Records that fail to get loaded into the CarbonData due to data type incompatibility or are empty or have incompatible format are classified as Bad Records.
The bad records are stored at the location set in carbon.badRecords.location in carbon.properties file.
By default carbon.badRecords.location specifies the following location /opt/Carbon/Spark/badrecords.
While loading data we can specify the approach to handle Bad Records. In order to analyse the cause of the Bad Records the parameter BAD_RECORDS_LOGGER_ENABLE must be set to value TRUE. There are multiple approaches to handle Bad Records which can be specified by the parameter BAD_RECORDS_ACTION.
'BAD_RECORDS_ACTION'='FORCE'
'BAD_RECORDS_ACTION'='REDIRECT'
To ignore the Bad Records from getting stored in the raw csv, we need to set the following in the query :
'BAD_RECORDS_ACTION'='IGNORE'
The store location specified while creating carbon session is used by the CarbonData to store the meta data like the schema, dictionary files, dictionary meta data and sort indexes.
Try creating carbonsession with storepath specified in the following manner :
val carbon = SparkSession.builder().config(sc.getConf).getOrCreateCarbonSession(<carbon_store_path>)
Example:
val carbon = SparkSession.builder().config(sc.getConf).getOrCreateCarbonSession("hdfs://localhost:9000/carbon/store")
The Apache CarbonData acquires lock on the files to prevent concurrent operation from modifying the same files. The lock can be of the following types depending on the storage location, for HDFS we specify it to be of type HDFSLOCK. By default it is set to type LOCALLOCK. The property carbon.lock.type configuration specifies the type of lock to be acquired during concurrent operations on table. This property can be set with the following values :
In order to build CarbonData project it is necessary to specify the spark profile. The spark profile sets the Spark Version. You need to specify the spark version while using Maven to build project.
Carbon support insert operation, you can refer to the syntax mentioned in DML Operations on CarbonData. First, create a source table in spark-sql and load data into this created table.
CREATE TABLE source_table(
id String,
name String,
city String)
ROW FORMAT DELIMITED FIELDS TERMINATED BY ",";
SELECT * FROM source_table;
id name city
1 jack beijing
2 erlu hangzhou
3 davi shenzhen
Scenario 1 :
Suppose, the column order in carbon table is different from source table, use script "SELECT * FROM carbon table" to query, will get the column order similar as source table, rather than in carbon table's column order as expected.
CREATE TABLE IF NOT EXISTS carbon_table(
id String,
city String,
name String)
STORED AS carbondata;
INSERT INTO TABLE carbon_table SELECT * FROM source_table;
SELECT * FROM carbon_table;
id city name
1 jack beijing
2 erlu hangzhou
3 davi shenzhen
As result shows, the second column is city in carbon table, but what inside is name, such as jack. This phenomenon is same with insert data into hive table.
If you want to insert data into corresponding column in carbon table, you have to specify the column order same in insert statement.
INSERT INTO TABLE carbon_table SELECT id, city, name FROM source_table;
Scenario 2 :
Insert operation will be failed when the number of column in carbon table is different from the column specified in select statement. The following insert operation will be failed.
INSERT INTO TABLE carbon_table SELECT id, city FROM source_table;
Scenario 3 :
When the column type in carbon table is different from the column specified in select statement. The insert operation will still success, but you may get NULL in result, because NULL will be substitute value when conversion type failed.
Spark executor shows task as failed after the maximum number of retry attempts, but loading the data having bad records and BAD_RECORDS_ACTION (carbon.bad.records.action) is set as "FAIL" will attempt only once but will send the signal to driver as failed instead of throwing the exception to retry, as there is no point to retry if bad record found and BAD_RECORDS_ACTION is set to fail. Hence the Spark executor displays this one attempt as successful but the command has actually failed to execute. Task attempts or executor logs can be checked to observe the failure reason.
SDK writer is an independent entity, hence SDK writer can generate carbondata files from a non-cluster machine that has different time zones. But at cluster when those files are read, it always takes cluster time-zone. Hence, the value of timestamp and date datatype fields are not original value. If wanted to control timezone of data while writing, then set cluster's time-zone in SDK writer by calling below API.
TimeZone.setDefault(timezoneValue)
Example:
cluster timezone is Asia/Shanghai
TimeZone.setDefault(TimeZone.getTimeZone("Asia/Shanghai"))
To observe the LRU cache memory footprint in the logs, configure the below properties in log4j.properties file.
log4j.logger.org.apache.carbondata.core.cache.CarbonLRUCache = DEBUG
This property will enable the DEBUG log for the CarbonLRUCache and UnsafeMemoryManager which will print the information of memory consumed using which the LRU cache size can be decided. Note: Enabling the DEBUG log will degrade the query performance. Ensure carbon.max.driver.lru.cache.size is configured to observe the current cache size.
Example:
18/09/26 15:05:29 DEBUG CarbonLRUCache: main Required size for entry /home/target/store/default/stored_as_carbondata_table/Fact/Part0/Segment_0/0_1537954529044.carbonindexmerge :: 181 Current cache size :: 0
18/09/26 15:05:30 INFO CarbonLRUCache: main Removed entry from InMemory lru cache :: /home/target/store/default/stored_as_carbondata_table/Fact/Part0/Segment_0/0_1537954529044.carbonindexmerge
Note: If Removed entry from InMemory LRU cache are frequently observed in logs, you may have to increase the configured LRU size.
To observe the LRU cache from heap dump, check the heap used by CarbonLRUCache class.
