-catalog.hive-metastore-uris Yes Fully qualified URI of the replica cluster's Hive metastore Thrift service. On AWS this usually comprises of the EMR master node public hostname and metastore thrift port. This property mimics the Hive property "hive.metastore.uris" and allows multiple comma separated URIs.
replica-catalog.site-xml No A list of Hadoop configuration XML files to add to the configuration for the replica.
replica-catalog.configuration-properties No A list of key:value pairs to add to the Hadoop configuration for the replica.
replica-catalog.metastore-tunnel.* No See metastore tunnel configuration values below.
security.credential-provider No URL(s) to the Java Keystore Hadoop Credential Provider(s) that contain the S3 access.key and secret.key for the source or destination S3 buckets.
copier-options No Globally applied Copier options. See Copier options for details.
table-replications[n].source-table.database-name Yes The name of the database in which the table you wish to replicate is located.
table-replications[n].source-table.table-name Yes The name of the table which you wish to replicate.
table-replications[n].source-table.table-location No The base path of the table (fully qualified URI). Required only if your table is partitioned, external, and has its location set to a path different to that of the base path of its partitions.
table-replications[n].source-table.partition-filter No A filter to select which partitions to replicate. Used for partitioned tables only. See Partition filters for more information.
table-replications[n].source-table.generate-partition-filter No Set to true to enable the "Hive Diff" feature. See Partition filter generation for details. Default is false. If true the table-replications[n].source-table.partition-filter will be ignored and instead a generated filter will be used.
table-replications[n].source-table.partition-limit No A limit on the number of partitions that will be replicated. Used for partitioned tables only.
table-replications[n].replica-table.table-location Yes The base path of the replica table (fully qualified URI). Please note this is a required parameter only if the replication-mode is FULL
table-replications[n].replica-table.database-name No The name of the destination database in which to replicate the table. Defaults to source database name.
table-replications[n].replica-table.table-name No The name of the table at the destination. Defaults to source table name.
table-replications[n].copier-options No Table specific Copier options which override any global options. See Copier options for details.
table-replications[n].orphaned-data-strategy No Orphaned data strategy for replication. See Orphaned data strategy for more information. Defaults to HOUSEKEEPING.
table-replications[n].replication-mode No Table replication mode. See Replication Mode for more information. Defaults to FULL.
table-replications[n].replication-strategy No Table replication strategy. See Replication Strategy for more information. Defaults to UPSERT.
table-replications[n].transform-options No Map of optional options that can be used to set configuration for a custom transformation per table replication.
table-replications[n].table-mappings No Only used by view replications. This is a map of source tables used by the view and their equivalent name in the replica metastore.

The table below describes the tunnel configuration values for source/replica catalog:

Property Required Description
*.metastore-tunnel.route No A SSH tunnel can be used to connect to source/replica metastores. The tunnel may consist of one or more hops which must be declared in this property. See Configuring a SSH tunnel for details.
*.metastore-tunnel.private-keys No A comma-separated list of paths to any SSH keys required in order to set up the SSH tunnel.
*.metastore-tunnel.known-hosts No Path to a known hosts file.
*.metastore-tunnel.port No The port on which SSH runs on the replica master node. Default is 22.
*.metastore-tunnel.localhost No The address on which to bind the local end of the tunnel. Default is 'localhost'.
*.metastore-tunnel.timeout No The SSH session timeout in milliseconds, 0 means no timeout. Default is 60000 milliseconds, i.e. 1 minute.
*.metastore-tunnel.strict-host-key-checking No Whether the SSH tunnel should be created with strict host key checking. Can be set to yes or no. The default is yes.

Partition filters

Control over which data and metadata Circus Train replicates can be achieved through the use of partition filters which return a list of partitions matching the filter and only data in these partitions will be replicated. This is useful if you have a very large table where only a few partitions change every day. Instead of replicating the entire table on each run of Circus Train you could create a partition filter which matches the changed partitions and only these would then be replicated.

Partition filters can be specified using the syntax described in HIVE-1609 which states:

The filter supports "=", "!=", ">", "<", ">=", "<=" and "LIKE" operations on partition keys of type string. "AND" and "OR" logical operations are supported in the filter. So for example, for a table having partition keys country and state, the filter can be 'country = "USA" AND (state = "CA" OR state = "AZ")'

In particular notice that it is possible to nest sub-expressions within parentheses. Wildcards are allowed in the form of .*, for example: year like '201.*'. Full regular expressions don't seem to be supported by Hive.

Encoding a constant literal expression into a replication configuration is not especially useful unless you always want to copy the same partitions each time. Therefore it is possible to introduce dynamically evaluated elements using Spring SpELs. We refer you to the SpEL documentation for the specific syntax and rules, however they quite simply allow you to embed Java language fragments into the filter which are dynamically evaluated to a string value. As most partition filters are date based, Circus Train includes the Joda Time API on its classpath and we've also statically imported some convenience methods into the SpEL root context:

  • DateTime nowUtc()
  • DateTime nowEuropeLondon()
  • DateTime nowAmericaLosAngeles()
  • DateTime nowInZone(String zone)
  • DateTime nowInZone(DateTimeZone zone)
  • String zeroPadLeft(int value, int width)
  • String zeroPadLeft(String value, int width)

It's also worth stating again that this is Java, encoded in a SpEL, contained within YAML. Therefore due consideration is needed with respect to the quoting, commenting, and escaping rules of all of these syntaxes. In particular bear in mind that the character # is heavily overloaded. Its uses include (but are not limited to): initiating a SpEL, beginning a YAML comment, and referencing the root context within a SpEL. We have also found some sensitivity to the space character; #{#nowUtc()... is parsed correctly whereas #{ #nowUtc()... fails. We believe this occurs because the YAML parser is keen to interpret the character sequence of a space followed by a hash as a comment, wherever it may be.

An example partition filter entry to match partitions that have been added in the last three days for a data set partitioned by a field called "local_date" might look something like:

partition-filter: local_date >= '#{#nowEuropeLondon().minusDays(3).toString("yyyy-MM-dd")}'

By default characters in the property value are treated simply as string literals and this is the case with the character sequence: local_date >= . In this example we use single quotes to ensure that the first #{ is interpreted as a SpEL prefix and not a YAML comment. The second # is a reference to the SpEL root context from which we can access the nowEuropeLondon convenience method. The rest is just regular Joda API in Java. Once evaluated this expression yields a result such as:

local_date >= '2016-05-13'

You can use the check-filters.sh tool in circus-train-tool to test your expressions against a real metastore without actually invoking any table replications. This is a read-only operation and will not modify any data or metadata:

<circus-train-tool-dir>/bin/check-filters.sh --config=your.yml
Partition filter generation

This is also known as "Hive Diff".

