cassandra-data-migrator (also known as CDM)

Migrate and Validate Tables between Origin and Target Cassandra Clusters.

:warning: Please note this job has been tested with spark version 3.5.3

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

• Install Java11 (minimum) as Spark binaries are compiled with it.
• Install Spark version 3.5.3 on a single VM (no cluster necessary) where you want to run this job. Spark can be installed by running the following: -

  wget https://archive.apache.org/dist/spark/spark-3.5.3/spark-3.5.3-bin-hadoop3-scala2.13.tgz
  tar -xvzf spark-3.5.3-bin-hadoop3-scala2.13.tgz

:warning: If the above Spark and Scala version is not properly installed, you'll then see a similar exception like below when running the CDM jobs,

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

Steps for Data-Migration:

:warning: Note that Version 4 of the tool is not backward-compatible with .properties files created in previous versions, and that package names have changed.

1. cdm.properties file needs to be configured as applicable for the environment. Parameter descriptions and defaults are described in the file. The file can have any name, it does not need to be cdm.properties.

   • A simplified sample properties file configuration can be found here as cdm.properties
   • A complete sample properties file configuration 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 below 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-4.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 in the target keyspace, pass param --conf spark.cdm.trackRun=true
• To filter and migrate data only in a specific token range, you can pass the below two additional params to the Migration or Validation jobs

--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-4.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-4.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 (makes target same as origin).
• Enable/disable this feature using one or both of the below setting in the config file

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

  Note:

• 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-4.x.x.jar &> logfile_name_$(date +%Y%m%d_%H_%M).txt

Perform large-field Guardrail violation checks

• The tool can be used to identify large fields from a 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-4.x.x.jar &> logfile_name_$(date +%Y%m%d_%H_%M).txt

Features

• Auto-detects table schema (column names, types, keys, collections, UDTs, etc.)
  • Including counter table Counter tables
• Rerun job from where the previous job had 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
• 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, 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.

Performance FAQ

• Below recommendations may only be needed while 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
  • 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 & this usually fine. However you can reduce this if your table has many large rows (over 100KB).
  • ratelimit: Default is 20K. Once you set all the other properties appropriately, set this value to the highest possible value that your cluster (origin & target) is able to handle.
• Using schema manipulation features (like constantColumns, explodeMap, extractJson), transformation functions and/or where-filter-conditions (except partition min/max) may negatively impacts 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 cluster.
• Note: For additional performance tuning, refer to details mentioned in 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-4.x.x.jar) file should now be present in the target folder

Contributors

Checkout all our wonderful contributors here.