cassandra-data-migrator (also known as CDM)

Migrate and Validate Tables between Origin and Target Cassandra Clusters.

Important

Please note this job has been tested with spark version 3.5.4

Install as a Container

• Get the latest image that includes all dependencies from DockerHub
  • All migration tools (cassandra-data-migrator + dsbulk + cqlsh) would be available in the /assets/ folder of the container

Install as a JAR file

• Download the latest jar file from the GitHub packages area here

Prerequisite

• Java11 (minimum) as Spark binaries are compiled with it.
• Spark 3.5.x with Scala 2.13 and Hadoop 3.3
  • Typically installed using this binary on a single VM (no cluster necessary) where you want to run this job. This simple setup is recommended for most one-time migrations.
  • However we recommend using a Spark Cluster or a Spark Serverless platform like Databricks or Google Dataproc (that supports the above mentioned versions) for large (e.g. several terabytes) complex migrations OR when CDM is used as a long-term data-transfer utility and not a one-time job.

Spark can be installed by running the following: -

wget https://archive.apache.org/dist/spark/spark-3.5.4/spark-3.5.4-bin-hadoop3-scala2.13.tgz
tar -xvzf spark-3.5.4-bin-hadoop3-scala2.13.tgz

Caution

If the above Spark and Scala version does not match, you may see an exception like below when running the CDM jobs,

Exception in thread "main" java.lang.NoSuchMethodError: scala.runtime.Statics.releaseFence()V

Note

When deploying CDM on a Spark cluster, replace the params --master "local[*]" with --master "spark://master-host:port" and remove any params (e.g. --driver-memory, --executor-memory, etc.) related to a single VM run

Steps for Data-Migration:

1. cdm.properties file needs to be configured as applicable for the environment. The file can have any name, it does not need to be cdm.properties.

   • A sample properties file with default values can be found here as cdm.properties
   • A complete reference properties file with default values can be found here as cdm-detailed.properties

2. Place the properties file where it can be accessed while running the job via spark-submit.
3. Run the job using spark-submit command as shown below:

spark-submit --properties-file cdm.properties \
--conf spark.cdm.schema.origin.keyspaceTable="<keyspacename>.<tablename>" \
--master "local[*]" --driver-memory 25G --executor-memory 25G \
--class com.datastax.cdm.job.Migrate cassandra-data-migrator-5.x.x.jar &> logfile_name_$(date +%Y%m%d_%H_%M).txt

Note:

• Above command generates a log file logfile_name_*.txt to avoid log output on the console.
• Update the memory options (driver & executor memory) based on your use-case
• To track details of a run (recorded on the target keyspace), pass param --conf spark.cdm.trackRun=true
• To filter records only for a specific token range, pass the below two additional params to the Migration OR Validation job

--conf spark.cdm.filter.cassandra.partition.min=<token-range-min>
--conf spark.cdm.filter.cassandra.partition.max=<token-range-max>

Steps for Data-Validation:

• To run the job in Data validation mode, use class option --class com.datastax.cdm.job.DiffData as shown below

spark-submit --properties-file cdm.properties \
--conf spark.cdm.schema.origin.keyspaceTable="<keyspacename>.<tablename>" \
--master "local[*]" --driver-memory 25G --executor-memory 25G \
--class com.datastax.cdm.job.DiffData cassandra-data-migrator-5.x.x.jar &> logfile_name_$(date +%Y%m%d_%H_%M).txt

• Validation job will report differences as “ERRORS” in the log file as shown below.

23/04/06 08:43:06 ERROR DiffJobSession: Mismatch row found for key: [key3] Mismatch: Target Index: 1 Origin: valueC Target: value999) 
23/04/06 08:43:06 ERROR DiffJobSession: Corrected mismatch row in target: [key3]
23/04/06 08:43:06 ERROR DiffJobSession: Missing target row found for key: [key2]
23/04/06 08:43:06 ERROR DiffJobSession: Inserted missing row in target: [key2]