There are certain cases where one may not be able to create a partition filter as described above but there is still a need to limit the amount of data that is transferred with every run to be only a "delta" of what has changed since the last run (as opposed to replicating the entire table). For example one might have a very large table that isn't partitioned by a date or the partitions themselves don't change but the data within them does.

For these situations Circus Train provides the option to automatically generate the filter by comparing and then computing a difference between source and replica - this is referred to as its "Hive Diff" feature. Circus Train will compare the relevant tables and only replicate partitions that are missing or changed in the replica. It does this by looking at table meta data as well as the actual files on both the replica and source. This is a potentially expensive operation as Circus Train needs to fetch all partitions from the metastore and possibly make many calls to the file system so should be used with care. To enable auto generation set the following property: table-replications[n].source-table.generate-partition-filter: true

The properties below are relevant for the filter generation:

Property Required Description
table-replications[n].source-table.generate-partition-filter No Set this to true to automatically generate partition filters based on changed data/metadata. The default is false.
table-replications[n].partition-iterator-batch-size No Number of partition objects that will be stored in memory from the source table. The default is 1000.
table-replications[n].partition-fetcher-buffer-size No Number of partition objects that will be stored in memory from the replica table. The default is 1000.
table-replications[n].source-table.partition-limit No Number of partitions that will be replicated. Used for partitioned tables only. When used in conjunction with a generated partition filter this also limits the generated partitions.

Replication Mode

Circus Train provides configurable replication modes which can be used to control whether to replicate data and metadata or just metadata (and if just metadata how this is replicated). This is configured using via the table-replication setting which can have an optional property replication-mode.

It has the following options:

  • FULL: Default behaviour, all metadata and data of the source table will be copied to the destination table and destination location. Metadata will be updated to point to the replicated data.
  • FULL_OVERWRITE: All metadata and data of the source table will be replicated in full, overwriting the existing replica. Any existing replica table will be dropped and its data deleted.. If no replica table exists, this mode will behave as FULL.
  • METADATA_MIRROR: Only metadata will be copied (mirrored) from the source to the replica. Replica metadata will not be modified so your source and replica will have the same data location. NOTE: The replica table will be marked as EXTERNAL. This is done to prevent accidental data loss when dropping the replica. For example, this can be used for copying someone else's metadata into your Hive Metastore without copying the data or to replicate a view. You still need to have access to the data in order to query it.
  • METADATA_UPDATE: Metadata only update for a table that was previously fully replicated. No data will be copied but any metadata from the source will be copied and table/partition locations will keep pointing to previously replicated data. Example use case: Update the metadata of a Hive Table (for instance to change the Serde used) without having the overhead of re-replicating all the data.

Restrictions:

  • METADATA_MIRROR the only supported replication mode for Hive views.
  • You cannot METADATA_MIRROR a previously fully replicated table (use METADATA_UPDATE if you want to do that). Similarly you cannot go from METADATA_MIRROR to FULL/METADATA_UPDATE as you then might end up with a table that points both to original and replicated data. Replicate to new a table if METADATA_UPDATE doesn't suffice for your use case. Circus Train will fail with an exception if any of these restrictions are not met.

Replication Strategy

Circus Train provides configurable replication strategies which can be used to control whether "destructive actions" like dropping tables or partitions should be propagated to the replica table.

It has the following options:

  • UPSERT: Default behaviour, data is only added to the replica. If the source tables or partitions are deleted these changes are not propagated to the replica.
  • PROPAGATE_DELETES: Like UPSERT but Circus Train will also propagate deletes from the source to the replica. If a source table is deleted then the replica table will also be deleted. Similarly if there are any partitions in the source table that have been deleted they will also be deleted from the replica table. The deletes apply to both metadata and the underlying data (which is scheduled for deletion using Circus Train's Housekeeping mechanism).

Copier options

Circus Train uses highly configurable means to copy the actual data between clusters. Control over this is provided by "copier options" which allow fine grained configuration of the copier processes. The default values should suffice for most use cases but the below sections describe the various options available as they might be useful in certain situations.

DistCp copier options

If data is being replicated to HDFS then Circus Train will use DistCp to copy the data. The DistCp documentation provides detailed information on the options and their meanings but these are also summarised below.

      copier-options:
        file-attribute: replication, blocksize, user, group, permission, checksumtype, acl, xattr, times
        preserve-raw-xattrs: true
        atomic-commit: false
        atomic-work-path: /foo/bar/work
        blocking: true
        copy-strategy: uniformsize
        filters-file: /foo/bar/filters
        ignore-failures: false
        log-path: /foo/bar/log
        task-bandwidth: 100
        max-maps: 50
        skip-crc: false
        ssl-configuration-file: /foo/bar/ssl-config
        copier-factory-class: com.hotels.bdp.circustrain.distcpcopier.DistCpCopierFactory
        data-manipulator-factory-class: com.hotels.bdp.circustrain.distcpcopier.HdfsDataManipulatorFactory
Property Required Description
copier-options.file-attribute No Controls which file attributes should be preserved in the destination. See org.apache.hadoop.tools.DistCpOptions.FileAttribute for allowed values. For replications to HDFS this defaults to replication, blocksize, user, group, permission, checksumtype, xattr.
copier-options.preserve-raw-xattrs No Flat that indicates that raw X attributes should be preserved in the destination. For replications to HDFS file systems this defaults to to true.
copier-options.atomic-commit No Instructs DistCp whether to copy the source data to a temporary target location, and then move the temporary target to the final-location atomically. Defaults to false.
copier-options.atomic-work-path No Path to store temporary files used in the atomic-commit option.
copier-options.copy-strategy No Which strategy to use when copying the data, valid values are dynamic, uniformsize. By default, uniformsize is used (i.e. map tasks are balanced on the total size of files copied by each map. Similar to legacy.) If dynamic is specified, DynamicInputFormat is used instead.
copier-options.ignore-failures No This option will keep more accurate statistics about the copy than the default case. It also preserves logs from failed copies, which can be valuable for debugging. Finally, a failing map will not cause the job to fail before all splits are attempted. Defaults to false.
copier-options.log-path No Location of the log files generated by the job. Defaults to null which means log files will be written to JobStagingDir/_logs.
copier-options.task-bandwidth No Number of MiB/second that map tasks can consume. Map throttles back its bandwidth consumption during a copy, such that the net bandwidth used tends towards the specified value. Defaults to 100.
copier-options.max-maps No Maximum number of map tasks used to copy files. Defaults to 50.
copier-options.skip-crc No Controls whether CRC computation is skipped. Defaults to false.
copier-options.ssl-configuration-file No Path to the SSL configuration file to use for hftps://. Defaults to null.
copier-options.ignore-missing-partition-folder-errors No Boolean flag, if set to true will ignore errors from DistCp that normally fail the replication. DistCp normally fails when a partition is found in the metadata that is missing on HDFS (Default DistCp behavior). Defaults to false (so replication will fail).
copier-options.copier-factory-class No Controls which copier is used for replication if provided.
copier-options.data-manipulator-factory-class No Controls which data manipulator is used to drop replica table data when using the FULL_OVERWRITE replication mode.
S3MapReduceCp copier options

