Storage Space Planning for Cloudera Manager

RDBMS:

Use any supported RDBMS to store the core configuration of your Cloudera Manager database and all cluster, service, and role configurations.

See Cloudera Manager and Managed Service Data Stores.

DISK:

Cloudera Manager Server Local Data Storage Directory (command_storage_path) on the host where the Cloudera Manager Server is configured to run. This local path is used by Cloudera Manager for storing data, including command result files. Critical configurations are not stored in this location.

/var/lib/cloudera-scm-server/

Command Eviction Age

Length of time after which inactive commands are evicted from the database.

Default is two years.

You should perform regular, verified, remotely-stored backups of the Cloudera Manager Server database.

See Cloudera Manager and Managed Service Data Stores.

You control Activity Monitor storage usage by configuring the number of days or hours of data to retain. Older data are purged.

To configure data retention in the Cloudera Manager Administration Console, navigate to Home > Cloudera Management Service > Configuration > Activity Monitor and edit the following properties:

Purge Activities Data at This Age

In Activity Monitor, purge data about MapReduce jobs and aggregate activities when the data reaches this age in hours. By default, Activity Monitor keeps data about activities for 336 hours (14 days).

Purge Attempts Data at This Age

In the Activity Monitor, purge data about MapReduce attempts when the data reaches this age in hours. Because attempt data may consume large amounts of database space, you may wish to purge it more frequently than activity data. By default, Activity Monitor keeps data about attempts for 336 hours (14 days).

Purge MapReduce Service Data at This Age

The number of hours of past service-level data to keep in the Activity Monitor database, such as total slots running. The default is to keep data for 336 hours (14 days).

The Activity Monitor only monitors MapReduce (MRv1) jobs, and does not monitor not YARN applications. If you no longer use MapReduce (MRv1) in your cluster, the Activity Monitor is not required for Cloudera Manager 5 (or later) or CDH 5 (or higher).

The amount of storage space needed for 14 days worth of MapReduce (MRv1) activities can vary greatly and directly depends on the size of your cluster and the level of activity that uses MapReduce (MRv1). It may be necessary to adjust and readjust the amount of storage as you determine the "stable state" and "burst state" of the MapReduce activity in your cluster.

For example, consider the following test cluster and usage:

• A simulated 1000-host cluster, each host with 32 slots

• Synthetic MapReduce (MRv1) jobs with 200 attempts (tasks) per activity (job)

Sizing observations for this cluster:

• Each attempt takes 10 minutes to complete.

• This usage results in roughly 20 thousand jobs a day with some 5 million total attempts.

• For a retention period of 7 days, this Activity Monitor database required 200 GB.

• 10 GiB Services Time Series Storage +

• 1 GiB Impala Query Storage +

• 1 GiB YARN Application Storage

Total: ~12 GiB Minimum (No Maximum)

Service Monitor data growth is controlled by configuring the maximum amount of storage space it may use.

To configure data retention in Cloudera Manager Administration Console, select Home > Cloudera Management Service > Configuration > Service Monitor Default Group and edit the following properties:

Time-Series Storage

The approximate amount of disk space dedicated to storing time series and health data. When the store has reached its maximum size, it deletes older data to make room for newer data. The disk usage is approximate because the store only begins deleting data once it reaches the limit.

Note that Cloudera Manager stores time-series data at a number of different data granularities, and these granularities have different effective retention periods. The Service Monitor stores metric data not only as raw data points but also as ten-minute, hourly, six-hourly, daily, and weekly summary data points. Raw data consumes the bulk of the allocated storage space and weekly summaries consume the least. Raw data is retained for the shortest amount of time while weekly summary points are unlikely to ever be deleted.

Select Cloudera Management Service > Charts Library tab in Cloudera Manager for information about how space is consumed within the Service Monitor. These pre-built charts also show information about the amount of data retained and time window covered by each data granularity.

Impala Storage

The approximate amount of disk space dedicated to storing Impala query data. When the store reaches its maximum size, it deletes older to make room for newer queries. The disk usage is approximate because the store only begins deleting data when it reaches the limit.

YARN Storage

The approximate amount of disk space dedicated to storing YARN application data. Once the store reaches its maximum size, it deletes older data to make room for newer applications. The disk usage is approximate because Cloudera Manager only begins deleting data when it reaches the limit.

Host Monitor data growth is controlled by configuring the maximum amount of storage space it may use.

See Data Storage for Monitoring Data.

To configure these data retention in Cloudera Manager Administration Console, select : Home > Cloudera Management Service > Configuration > Host Monitor Default Group

Time-Series Storage

The approximate amount of disk space dedicated to storing time series and health data. When the store reaches its maximum size, it deletes older data to make room for newer data. The disk usage is approximate because the store only begins deleting data when it reaches the limit.

Note that Cloudera Manager stores time-series data at a number of different data granularities, and these granularities have different effective retention periods. Host Monitor stores metric data not only as raw data points but also ten-minutely, hourly, six-hourly, daily, and weekly summary data points. Raw data consumes the bulk of the allocated storage space and weekly summaries consume the least. Raw data is retained for the shortest amount of time, while weekly summary points are unlikely to ever be deleted.