• Please grep for all ERROR from the output log files to get the list of missing and mismatched records.
  • Note that it lists differences by primary-key values.
• If you would like to redirect such logs (rows with details of missing and mismatched rows) into a separate file, you could use the log4j2.properties file provided here as shown below

spark-submit --properties-file cdm.properties \
--conf spark.cdm.schema.origin.keyspaceTable="<keyspacename>.<tablename>" \
--conf spark.executor.extraJavaOptions='-Dlog4j.configurationFile=log4j2.properties' \ 
--conf spark.driver.extraJavaOptions='-Dlog4j.configurationFile=log4j2.properties' \ 
--master "local[*]" --driver-memory 25G --executor-memory 25G \
--class com.datastax.cdm.job.DiffData cassandra-data-migrator-5.x.x.jar &> logfile_name_$(date +%Y%m%d_%H_%M).txt

• The Validation job can also be run in an AutoCorrect mode. This mode can
  • Add any missing records from origin to target
  • Update any mismatched records between origin and target
• Enable/disable this feature using one or both of the below params in the properties file

spark.cdm.autocorrect.missing                     false|true
spark.cdm.autocorrect.mismatch                    false|true

Important

The validation job will never delete records from target i.e. it only adds or updates data on target

Rerun (previously incomplete) Migration or Validation

• You can rerun/resume a Migration or Validation job to complete a previous run that could have stopped (or completed with some errors) for any reasons. This mode will skip any token-ranges from the previous run that were migrated (or validated) successfully. This is done by passing the spark.cdm.trackRun.previousRunId param as shown below

spark-submit --properties-file cdm.properties \
 --conf spark.cdm.schema.origin.keyspaceTable="<keyspacename>.<tablename>" \
 --conf spark.cdm.trackRun.previousRunId=<prev_run_id> \
 --master "local[*]" --driver-memory 25G --executor-memory 25G \
 --class com.datastax.cdm.job.<Migrate|DiffData> cassandra-data-migrator-5.x.x.jar &> logfile_name_$(date +%Y%m%d_%H_%M).txt

Perform large-field Guardrail violation checks

• This mode can help identify large fields on an origin table that may break you cluster guardrails (e.g. AstraDB has a 10MB limit for a single large field), use class option --class com.datastax.cdm.job.GuardrailCheck as shown below

spark-submit --properties-file cdm.properties \
--conf spark.cdm.schema.origin.keyspaceTable="<keyspacename>.<tablename>" \
--conf spark.cdm.feature.guardrail.colSizeInKB=10000 \
--master "local[*]" --driver-memory 25G --executor-memory 25G \
--class com.datastax.cdm.job.GuardrailCheck cassandra-data-migrator-5.x.x.jar &> logfile_name_$(date +%Y%m%d_%H_%M).txt

Note

This mode only operates on one database i.e. origin, there is no target in this mode

Features

• Auto-detects table schema (column names, types, keys, collections, UDTs, etc.)
  • Including counter table Counter tables
• Rerun/Resume a previous job that may have stopped for any reason (killed, had exceptions, etc.)
  • If you rerun a validation job, it will include any token-ranges that had differences in the previous run
• Preserve writetimes and TTLs
• Supports migration/validation of advanced DataTypes (Sets, Lists, Maps, UDTs)
• Filter records from Origin using writetime and/or CQL conditions and/or a list of token-ranges
• Perform guardrail checks (identify large fields)
• Supports adding constants as new columns on Target
• Supports expanding Map columns on Origin into multiple records on Target
• Supports extracting value from a JSON column in Origin and map it to a specific field on Target
• Can be deployed on a Spark Cluster or a single VM
• Fully containerized (Docker and K8s friendly)
• SSL Support (including custom cipher algorithms)
• Migrate from any Cassandra Origin (Apache Cassandra® / DataStax Enterprise™ / DataStax Astra DB™) to any Cassandra Target (Apache Cassandra® / DataStax Enterprise™ / DataStax Astra DB™)
• Supports migration/validation from and to Azure Cosmos Cassandra
• Validate migration accuracy and performance using a smaller randomized data-set
• Supports adding custom fixed writetime and/or ttl
• Track run information (start-time, end-time, run-metrics, status, etc.) in tables (cdm_run_info and cdm_run_details) on the target keyspace