If data is being replicated from HDFS to S3 then Circus Train will use a customized, improved version of DistCp to copy the data. The options are summarised below.

      copier-options:
        credential-provider:
        task-bandwidth: 100
        storage-class: Standard
        s3-server-side-encryption: true
        region:
        multipart-upload-chunk-size: 5
        multipart-upload-threshold: 16
        max-maps: 20
        num-of-workers-per-map: 20
        copy-strategy: uniformsize
        ignore-failures: false
        log-path:
        copier-factory-class: com.hotels.bdp.circustrain.s3mapreducecpcopier.S3MapReduceCpCopierFactory
        data-manipulator-factory-class: com.hotels.bdp.circustrain.s3mapreducecp.aws.AwsMapreduceDataManipulatorFactory
Property Required Description
copier-options.credential-provider No Path to the JCE key store with the AWS credentials. Defaults to the path specified in security.credential-provider. See Replication configuration reference for details.
copier-options.task-bandwidth No Number of MB/second that Mappers can consume. A Mapper will throttle back its bandwidth consumption during a copy, such that the net bandwidth used tends towards the specified value. No limit by default.
copier-options.storage-class No S3 storage class. See IDs in com.amazonaws.services.s3.model.StorageClass. Defaults to null which means default storage class, i.e. STANDARD.
copier-options.s3-server-side-encryption No Whether to enable server side encryption. Defaults to true.
copier-options.region No AWS Region for the S3 client. Defaults to null which means S3MapReduceCP will interrogate AWS for the target bucket location.
copier-options.multipart-upload-chunk-size No Size of multipart chunks in MB. Defaults to 5.
copier-options.multipart-upload-threshold No Size threshold in MB for Amazon S3 object after which multi-part copy is initiated. Defaults to 16.
copier-options.max-maps No Maximum number of map tasks used to copy files. Defaults to 20.
copier-options.num-of-workers-per-map No Number of upload workers to use for each Mapper. Defaults to 20.
copier-options.copy-strategy No Which strategy to use when copying the data, valid values are dynamic, static (A.K.A. uniformsize.) By default, uniformsize is used (i.e. map tasks are balanced on the total size of files copied by each map.) If dynamic is specified, DynamicInputFormat is used instead.
copier-options.ignore-failures No This option will keep more accurate statistics about the copy than the default case. It also preserves logs from failed copies, which can be valuable for debugging. Finally, a failing map will not cause the job to fail before all splits are attempted. Defaults to false.
copier-options.log-path No Location of the log files generated by the job. Defaults to null which means log files will be written to JobStagingDir/_logs.
copier-options.s3-endpoint-uri No URI of the S3 end-point used by the S3 client. Defaults to null which means the client will select the end-point.
copier-options.upload-retry-count No Maximum number of upload retries. Defaults to 3
copier-options.upload-retry-delay-ms No Milliseconds between upload retries. The actual delay will be computed as delay = attempt * copier-options.upload-retry-delay-ms where attempt is the current retry number. Defaults to 300 ms.
copier-options.upload-buffer-size No Size of the buffer used to upload the stream of data. If the value is 0 the upload will use the value of the HDFS property io.file.buffer.size to configure the buffer. Defaults to 0
copier-options.canned-acl No AWS Canned ACL name. See Access Control List (ACL) Overview for possible values. If not specified S3MapReduceCp will not specify any canned ACL.
copier-options.copier-factory-class No Controls which copier is used for replication if provided.
copier-options.data-manipulator-factory-class No Controls which data manipulator is used to drop replica table data when using the FULL_OVERWRITE replication mode.
copier-options.assume-role No ARN of an IAM role to assume when writing S3 data to the target replica. Useful when the target is in a different AWS account than Circus Train is running in. Note that if JCEKS is also configured, JCEKS credentials will be used instead of assuming a role. If assume-role is not specified, the copier will use instance credentials.
S3 to S3 copier options

If data is being replicated from S3 to S3 then Circus Train will use the AWS S3 API to copy data between S3 buckets. Using the AWS provided APIs no data needs to be downloaded or uploaded to the machine on which Circus Train is running but is copied by AWS internal infrastructure and stays in the AWS network boundaries. Assuming the correct bucket policies are in place cross region and cross account replication is supported. We are using the TransferManager to do the copying and we expose its options via copier-options see the table below. Given the source and target buckets Circus-Train will try to infer the region from them.

