Skip to main content



Apache Kudu comparison with Hive (HDFS Parquet) with Impala & Spark


Need

  1. Tight integration with Apache Impala, making it a good, mutable alternative to using HDFS with Apache Parquet.
  2. High availability like other Big Data technologies.
  3. Structured Data Model.
  4. Unlike Hadoop- which is Schema on Read, Kudu is Schema on write. That is following should be known before hand, - 
    1. Primary key,
    2. Data Types of columns
    3. Partition columns, etc.
  5. It is considered as bridging gap between Hive & HBase.
  6. Can integrate with Hive Meta store.

Note

  • Following document is prepared –
    • Not considering any future Cloudera Distribution Upgrades.
    • Considering, we have 2.2.0.cloudera2, Hive 1.1.0-cdh5.12.2, Hadoop 2.6.0-cdh5.12.2
    • Kudu is just supported by Cloudera. So, we consider that, we will have an ongoing Cloudera Cluster.
Kudu is considered good for – 
  1. Reporting applications where new data must be immediately available for end users.
  2. Time-series applications that must support queries across large amounts of historic data while simultaneously returning granular queries about an individual entity.
  3. Applications that use predictive models to make real-time decisions, with periodic refreshes of the predictive model based on all historical data.
Limitations –

  • HDFS- It is a data store that keeps data on multiple Edge Nodes. We use execution engines like MR, Spark, and Tez to access data. So, the process can run on any data node. Kudu – Will limit read/ write access to lead tablet server
  • Data locality – A MR JOB / SPARK JOB is good if it is co-located with the Node that serves data. So, if there are too many Jobs running at same time than it can be a bottleneck. Since, there is one lead tablet server at given moment of time.
  • As said above, Hortonworks does not support Kudu. A question to answer will be do we want to incur cost of Cloudera license (, if any).
  • Primary key has to be present and can't be altered once defined for a table.
  • Kudu tables can have a maximum of 300 columns.
  • No individual cell may be larger than 64KB before encoding or compression. The cells making up a composite key are limited to a total of 16KB after the internal composite-key encoding done by Kudu.
  • Complex data types are not supported. Array, Map, Structured
  • As pointed above, it is schema on write vs Hive which is schema on read.
  • Secondary indexes are not supported.
  • Multi-row transactions are not supported.
  • Relational features, such as foreign keys, are not supported.
  • Impala cannot create Kudu tables with VARCHAR or nested-typed columns.
  • Impala- “!= “and LIKE predicates are not pushed to Kudu, and instead will be evaluated by the Impala scan node.
  • There is limitation of partitioning. It is not automatic.
  • Recommended maximum number of tablet servers is 100.
  • Recommended maximum number of masters is 3.
  • Maximum number of tablets per table is 60, per tablet server, at table-creation time.
  • Tablet servers cannot change their address or port.
  • Spark - <> and OR predicates are not pushed to Kudu, and instead will be evaluated by the Spark task. Only LIKE predicates with a suffix wildcard are pushed to Kudu. This means LIKE "FOO%" will be pushed, but LIKE "FOO%BAR" won't.
  • Impala gives poor performance for multiple inserts. But, is good for one insert with multi-value clause.
  • Impala – On insert, if primary key is present then it will print warning but not fail.



Performance comparison
Insert 100000000 rows to Hive & Kudu Table using Spark.
Hive (ORC, 18 Partition, bloom filter on ID)
Kudu (18 hash partition)
Hive(Parquet)
103314 ms
195357 ms
105608 ms

Random Access Performance - Select a key randomly using Impala.
**Impala doesn’t support ORC tables
Id Primary key
Hive (Parquet)
Kudu
1
0.41s
0.17s
2
0.19s
0.16s
3
0.24s
0.21s
4
0.24s
0.17s
5
0.23s
0.07s



Insert a key into table using Impala.
Id Primary key
Hive (Parquet)
Kudu
1
0.32s
0.11s
2
0.21s
0.11s
3
0.11s
0.11s
4
0.21s
0.11s
5
0.11s
0.11s

Update/Delete a key into table using Impala.
**Impala does not support modifying a non-Kudu table
Id Primary key
Kudu (Update)
Kudu (Delete)
1
0.14s
0.21s
2
0.14s
0.19s
3
0.22s
0.14s
4
0.15s
0.19s
5
0.22s
0.14s

Select random keys using Spark
Hive (Parquet)
Hive (ORC)
Kudu
4172ms
10040ms
2228ms
   
     **Spark cannot be used to update/ delete rows (ACID).
     **Hive supports update/delete rows with ORC file format.
    
   Join performance with Spark –
    1)      Join Hive with Hive on primary key and save results into Hive table.
    2)      Join Hive with Hive on non-primary key column and save results to Hive table.
    3)      Join Hive with Kudu on primary key and save results into Hive table.
    4)      Join Hive with kudu on non-primary key column and save results into Hive table.
        1
       2
      3
       4
      114256ms
      220895ms
     128081ms
      Notes: refer below

     Note- Joining Kudu table on non-primary key column is a very expensive operation as it will scan complete table and all tablet servers. And, it may fail due to out of Memory.