See the Cloudera Management Service > Charts Library tab in Cloudera Manager for information on how space is consumed within the Host Monitor. These pre-built charts also show information about the amount of data retained and the time window covered by each data granularity.

The amount of storage space the Event Server uses is influenced by configuring how many discrete events it may retain.

To configure data retention in Cloudera Manager Administration Console, select Home > Cloudera Management Service > Configuration > Event Server Group and edit the following property:

Maximum Number of Events in the Event Server Store

The maximum size of the Event Server store, in events. Once this size is exceeded, events are deleted starting with the oldest first until the size of the store is below this threshold

The Event Server is a managed Lucene index that collects relevant events that happen within your cluster, such as results of health tests, log events that are created when a log entry matches a set of rules for identifying messages of interest and makes them available for searching, filtering and additional action. You can view and filter events on the Diagnostics > Events tab of the Cloudera Manager Administration Console. You can also poll this data using the Cloudera Manager API.

RDBMS:

Any Supported RDBMS.

See Installing and Configuring Databases.

DISK:

/var/lib/cloudera-scm-headlamp/ on the host where the Reports Manager role is configured to run.

RDBMS:

There are no exposed defaults or configurations to directly cull or purge the size of this data set.

DISK:

There are no configuration defaults to influence the size of this location. The size of the data in this location depends not only on the size of the HDFS fsimage, but also on the HDFS path complexity.

The Reports Manager uses space in two main locations, one local on the host where Reports Manager runs, and the other in the RDBMS provided to it for its historical aggregation. The RDBMS is not required to be on the same host where the Reports Manager runs.

Reports Manager downloads the fsimage from the Namenode every 60 minutes (default) and stores it locally to perform operations against, including indexing the HDFS filesystem structure represented in the fsimage. A larger fsimage, or more deep and complex paths within HDFS consume more disk space.

Reports Manager has no control over the size of the fsimage. If your total HDFS usage trends upward notably or you add excessively long paths in HDFS, it may be necessary to revisit and adjust the amount of space allocated to the Reports Manager for its local storage. Periodically monitor, review and readjust the local storage allocation.

Any Supported RDBMS.

See Installing and Configuring Databases.

Navigator Audit Server storage usage is controlled by configuring how many days of data it may retain. Any older data are purged.

To configure data retention in the Cloudera Manager Administration Console, select Home > Cloudera Management Service > Configuration > Navigator Audit Server Default Group and edit the following property:

Navigator Audit Server Data Expiration Period

In Navigator Audit Server, purge audit data of various auditable services when the data reaches this age in days. By default, Navigator Audit Server keeps data about audits for 90 days.

The size of the Navigator Audit Server database directly depends on the number of audit events the cluster’s audited services generate. Normally the volume of HDFS audits exceed the volume of other audits (all other components like MRv1, Hive and Impala read from HDFS, which generates additional audit events).

The average size of a discrete HDFS audit event is ~1 KB. For a busy cluster of 50 hosts with ~100K audit events generated per hour, the Navigator Audit Server database would consume ~2.5 GB per day. To retain 90 days of audits at that level, plan for a database size of around 250 GB. If other configured cluster services generate roughly the same amount of data as the HDFS audits, plan for the Navigator Audit Server database to require around 500 GB of storage for 90 days of data.

Notes:

• Individual Hive and Impala queries themselves can be very large. Since the query itself is part of an audit event, such audit events consume space in proportion to the length of the query.

• The amount of space required increases as activity on the cluster increases. In some cases, Navigator Audit Server databases can exceed 1TB for 90 days of audit events. Benchmark your cluster periodically and adjust accordingly.

Use this table to map Product Compatibility Matrix for Cloudera Navigator versions to Cloudera Manager versions.

RDBMS:

Any Supported RDBMS.

See Installing and Configuring Databases.

DISK:

/var/lib/cloudera-scm-navigator/ on the host where the Navigator Metadata Server role is configured to run,

RDBMS:

There are no exposed defaults or configurations to directly cull or purge the size of this data set.

DISK:

There are no configuration defaults to influence the size of this location. You can change the location itself with the Navigator Metadata Server Storage Dir property. The size of the data in this location depends on the amount of metadata in the system (HDFS fsimage size, Hive Metastore size) and activity on the system (the number of MapReduce Jobs run, Hive queries executed, etc).

RDBMS:

There is no maximum size of this data and no way to purge data that is old.

DISK:

There is no maximum size of this data. As data in the cluster grows its metadata is captured and stored in the location specified by the Navigator Metadata Server Storage Dir property.

RDBMS:

The database is used to store policies and authorization data. The dataset is small, but this database is also used during a Solr schema upgrade, where Solr documents are extracted and inserted again in Solr. This has same space requirements as above use case, but the space is only used temporarily during product upgrades.

Use this matrix to map Cloudera Navigator and Cloudera Manager versions.

DISK:

This filesystem location contains all the metadata that is extracted from managed clusters. The data is stored in Solr, so this is the location where Solr stores its index and documents. Depending on the size of the cluster, this data can occupy tens of gigabytes. A guideline is to look at the size of HDFS fsimage and allocate two to three times that size as the initial size. The data here is incremental and continues to grow as activity is performed on the cluster. The rate of growth can be on order of tens of megabytes per day.