Property Required Description
copier-options.s3s3-multipart-copy-threshold-in-bytes No Default value should be OK for most replications. See TransferManagerConfiguration
copier-options.s3s3-multipart-copy-part-size-in-bytes No Default value should be OK for most replications. See TransferManagerConfiguration
copier-options.s3-endpoint-uri No URI of the S3 end-point used by the S3 client. Defaults to null which means the client will select the end-point.
copier-options.s3-server-side-encryption No Whether to enable server side encryption. Defaults to false.
copier-options.canned-acl No AWS Canned ACL name. See Access Control List (ACL) Overview for possible values. If not specified S3S3Copier will not specify any canned ACL.
copier-options.copier-factory-class No Controls which copier is used for replication if provided.
copier-options.data-manipulator-factory-class No Controls which data manipulator is used to drop replica table data when using the FULL_OVERWRITE replication mode.
copier-options.s3s3-retry-max-copy-attempts No Controls the maximum number of attempts if AWS throws an error during copy. Default value is 3.
copier-options.s3s3-max-thread-pool-size No Max number of threads to use for the transferManager thread pool. Default value is 10.
copier-options.assume-role No ARN of an IAM role to assume when writing S3 data to the target S3 location. Useful when the target is in a different AWS account than Circus Train is running in. Note that if JCEKS is also configured, JCEKS credentials will be used instead of assuming a role. If assume-role is not specified, the copier will use instance credentials. The role provided must have read access to the S3 source and write access to the S3 target.
copier-options.assume-role-session-duration-seconds No Amount of time (in seconds) that the AWS SDK should assume the given role for. Default value is 12 hours.

S3 Secret Configuration

When configuring a job for replication to or from S3, the AWS access key and secret key with read/write access to the configured S3 buckets must be supplied. Circus train has a couple of options depending on where you run Circus Train.

  • Running on EMR: We provide a com.hotels.bdp.circustrain.aws.HadoopAWSCredentialProviderChain that uses Jceks (explained below), STS assume role credentials (see copier-options.assume-role) and Instance Profile credentials. If the EMR cluster is running with a role that has read write access to the necessary buckets no configuration is necessary and instance profile credentials will be used.
  • Running on on-premises: It is possible to provide s3 secret configuration with jceks (explained below) or STS assume role credentials (see copier-options.assume-role).

S3 Secret Configuration with Jceks

The AWS access key and secret key can be protected from being exposed in the job's Hadoop configuration, Circus Train expects them to be stored using the Hadoop Credential Provider and the JCEKS URL provided in the Circus Train configuration security.credential-provider property. This property is only required if a specific set of credentials is needed or if Circus Train runs on a non-AWS environment. If it is not set then the credentials of the instance where Circus Train runs will be used - note this scenario is only valid when Circus Train is executed on an AWS environment, i.e. EC2/EMR instance.

To add your existing AWS keys for a new replication job run the following commands as the user that will be executing Circus Train and pass in your keys when prompted:

    hadoop credential create access.key -provider <provider>
    hadoop credential create secret.key -provider <provider>

This needs to be done on a machine where the hadoop binary is present (e.g. a tenant node).

Note:

  • For each command you will be prompted to enter the secret twice.
  • We recommend that the <provider> be of the format jceks://hdfs[@namenode:port]/foo/bar/aws-<account>.jceks.
  • To verify the credentials have been stored, run hadoop credential list -provider <provider>
  • See the Hadoop documentation for further information about the Hadoop Credential Provider.

Configuring a SSH tunnel

Circus Train can be configured to use a SSH tunnel to access a remote Hive metastore in cases where certain network restrictions prevent a direct connection from the machine running Circus Train to the machine running the Thrift Hive metastore service. A SSH tunnel consists of one or more hops or jump-boxes. The connection between each pair of nodes requires a user - which if not specified defaults to the current user - and a private key to establish the SSH connection.

As outlined above the metastore-tunnel property is used to configure Circus train to use a tunnel. The tunnel route expression is described with the following EBNF:

path = path part, {"->", path part} ;
path part = {user, "@"}, hostname ;
user = ? user name ? ;
hostname = ? machine name with or without domain name ? ;

For example, if the Hive metastore runs on the host hive-server-box which can only be reached first via bastion-host and then jump-box then the SSH tunnel route expression will be:

bastion-host -> jump-box -> hive-server-box

If bastion-host is only accessible by user ec2-user, jump-box by user user-a and hive-server-box by user hadoop then the expression above becomes:

ec2-user@bastion-host -> user-a@jump-box -> hadoop@hive-server-box

Once the tunnel is established Circus Train will set up port forwarding from the local machine specified in replica-catalog.metastore-tunnel.localhost to the remote machine specified in replica-catalog.hive-metastore-uris. The last node in the tunnel expression doesn't need to be the Thrift server, the only requirement is that the this last node must be able to communicate with the Thrift service. Sometimes this is not possible due to firewall restrictions so in these cases they must be the same.

Note that all the machines in the tunnel expression must be included in the known_hosts file and the keys required to access each box must be set in replica-catalog.metastore-tunnel.private-keys. For example, if bastion-host is authenticated with bastion.pem and both jump-box and hive-server-box are authenticated with emr.pem then the property must be set asreplica-catalog.metastore-tunnel.private-keys=<path-to-ssh-keys>/bastion.pem, <path-to-ssh-keys>/emr.pem.

The following configuration snippets show a few examples of valid tunnel expressions. These fragments can be included under either source-catalog or replica-catalog depending on the use case.

Simple tunnel to metastore server
    hive-metastore-uris: thrift://metastore.domain:9083
    metastore-tunnel:
      route: user@metastore.domain
      private-keys: /home/user/.ssh/user-key-pair.pem
      known-hosts: /home/user/.ssh/known_hosts
Simple tunnel to cluster node with current user
    hive-metastore-uris: thrift://metastore.domain:9083
    metastore-tunnel:
      route: cluster-node.domain
      private-keys: /home/run-as-user/.ssh/key-pair.pem
      known-hosts: /home/run-as-user/.ssh/known_hosts
Bastion host to cluster node with different users and key-pairs
    hive-metastore-uris: thrift://metastore.domain:9083
    metastore-tunnel:
      route: bastionuser@bastion-host.domain -> user@cluster-node.domain
      private-keys: /home/run-as-user/.ssh/bastionuser-key-pair.pem, /home/run-as-user/.ssh/user-key-pair.pem
      known-hosts: /home/run-as-user/.ssh/known_hosts
Bastion host to cluster node with same user
    hive-metastore-uris: thrift://metastore.domain:9083
    metastore-tunnel:
      route: user@bastion-host.domain -> user@cluster-node.domain
      private-keys: /home/user/.ssh/user-key-pair.pem
      known-hosts: /home/user/.ssh/known_hosts
Bastion host to cluster node with current user
    hive-metastore-uris: thrift://metastore.domain:9083
    metastore-tunnel:
      route: bastion-host.domain -> cluster-node.domain
      private-keys: /home/run-as-user/.ssh/run-as-user-key-pair.pem
      known-hosts: /home/run-as-user/.ssh/known_hosts
Bastion host to metastore via jump-box with different users and key-pairs
    hive-metastore-uris: thrift://metastore.domain:9083
    metastore-tunnel:
      route: bastionuser@bastion-host.domain -> user@jump-box.domain -> hive@metastore.domain
      private-keys: /home/run-as-user/.ssh/bastionuser-key-pair.pem, /home/run-as-user/.ssh/user-key-pair.pem, /home/run-as-user/.ssh/hive-key-pair.pem
      known-hosts: /home/run-as-user/.ssh/known_hosts

Instance Globals

Circus Train declares some global configuration properties with sensible default values that can be overridden. These mainly control where Circus Train stores the data files it needs in order to operate and are described in more detail below.