   Performance of Analytical queries from Impala-
   
    Conclusion
    Kudu has its own pros and cons. It is not to replace HDFS or HBase or Hive. It has its own significance and is wroth using depending upon the use case.

    Refer
    https://github.com/dinesh028/engineering/blob/master/resources/samples/Code%20used%20for%20testing.txt

Comments

Post a Comment

Popular posts

Spring MongoDB Rename field with derived Value of another field

Input Collection -  [ { 'k' : 'Troubleshooting' , 'hour' : '2024-10-10T16' , 'v' : [ 'WebPage, Login' ] }, { 'k' : 'TroubleshootingMe' , 'hour' : '2024-10-07T01' , 'v' : [ 'Accounts, etc' ] }  ] Expected Output -  [ { 'hour' : '2024-10-10T16' , 'Troubleshooting' : [ 'WebPage, Login' ] }, { 'hour' : '2024-10-07T01' , 'TroubleshootingMe' : [ 'Accounts, etc' ] }  ]   Above Can be achieved by  $replaceRoot / $replaceWith as follows - { $replaceWith : { $mergeObjects : [ { hour : "$hour" }, { "$arrayToObject" : [ [ { k : "$k" , v : "$v" } ] ] } ] } } or { $replaceRoo...
Post a Comment
Read more




Spark MongoDB Connector Not leading to correct count or data while reading

  We are using Scala 2.11 , Spark 2.4 and Spark MongoDB Connector 2.4.4 Use Case 1 - We wanted to read a Shareded Mongo Collection and copy its data to another Mongo Collection. We noticed that after Spark Job successful completion. Output MongoDB did not had many records. Use Case 2 -  We read a MongoDB collection and doing count on dataframe lead to different count on each execution. Analysis,  We realized that MongoDB Spark Connector is missing data on bulk read as a dataframe. We tried various partitioner, listed on page -  https://www.mongodb.com/docs/spark-connector/v2.4/configuration/  But, none of them worked for us. Finally, we tried  MongoShardedPartitioner  this lead to constant count on each execution. But, it was greater than the actual count of records on the collection. This seems to be limitation with MongoDB Spark Connector. But,  MongoShardedPartitioner  seemed closest possible solution to this kind of situation. But, it per...
Post a Comment
Read more




Experience with MongoDB and Optimizations

  Experience with MongoDB and Optimizations Before reading below. I would like to point out that this  experience  is related to version  6.0.14-ent, having 6 shards, each shard having 3 machines, each machine is VM with 140 GB RAM and 2TB SSD. And, we had been hosting almost 36 TB of data. MongoDB is not good with Big Data Joins and/ or Big Data OLAP processing. It is mainly meant for OLTP purposes.  Instead of joining millions of keys between 2 collections. It is better to lookup data of one key from one collection then lookup it in other collection. Thus, merging data from 2 collection for same key. Its better to keep De-normalized data in one document.  Updating a document later is cumbersome.  MongoDB crash if data is overloaded. And, it has long downtime if crashed unlike other databases which fails write to database if disk space achieves certain limit. Thus, keeping database active and running for read traffic. MongoDB needs indexes for fast qu...
Post a Comment
Read more




Spring MongoDB Log Connection Pool Details - Active, Used, Waiting

  We couldn't find any direct way to log Mongo Connection pool Size. So, we did implement an indirect way as below.  This may be incorrect at times when dealing with Sharded MongoDB having Primaty & Secondary nodes. Because, connection may be used based on read prefrence - Primary, primaryPreferred, Secondary, etc. But, this gives an understanding if connections are used efficiently and there is no wait to acquire connections from pool. This can be further enhanced to log correct connection pool statistics.  1) Implement  MyConnectionPoolListener  as below -  import java.util.concurrent.atomic.AtomicInteger; import org.slf4j.Logger; import org.slf4j.LoggerFactory; import com.mongodb.event.ConnectionCheckOutFailedEvent; import com.mongodb.event.ConnectionCheckOutStartedEvent; import com.mongodb.event.ConnectionCheckedInEvent; import com.mongodb.event.ConnectionCheckedOutEvent; import com.mongodb.event.ConnectionClosedEvent; import com.mongodb.event.Conne...
Post a Comment
Read more




Spark Streaming with Kafka Leading to increase in Open File Descriptors ( Kafka )

  Open File Descriptors w.r.t Kafka brokers relates with following -  number of file descriptors to just track log segment files. Additional file descriptors to communicate via network sockets with external parties (such as clients, other brokers, Zookeeper, and Kerberos). For # 1 this is formula -  (number of partitions)*(partition size / segment size) Reference -  https://docs.cloudera.com/cdp-private-cloud-base/7.1.6/kafka-performance-tuning/topics/kafka-tune-broker-syslevel-file-descriptors.html For #2, every connection made my consumer or producer or zookeeper or  Kerberos  opens file descriptors. Note that each TCP connection creates 2 file descriptors. These connections can be for internal communication of heartbeat, or  security handshake , or data transfer to or from client (producer or consumer) When we run a Spark application integrating it with  Kafka . And, if it is not stable, meaning -  Streaming window for micro batches is les...
Post a Comment
Read more