Things to know

• Each run (Migration or Validation) can be tracked (when enabled). You can find summary and details of the same in tables cdm_run_info and cdm_run_details in the target keyspace.
• CDM does not migrate ttl & writetime at the field-level (for optimization reasons). It instead finds the field with the highest ttl & the field with the highest writetime within an origin row and uses those values on the entire target row.
• CDM ignores using collection and UDT fields for ttl & writetime calculations by default for performance reasons. If you want to include such fields, set spark.cdm.schema.ttlwritetime.calc.useCollections param to true.
• If a table has only collection and/or UDT non-key columns and no table-level ttl configuration, the target will have no ttl, which can lead to inconsistencies between origin and target as rows expire on origin due to ttl expiry. If you want to avoid this, we recommend setting spark.cdm.schema.ttlwritetime.calc.useCollections param to true in such scenarios.
• If a table has only collection and/or UDT non-key columns, the writetime used on target will be time the job was run. If you want to avoid this, we recommend setting spark.cdm.schema.ttlwritetime.calc.useCollections param to true in such scenarios.
• When CDM migration (or validation with autocorrect) is run multiple times on the same table (for whatever reasons), it could lead to duplicate entries in list type columns. Note this is due to a Cassandra/DSE bug and not a CDM issue. This issue can be addressed by enabling and setting a positive value for spark.cdm.transform.custom.writetime.incrementBy param. This param was specifically added to address this issue.
• When you rerun job to resume from a previous run, the run metrics (read, write, skipped, etc.) captured in table cdm_run_info will be only for the current run. If the previous run was killed for some reasons, its run metrics may not have been saved. If the previous run did complete (not killed) but with errors, then you will have all run metrics from previous run as well.
• When running on a Spark Cluster (and not a single VM), the rate-limit values (spark.cdm.perfops.ratelimit.origin & spark.cdm.perfops.ratelimit.target) applies to individual Spark worker nodes. Hence this value should be set to the effective-rate-limit-you-need/number-of-spark-worker-nodes . E.g. If you need an effective rate-limit of 10000, and the number of Spark worker nodes are 4, then you should set the above rate-limit params to a value of 2500.

Performance recommendations

Below recommendations may only be useful when migrating large tables where the default performance is not good enough

• Performance bottleneck are usually the result of
  • Low resource availability on Origin OR Target cluster
  • Low resource availability on CDM VMs, see recommendations here
  • Bad schema design which could be caused by out of balance Origin cluster, large partitions (> 100 MB), large rows (> 10MB) and/or high column count.
• Incorrect configuration of below properties may negatively impact performance
  • numParts: Default is 5K, but ideal value is usually around table-size/10MB.
  • batchSize: Default is 5, but this should be set to 1 for tables where primary-key=partition-key OR where average row-size is > 20 KB. Similarly, this should be set to a value > 5, if row-size is small (< 1KB) and most partitions have several rows (100+).
  • fetchSizeInRows: Default is 1K and this usually works fine. However you can reduce this as needed if your table has many large rows (over 100KB).
  • ratelimit: Default is 20000, but this property should usually be updated (after updating other properties) to the highest possible value that your origin and target clusters can efficiently handle.
• Using schema manipulation features (like constantColumns, explodeMap, extractJson), transformation functions and/or where-filter-conditions (except partition min/max) may negatively impact performance
• We typically recommend this infrastructure for CDM VMs and this starter conf. You can then optimize the job further based on CDM params info provided above and the observed load and throughput on Origin and Target clusters
• We recommend using a Spark Cluster or a Spark Serverless platform like Databricks or Google Dataproc for large (e.g. several terabytes) complex migrations OR when CDM is used as a long-term data-transfer utility and not a one-time job.

Note

For additional performance tuning, refer to details mentioned in the cdm-detailed.properties file here

Building Jar for local development

1. Clone this repo
2. Move to the repo folder cd cassandra-data-migrator
3. Run the build mvn clean package (Needs Maven 3.9.x)
4. The fat jar (cassandra-data-migrator-5.x.x.jar) file should now be present in the target folder

Contributors

Checkout all our wonderful contributors here.

---