Property Required Description
instance.home No The path to the home directory of the Circus Train instance. Typically this would be the folder containing the conf and housekeeping data directories. Defaults to the user's home directory (${user.home}.
instance.name No A name to be used for constructing the default housekeeping H2 database location. Defaults to ${source-catalog.name}_${replica-catalog.name}.

Profiles

Circus Train is able to use Spring Boot Profiles to further control configuration by allowing the activation of different sections of configuration for different operating conditions (e.g. environments).

Profiles within files

A YAML file is made up of multiple documents, separated by three hyphens (---). It is therefore possible to put the configuration for multiple environments in a single circus-train.yml file and use profiles within each document to conditionally activate these configuration elements. The example below shows how you could use profiles to control the Hive metastore URI based on whether a dev, qa or prod profile is active:

    source-catalog:
      name: on-premises-cluster
    replica-catalog:
      name: aws-data-warehouse
    table-replications:
      -
        ...
    ---
    spring:
      profiles: dev
    replica-catalog:
      hive-metastore-uris: thrift://<dev-host>.com:9083
    ---
    spring:
      profiles: qa
    replica-catalog:
      hive-metastore-uris: thrift://<qa-host>.com:9083
    ---
    spring:
      profiles: prod
    replica-catalog:
      hive-metastore-uris: thrift://<prod-host>.com:9083

Configuring housekeeping

Housekeeping is the process that removes expired and orphaned data on the replica. Below is a YAML configuration fragment and description of the available configuration properties. Implementation of Housekeeping can be found here

    housekeeping:
      expired-path-duration: P3D
      db-init-script: classpath:/schema.sql
      data-source:
        driver-class-name: org.h2.Driver
        url: jdbc:h2:${housekeeping.h2.database};AUTO_SERVER=TRUE;DB_CLOSE_ON_EXIT=FALSE
        username: user
        password: secret
Property Required Description
housekeeping.h2.database No The location of the h2 housekeeping database. Defaults to ${instance.home}/data/${instance.name}/${housekeeping.schema-name}. Refer to the Instance Globals section for more details.
housekeeping.schema-name No Database schema name to use. Circus Train default: 'circus_train'
housekeeping.db-init-script No Database init script to use. Circus Train comes with schema.sql on the classpath by default to create a circus_train schema, but to use it, it needs to be specified for this parameter.

For more details on Housekeeping configuration, including examples on how to override the H2 default database with something a bit more robust, please consult the Housekeeping documentation.

Metric Reporting

Circus Train can be configured to output metrics to "standard out" as well as to Graphite. Most metrics generated by Circus Train are reported after a job has run and represent totals or averages for the duration of the entire job. However, the number of bytes replicated is reported while the copy process is running to allow a more granular view of bandwidth over time instead of just one value at the end. The frequency of reporting the metric can be controlled via the following properties (which show the default values, sending metrics every minute):

metrics-reporter:
    period: 1
    time-unit: MINUTES
Property Required Description
metrics-reporter.period No Period in between measurements, defaults to 1.
metrics-reporter.time-unit No TimeUnit for period e.g. MILLISECONDS, SECONDS, MINUTES, etc (see the enum java.util.concurrent.TimeUnit), defaults to MINUTES (needs to be upper-case to match the enum)

Graphite

Circus Train can be configured to send various metrics about a replication job to Graphite. Metrics are sent in the format:

    <prefix>.<namespace>[.database.table].<metric_name>

Note that database and table are only present for table level metrics and are automatically filled in by Circus Train. The following metrics are sent for each table/job.

Metric Name Description
completion_code See Exit codes for possible values.
[database].[table].completion_code 1 for success or -1 for failure.
[database].[table].replication_time Milliseconds taken to complete the replication. This includes all metastore interactions as well as the replication process itself.
[database].[table].bytes_replicated Number of bytes transmitted to the replica table location.
[database].[table].[hadoop_counters] All other metrics (typically Hadoop Counters from DistCp and S3MepReduceCp jobs).

The configuration options for Graphite are:

    graphite:
      config: hdfs:///path/to/your/conf/cluster.properties
      host:
      prefix:
      namespace: com.company.<team>.circus-train.<application>
Property Required Default/Example Description
graphite.config No e.g. hdfs:///foo/bar/cluster.properties The path to a Java Properties file containing the properties described below (i.e. you can set them in a separate file or directly in the configuration here).
graphite.host No hostname:port The hostname and port of the Graphite server. By default this is picked up from the config if provided.
graphite.prefix No e.g. dev All metrics are prefixed with prefix.namespace. By default this is picked up from the config if provided.
graphite.namespace Yes e.g. com.company.team.circus-train.application All metrics are prefixed with prefix.namespace. Formatting of this is important for reporting from Graphite.

Metadata transformations

Circus Train can transform the metadata of a newly created replica table by adding table parameters during the replication.