When tuning query performance, user may found that a few tasks slow down the overall query progress. To improve performance in such case, user can set spark.locality.wait and spark.speculation=true to enable speculation in spark, which will launch multiple task and get the result the one of the task which is finished first. Besides, user can also consider following configurations to further improve performance in this case.
Example:
spark.locality.wait = 500
spark.speculation = true
spark.speculation.quantile = 0.75
spark.speculation.multiplier = 5
spark.blacklist.enabled = false
Note:
spark.locality control data locality the value of 500 is used to shorten the waiting time of spark.
spark.speculation is a group of configuration, that can monitor trailing tasks and start new tasks when conditions are met.
spark.blacklist.enabled, avoid reduction of available executors due to blacklist mechanism.
Symptom
17/11/11 16:48:13 ERROR LocalFileLock: main hdfs:/localhost:9000/carbon/store/default/hdfstable/tablestatus.lock (No such file or directory)
java.io.FileNotFoundException: hdfs:/localhost:9000/carbon/store/default/hdfstable/tablestatus.lock (No such file or directory)
at java.io.FileOutputStream.open0(Native Method)
at java.io.FileOutputStream.open(FileOutputStream.java:270)
at java.io.FileOutputStream.<init>(FileOutputStream.java:213)
at java.io.FileOutputStream.<init>(FileOutputStream.java:101)
Possible Cause
If you use <hdfs path> as store path when creating carbonsession, may get the errors,because the default is LOCALLOCK.
Procedure Before creating carbonsession, sets as below:
import org.apache.carbondata.core.util.CarbonProperties
import org.apache.carbondata.core.constants.CarbonCommonConstants
CarbonProperties.getInstance().addProperty(CarbonCommonConstants.LOCK_TYPE, "HDFSLOCK")
Symptom
Thrift throws following exception :
thrift: error while loading shared libraries:
libthriftc.so.0: cannot open shared object file: No such file or directory
Possible Cause
The complete path to the directory containing the libraries is not configured correctly.
Procedure
Follow the Apache thrift docs at https://thrift.apache.org/docs/install to install thrift correctly.
Symptom
The shell prompts the following error :
org.apache.spark.sql.CarbonContext$$anon$$apache$spark$sql$catalyst$analysis
$OverrideCatalog$_setter_$org$apache$spark$sql$catalyst$analysis
$OverrideCatalog$$overrides_$e
Possible Cause
The Spark Version and the selected Spark Profile do not match.
Procedure
Ensure your spark version and selected profile for spark are correct.
Use the following command :
mvn -Pspark-2.4 -Dspark.version={yourSparkVersion} clean package
Note : Refrain from using "mvn clean package" without specifying the profile.
Symptom
Load query failed with the following exception:
Dictionary file is locked for updation.
Possible Cause
The carbon.properties file is not identical in all the nodes of the cluster.
Procedure
Follow the steps to ensure the carbon.properties file is consistent across all the nodes:
Copy the carbon.properties file from the master node to all the other nodes in the cluster. For example, you can use ssh to copy this file to all the nodes.
For the changes to take effect, restart the Spark cluster.
Symptom
Load query failed with the following exception:
Dictionary file is locked for updation.
Possible Cause
The carbon.properties file is not identical in all the nodes of the cluster.
Procedure
Follow the steps to ensure the carbon.properties file is consistent across all the nodes:
Copy the carbon.properties file from the master node to all the other nodes in the cluster. For example, you can use scp to copy this file to all the nodes.
For the changes to take effect, restart the Spark cluster.
Symptom
We get the following exception :
Cannot connect to hiveuser.
Possible Cause
The external process does not have permission to access.
Procedure
Ensure that the Hiveuser in mysql must allow its access to the external processes.
Symptom
We get the following exception on trying to connect :
Cannot read the metastore db
Possible Cause
The metastore db is dysfunctional.
Procedure
Remove the metastore db from the carbon.metastore in the Spark Directory.
Symptom
Data loading fails with the following exception :
Data Load failure exception
Possible Cause
The following issue can cause the failure :
The core-site.xml, hive-site.xml, yarn-site and carbon.properties are not consistent across all nodes of the cluster.
Path to hdfs ddl is not configured correctly in the carbon.properties.
Procedure
Follow the steps to ensure the following configuration files are consistent across all the nodes:
Copy the core-site.xml, hive-site.xml, yarn-site, carbon.properties files from the master node to all the other nodes in the cluster. For example, you can use scp to copy this file to all the nodes.
Note : Set the path to hdfs ddl in carbon.properties in the master node.
For the changes to take effect, restart the Spark cluster.
Symptom
Insertion fails with the following exception :
Data Load failure exception
Possible Cause
The following issue can cause the failure :
The core-site.xml, hive-site.xml, yarn-site and carbon.properties are not consistent across all nodes of the cluster.
Path to hdfs ddl is not configured correctly in the carbon.properties.
Procedure
Follow the steps to ensure the following configuration files are consistent across all the nodes:
Copy the core-site.xml, hive-site.xml, yarn-site, carbon.properties files from the master node to all the other nodes in the cluster. For example, you can use scp to copy this file to all the nodes.
Note : Set the path to hdfs ddl in carbon.properties in the master node.
For the changes to take effect, restart the Spark cluster.
Symptom
Execution fails with the following exception :
Table is locked for updation.
Possible Cause
Concurrency not supported.
Procedure
Worker must wait for the query execution to complete and the table to release the lock for another query execution to succeed.
Symptom
Execution fails with the following exception :
HDFS Quota Exceeded
Possible Cause
HDFS Quota is set, and it is not letting carbondata write or modify any files.
Procedure
Drop that particular index using Drop Index command so as to clear the stale folders.