Local Parcel Repository Path

/opt/cloudera/parcel-repo

This path exists only on the host where Cloudera Manager Server (cloudera-scm-server) runs. The Cloudera Manager Server stages all new parcels in this location as it fetches them from any external repositories. Cloudera Manager Agents are then instructed to fetch the parcels from this location when the administrator distributes the parcel via the Cloudera Manager Administration Console or the Cloudera Manager API.

The default location is /opt/cloudera/parcel-repo but you can configure another local filesystem location on the host where Cloudera Manager Server runs. See Parcel Configuration Settings.

Provide sufficient space to hold all the parcels you download from all configured Remote Parcel Repository URLs (See Parcel Configuration Settings). Cloudera Manager deployments that manage multiple clusters store all applicable parcels for all clusters.

Parcels are provided for each operating system, so be aware that heterogeneous clusters (distinct operating systems represented in the cluster) require more space than clusters with homogeneous operating systems.

For example, a cluster with both RHEL5.x and 6.x hosts must hold -el5 and -el6 parcels in the Local Parcel Repository Path, which requires twice the amount of space.

Delete any parcels that are no longer in use from the Cloudera Manager Administration Console, (never delete them manually from the command line) to recover disk space in the Local Parcel Repository Path and simultaneously across all managed cluster hosts which hold the parcel.

Perform regular backups of this path, and consider it a non-optional accessory to backing up Cloudera Manager Server. If you migrate Cloudera Manager Server to a new host or restore it from a backup (for example, after a hardware failure), recover the full content of this path to the new host, in the /opt/cloudera/parcel-repo directory before starting any cloudera-scm-agent or cloudera-scm-server processes.

Parcel Cache

/opt/cloudera/parcel-cache

Managed Hosts running a Cloudera Manager Agent stage distributed parcels into this path (as .parcel files, unextracted). Do not manually manipulate this directory or its files.

Provide sufficient space per-host to hold all the parcels you distribute to each host.

You can configure Cloudera Manager to remove these cached .parcel files after they are extracted and placed in /opt/cloudera/parcels/. It is not mandatory to keep these temporary files but keeping them avoids the need to transfer the .parcel file from the Cloudera Manager Server repository should you need to extract the parcel again for any reason.

To configure this behavior in the Cloudera Manager Administration Console, select Administration > Settings > Parcels > Retain Downloaded Parcel Files

Host Parcel Directory

/opt/cloudera/parcels

Managed cluster hosts running a Cloudera Manager Agent extract parcels from the /opt/cloudera/parcel-cache directory into this path upon parcel activation. Many critical system symlinks point to files in this path and you should never manually manipulate its contents.

Provide sufficient space on each host to hold all the parcels you distribute to each host. Be aware that the typical CDH parcel size is slightly larger than 1 GB per parcel. If you maintain various versions of parcels staged before and after upgrading, be aware of the disk space implications.

You can configure Cloudera Manager to automatically remove older parcels once they are no longer in use. As an administrator you can always manually delete parcel versions not in use, but configuring these settings can handle the deletion automatically, in case you forget.

To configure this behavior in the Cloudera Manager Administration Console, select Administration > Settings > Parcels and configure the following property:

Automatically Remove Old Parcels

This parameter controls whether parcels for old versions of an activated product should be removed from a cluster when they are no longer in use.

The default value is Disabled.

Number of Old Parcel Versions to Retain

If you enable Automatically Remove Old Parcels, this setting specifies the number of old parcels to keep. Any old parcels beyond this value are removed. If this property is set to zero, no old parcels are retained.

The default value is 3.

Activity Monitor (One-time)

The Activity Monitor only works against a MapReduce (MR1) service, not YARN. So if your deployment has fully migrated to YARN and no longer uses a MapReduce (MR1) service, your Activity Monitor database is no longer growing. If you have waited longer than the default Activity Monitor retention period (14 days) to address this point, then the Activity Monitor has already purged it all for you and your database is mostly empty. If your deployment meets these conditions, consider cleaning up by dropping the Activity Monitor database (again, only when you are satisfied that you no longer need the data or have confirmed that it is no longer in use) and the Activity Monitor role.

For those who used Cloudera Manager version 4.x and have now upgraded to version 5.x: The Service Monitor and Host Monitor were migrated from their previously-configured RDBMS into a dedicated time series store used solely by each of these roles respectively. After this happens, there is still legacy database connection information in the configuration for these roles. This was used to allow for the initial migration but is no longer being used for any active work.

After the above migration has taken place, the RDBMS databases previously used by the Service Monitor and Host Monitor are no longer used. Space occupied by these databases is now recoverable. If appropriate in your environment (and you are satisfied that you have long-term backups or do not need the data on disk any longer), you can drop those databases.

Cloudera Management Services are automatically rolling up, purging or otherwise consolidating aged data for you in the background. Configure retention and purging limits per-role to control how and when this occurs. These configurations are discussed per-entity above. Adjust the default configurations to meet your space limitations or retention needs.