To use this feature, add properties to your configuration:

    table-replications:
       - ...
    transform-options:
      table-properties:
        my-custom-property: my-custom-value
        my-custom-property2: my-custom-value2

This will add my-custom-property and my-custom-property2 to the metadata of all tables in your replication. Surround each value with single quotes to ensure that Circus Train reads them as a String. Default properties can also be overridden by individual replications:

     table-replications:
       - source-table:
           ...
         replica-table:
           ...
         transform-options:
           table-properties:
             my-custom-override: my-custom-value
       - source-table:
           ...
         replica-table:
           ...
     transform-options:
       table-properties:
        my-custom-property: my-custom-value
        my-custom-property2: my-custom-value2

In this case, the first table replication will add just one property, my-custom-override, while the second will add the two default properties. Adding any number of override properties to a replication will override all default properties. To add properties to your table with non-alphanumeric characters (other than '-'), surround the key with single quotes and brackets like so:

    transform-options:
      table-properties:
        '[my.custom.property]': my-custom-value
        '[my.custom.property2]': my-custom-value2

This feature is a part of Circus Train and no further configuration is required. In addition to this, Circus Train supports custom metadata transformations which are built separately and placed on the Classpath when Circus Train is run. These are documented below.

Important Notes

  • By default, the source Hadoop and Hive configurations are loaded from the environment.
  • The replica database you are replicating into must already exist in the remote Hive metastore (i.e. Circus Train won't create this for you).
  • You will need to have the relevant permissions to create snapshots on the source if you have this feature enabled.
  • Removal of partitions is not replicated - i.e. if you remove a partition on the source it will remain on the replica.
  • If replicating to/from AWS ensure that the EMR master node is in your 'known_hosts' file.
  • Circus Train generally supports backwards compatible schema evolutions, for more information on how Circus Train handles schema evolution see this readme.

Notes on Hadoop S3 FileSystems

The Hadoop S3 FileSystem landscape is chaotic and confusing. What truths apply in a vanilla Apache Hadoop cluster do not necessarily hold in an EMR cluster. Protocols map to different implementations in different environments, and implementations can lack necessary features. For Circus Train we are interested in the following properties:

  • Optimal performance in EMR environments.
  • MD5 data verification for transmissions.
  • High-throughput transmission.
  • Support for very large files.
  • Copier execution taking place on-premises.
  • Adherence to Amazon EMR recommendations.

We have spent time investigating the S3 FileSystem implementations provided by both the Apache Hadoop project and the Amazon EMR project. Additionally we've analysed the operation of multiple versions of both DistCp and S3MapReduceCp. With this in mind we believe that the best way to move data using Circus Train from an on-premises cluster to S3 for consumption by an EMR cluster is using S3MapReduceCp.

We have verified its operation end-to-end, from a file based in HortonWorks HDFS to a query executing in EMR Hive. As a user all you need do is specify your S3 paths using only the s3:// protocol, Circus Train will handle the details.

Project layout

Below is a high level summary of the various modules that comprise Circus Train as they are laid out in version control.

  • circus-train-api: Connector contracts to allow alternative integrations.
  • circus-train-avro: Circus Train transformations to replicate Avro tables. readme
  • circus-train-aws: AWS specific connector implementations. Scripts for setting up a test AWS/EMR replica.
  • circus-train-aws-sns: Publishes replication events to Amazon SNS for workflow orchestration. readme
  • circus-train-comparator: Module used internally to compute a "Hive Diff" for comparison of source and replica data and metadata.
  • circus-train-core: Main class, replication strategies.
  • circus-train-distcp-copier: Builds a DistCp file copier.
  • circus-train-gcp: Builds Google Storage file copier. readme
  • circus-train-hive: Common Hive-related helper classes.
  • circus-train-hive-view: Components used to replicate views.
  • circus-train-housekeeping: A database-backed module that stores orphaned replica paths in a table for later clean up. This module is loaded by activating the housekeeping Spring Boot Profile
  • circus-train-integration-tests: Circus Train integration tests.
  • circus-train-metrics: Code used to enable reporting of Circus Train metrics to Graphite.
  • circus-train-package: Package Circus Train a tgz.
  • circus-train-s3-mapreduce-cp: Hadoop job based on DistCp optimised for AWS S3 data transfers. readme
  • circus-train-s3-mapreduce-cp-copier: Builds a S3MapReduceCp file copier using the package built in the circus-train-s3-mapreduce-cp module.
  • circus-train-s3-s3-copier: Delegates file copy from on S3 location to another S3 location to the AWS infrastructure.
  • circus-train-tool-parent: Sub-module for debug and maintenance tools. readme
    • circus-train-comparison-tool: Tool that performs data and metadata comparisons, useful for debugging the "Hive Diff" feature.
    • circus-train-filter-tool: Tool that checks partition filter expressions in table replication configurations.
    • circus-train-tool-core: Code common across tool implementations.
    • circus-train-tool: Packages tools tgz.

Extension points

Circus Train can be extended in various ways. This is an advanced feature that typically involves an implementation of an interface or an extension of existing classes. These are built separately from Circus Train and then placed on the classpath when Circus Train is run. Below is a summary of some common extension points.

  • ReplicationEventListeners can send replication start, success, and failure notifications to other systems. They are easy to develop and are dynamically loaded.
  • Source represents the cluster from which we are loading data. Designed to be reused for multiple tables/replications.
  • Replica represents the cluster to which we are copying data. Designed to be reused for multiple tables/replications.
  • Replications provide strategies for copying high level entities such as tables and partitions.
  • CopierFactories provide means of abstracting the complexity of creating Copiers.
  • Copiers move table data between cluster file systems.
  • LocationManagers look after the source and replica locations and are an ideal point to implement both snapshot isolation and retired data clean-up.
  • CompositeCopierFactory allows the provision of multiple copiers for the same table in order to add functionality to the copy process, e.g. introduce compaction, copy to multiple destinations, etc.

External Extensions

Circus Train DataSqueeze Copier

The DataSqueeze extension provides new Copiers for Circus Train which compact the data being replicated in various ways.

Circus Train BigQuery To Hive Replication

The Circus Train BigQuery extension enables the conversion of Google BigQuery tables to Hive.

Loading Extensions

Circus Train loads extensions using Spring's standard ComponentScan mechanism. Users can add their own packages to be scanned by declaring extension-packages as a comma separated list of package names to their YAML.

Example:

    extension-packages: some.extension,another.extension

In order not to clash with Circus Train's internal components we recommended you use your own package structure for these extensions (i.e. do not use com.hotels.circustrain or any sub-packages of this). The classes involved in implementing the extensions need to be made available on Circus Train's CLASSPATH. If your extension implementations require any existing Circus Train Beans then CircusTrainContext can be @Autowired in.

Custom metadata transformations

The following transformation interfaces can be implemented to manipulate metadata during replication. Note that transformations are loaded by Spring - so they must be annotated with @Component - and only one transformation of each type is allowed on the classpath. In case multiple transformations are required you can compose them into a single transformation function.

  • TableTransformation is applied to the table metadata before the replica database is updated.
  • PartitionTransformation is applied to the partition metadata before the replica database is updated.
  • ColumnStatisticsTransformation is applied to the column statistics metadata of both table and partitions before the replica database is updated. The property statsDesc.tblLevel of ColumnStatistics can be used to determine if the statistics are for the table or a partition.

Relevant configuration: table-replications[n].transform-options Example:

    table-replications:
      - ...
        transform-options:
           custom-transform-config-key1: someValue

Developing a ReplicationEventListener

A high level summary that assumes some knowledge of the Spring Framework and Spring Boot. See circus-train-aws-sns for a concrete example.

  • Create a new Maven project.
  • Add com.hotels.bdp:circus-train-api as a dependency.
  • Create a Spring context class to instantiate your listener's dependencies (clients etc.), mark it with @Configuration and mark factory methods with @Bean.
  • Create a JavaBean that represents your listener's configuration, mark it with @Configuration. Optionally assign it a prefix with @ConfigurationProperties.
  • Create your listener, implementing the com.hotels.bdp.circustrain.api.event.*Listener interfaces as needed, mark it as a @Component and @Autowire its dependencies.
  • Try to prevent blocking callers to your listener by invoking other systems asynchronously.
  • In a script that invokes Circus Train, append an entries to the CIRCUS_TRAIN_CLASSPATH environment variable to add your listener jar file and its dependencies to the Circus Train classpath.
  • Add the package where your Listener is located to the extension-packages section of your replication YAML.

Providing a custom MetaStoreClientFactory

MetaStoreClientFactory is responsible for creating the instances of the source and replica CloseableMetaStoreClient. Simply mark your implementation with @Component or have a @Configuration create a @Bean and the package names containing these classes to the extension-packages section of your YAML configuration.

The accept method must analyse the given metastore URL to determine whether the class can create a client for the metastore protocol. Circus Train will choose the first implementation in the classpath that accepts the metastore URL - the order in which these classes are loaded and interrogated is not controlled by Circus Train, i.e. the interrogation order in unpredictable and depends on the class loader.

Circus Train provides a MetaStoreClientFactory implementation out of the box that creates clients for the thrift protocol.

Implementing your own data copiers

New CopierFactories can be added just by extending the CopierFactory interface. Circus Train relies on Spring to detect a CopierFactory so each instance of such interface can either be instantiated in a Spring @Configuration class or the class itself can be annotated with @Component and @Order. The @Order annotation is used to establish precedence with respect to other CopierFactories that support the same schema.

The copier factories DistCpCopierFactory, S3MapReduceCpCopierFactory and S3S3CopierFactory, which are provided out of the box, take the lowest precedence allowing users to override them with their own implementations.

Implementing data transformations in the copying phase

The copy functionality of existing Copiers can be extended by declaring an instance of CompositeCopierFactory which allows a set of Copiers to be executed for one replication. This opens the possibility of adding functionality to the replication, like performing compaction before or after moving the data to its target location.

CompositeCopierFactory will support the same schema supported by the first CopierFactory in the delegates list.

All Copiers in the delegates list share the same set of configuration properties specified in copier-options. This set of properties can be used to control the behaviour of specific functionalities of each Copier. Users can add custom properties in this configuration section as well as set the values of any out-of-the-box Copier - refer to the Copier options section for details.

Connecting to a housekeeping DB

By default Circus Train uses an H2 file-based database to store information about dereferenced partition locations in the replica catalog. This should be transparent to most end users.

Configuration Examples

Below are a number of configuration fragments which can be used as a starting point for creating your own configuration files. These are for reference and will obviously need host names and various other values changed to suit your setup. Also keep in mind that database and table locations in Hive are fully qualified paths therefore if using AWS then EMR must be able to translate DNS names. If a DNS service is not available then each EMR node - master and slaves - must include an entry in etc/hosts for each domain name in the Circus Train configuration file.

Circus Train running on on-premises cluster A synchronizing tables from on-premises cluster A to on-premises cluster B

This configuration below shows how you could run Circus Train to replicate data between two on-premises clusters - for example you may want to replicate data between the on-premises DEV cluster (referred to as "cluster A") and the on-premises QA cluster (referred to as "cluster B"). Circus Train runs on DEV and it will push data to QA.

source-catalog:
  name: dev
replica-catalog:
  name: qa
  hive-metastore-uris: thrift://hiveQA.onpremises.domain:9083
table-replications:
  - source-table:
      database-name: dev_db
      table-name: dev_table
    replica-table:
      database-name: qa_db
      table-name: qa_table
      table-location: hdfs://nameNodeQA.onpremises.domain:8020/<path-to-qa_db>/qa_table

Circus Train running on on-premises cluster A synchronizing tables from on-premises cluster B to on-premises cluster A

This configuration below shows how you could run Circus Train to replicate data between two on-premises clusters - for example you may want to replicate data between the on-premises DEV cluster (referred to as "cluster A") and the on-premises QA cluster (referred to as "cluster B"). Circus Train runs on DEV and it will pull data from QA.

source-catalog:
  name: qa
  hive-metastore-uris: thrift://hiveQA.onpremises.domain:9083
replica-catalog:
  name: dev
  hive-metastore-uris: thrift://hiveDEV.onpremises.domain:9083
table-replications:
  - source-table:
      database-name: qa_db
      table-name: qa_table
    replica-table:
      database-name: dev_db
      table-name: dev_table
      table-location: hdfs://nameNodeDEV.onpremises.domain/<path-to-dev_db>/dev_table

Circus Train running on on-premises cluster A synchronizing tables from on-premises cluster A to EMR cluster B

This configuration below shows how you could run Circus Train to replicate data between an on-premises cluster and EMR - for example you may want to replicate data between the on-premises DEV cluster (referred to as "cluster A") and the EMR AWS cluster (referred to as "cluster B"). Circus Train runs on DEV and it will push data to AWS.

source-catalog:
  name: dev
replica-catalog:
  name: aws
  hive-metastore-uris: thrift://hiveAWS.amazonaws.com:9083
security:
  credential-provider: jceks://hdfs/<path-to-aws-credentials>/circus-train-credentials.jceks
table-replications:
  - source-table:
      database-name: dev_db
      table-name: dev_table
    replica-table:
      database-name: aws_db
      table-name: aws_table
      table-location: s3://<data-bucket>/<path-to-aws_db>/aws_table

Circus Train running on on-premises cluster A synchronizing tables from EMR cluster B to on-premises cluster A

This configuration below shows how you could run Circus Train to replicate data between an on-premises cluster and EMR - for example you may want to replicate data between the on-premises DEV cluster (referred to as "cluster A") and the EMR AWS cluster (referred to as "cluster B"). Circus Train runs on DEV and it will pull data from AWS.

source-catalog:
  name: aws
  hive-metastore-uris: thrift://hiveAWS.amazonaws.com:9083
replica-catalog:
  name: dev
  hive-metastore-uris: thrift://hiveDEV.onpremises.domain:9083
security:
  credential-provider: jceks://hdfs//<path-to-aws-credentials>/circus-train-credentials.jceks
table-replications:
  - source-table:
      database-name: aws_db
      table-name: aws_table
    replica-table:
      database-name: dev_db
      table-name: dev_table
      table-location: hdfs://nameNodeDEV.onpremises.domain/<path-to-dev_db>/dev_table

Circus Train running on EMR cluster A synchronizing tables from EMR cluster A to on-premises cluster B

This configuration below shows how you could run Circus Train to replicate data between EMR and an on-premises cluster - for example you may want to replicate data between the AWS cluster (referred to as "cluster A") and the on-premises DEV cluster (referred to as "cluster B"). Circus Train runs on AWS and it will push data to DEV.

Note that usually the EMR user who runs Circus Train is the default EMR user hadoop so prior to running Circus Train you may need to export the user who owns the table on the on-premises clusters, e.g. export HADOOP_USER_NAME=hwwetl.

source-catalog:
  name: aws
replica-catalog:
  name: dev
  hive-metastore-uris: thrift://hiveDEV.onpremises.domain:9083
copier-options:
  tmp-dir: hdfs:///tmp/circus-train/
table-replications:
  - source-table:
      database-name: aws_db
      table-name: aws_table
    replica-table:
      database-name: dev_db
      table-name: dev_table
      table-location: hdfs://nameNodeDEV.onpremises.domain/<path-to-dev_db>/dev_table

Circus Train running on EMR cluster A synchronizing tables from on-premises cluster B to EMR cluster A

This configuration below shows how you could run Circus Train to replicate data between EMR and an on-premises cluster - for example you may want to replicate data between the AWS cluster (referred to as "cluster A") and the on-premises DEV cluster (referred to as "cluster B"). Circus Train runs on AWS and it will pull data from DEV.

source-catalog:
  name: dev
  hive-metastore-uris: thrift://hiveDEV.onpremises.domain:9083
replica-catalog:
  name: aws
  hive-metastore-uris: thrift://hiveAWS.amazonaws.com:9083
copier-options:
  tmp-dir: hdfs:///tmp/circus-train/
table-replications:
  - source-table:
      database-name: dev_db
      table-name: dev_table
    replica-table:
      database-name: aws_db
      table-name: aws_table
      table-location: s3://<data-bucket>/<path-to-aws_db>/aws_table

Circus Train running on EMR cluster A synchronizing tables from EMR cluster A to EMR cluster B

This configuration below shows how you could run Circus Train to replicate data between two EMR clusters - for example you may want to replicate data between EMR AWS-A cluster (referred to as "cluster A") and the EMR AWS-B cluster (referred to as "cluster B"). Circus Train runs on AWS-A and it will push data to AWS-B.

Note that the Circus Train configuration only supports one set of AWS credentials so the AWS user must have access to both the source and destination buckets. This can be achieved via IAM roles and policies.

source-catalog:
  name: aws-a
  hive-metastore-uris: thrift://hiveAWS-A.amazonaws.com:9083
replica-catalog:
  name: aws-b
  hive-metastore-uris: thrift://hiveAWS-B.amazonaws.com:9083
copier-options:
  tmp-dir: hdfs:///tmp/circus-train/
table-replications:
  - source-table:
      database-name: aws-a_db
      table-name: aws-a_table
    replica-table:
      database-name: aws-b_db
      table-name: aws-b_table
      table-location: s3://<aws-b-data-bucket>/<pata-to-aws-b_db>/aws-b_table

Circus Train running on EMR cluster A synchronizing tables from EMR cluster B to EMR cluster A

This configuration below shows how you could run Circus Train to replicate data between two EMR clusters - for example you may want to replicate data between EMR AWS-A cluster (referred to as "cluster A") and the EMR AWS-B cluster (referred to as "cluster B"). Circus Train runs on AWS-A and it will pull data from AWS-B.

Note that the Circus Train configuration only supports one set of AWS credentials so the AWS user must have access to both the source and destination buckets. This can be achieved via IAM roles and policies.

source-catalog:
  name: aws-b
  hive-metastore-uris: thrift://hiveAWS-B.amazonaws.com:9083
replica-catalog:
  name: aws-a
  hive-metastore-uris: thrift://hiveAWS-A.amazonaws.com:9083
copier-options:
  tmp-dir: hdfs:///tmp/circus-train/
security:
  credential-provider: jceks://hdfs/<path-to-aws-credentials-on-hdfs-A>/circus-train-credentials.jceks
table-replications:
  - source-table:
      database-name: aws-b_db
      table-name: aws-b_table
    replica-table:
      database-name: aws-a_db
      table-name: aws-a_table
      table-location: s3://<aws-a-data-bucket>/<pata-to-aws-a_db>/aws-a_table

Contact

Mailing List

If you would like to ask any questions about or discuss Circus Train please join our mailing list at

https://groups.google.com/forum/#!forum/circus-train-user

Credits

Created by Elliot West, Daniel del Castillo, Patrick Duin, Dave Maughan & Courtney Edwards with thanks to: Adrian Woodhead, Dave Bauman, Jose Nuñez Izu and Oscar Mateos Ventura.

The Circus Train logo uses the Ewert font by Johan Kallas under the SIL Open Font License (OFL).

Legal

This project is available under the Apache 2.0 License.

Copyright 2016-2020 Expedia, Inc.