A rather straight option is allow user to supply global dictionary generated 
somewhere else or we build a separate tool just for generating as well updating 
dictionary. Then the general normal data loading process will encode columns 
with local dictionary if not supplied.  This should cover majority of cases for 
low-medium cardinality column. For the cases we have to incorporate online 
dictionary update, use a lock mechanism to sync up should serve the purpose. 

In another words, generating global dictionary is an optional step, only 
triggered when needed, not a default step as we do currently.


-----Original Message-----
From: Ravindra Pesala [mailto:ravi.pes...@gmail.com] 
Sent: Tuesday, October 11, 2016 2:33 AM
To: dev
Subject: Discussion(New feature) regarding single pass data loading solution.

Hi All,

This discussion is regarding single pass data load solution.

Currently data is loading to carbon in 2 pass/jobs
 1. Generating global dictionary using spark job.
 2. Encode the data with dictionary values and create carbondata files.
This 2 pass solution has many disadvantages like it needs to read the data
twice in case of csv files input or it needs to execute dataframe twice if
data is loaded from dataframe.

In order to overcome from above issues of 2 pass dataloading, we can have
single pass dataloading and following are the alternate solutions.

Use local dictionary
 Use local dictionary for each carbondata file while loading data, but it
may lead to query performance degradation and more memory footprint.

Use KV store/distributed map.
*HBase/Cassandra cluster : *
  Dictionary data would be stored in KV store and generates the dictionary
value if it is not present in it. We all know the pros/cons of Hbase but
following are few.
  Pros : These are apache licensed
         Easy to implement to store/retreive dictionary values.
         Performance need to be evaluated.

  Cons : Need to maintain seperate cluster for maintaining global

*Hazlecast distributed map : *
  Dictionary data could be saved in distributed concurrent hash map of
hazlecast. It is in-memory map and partioned as per number of nodes. And
even we can maintain the backups using sync/async functionality to avoid
the data loss when instance is down. We no need to maintain seperate
cluster for it as it can run on executor jvm itself.
  Pros: It is apache licensed.
        No need to maintain seperate cluster as instances can run in
executor jvms.
        Easy to implement and store/retreive dictionary values.
        It is pure java implementation.
        There is no master/slave concept and no single point failure.

  Cons: Performance need to be evaluated.

*Redis distributed map : *
    It is also in-memory map but it is coded in c language so we should
have java client libraries to interact with redis. Need to maintain
seperate cluster for it. It also can partition the data.
  Pros : More feature rich than Hazlecast.
         Easy to implement and store/retreive dictionary values.
  Cons : Need to maintain seperate cluster for maintaining global
         May not be suitable for big data stack.
         It is BSD licensed (Not sure whether we can use or not)
  Online performance figures says it is little slower than hazlecast.

Please let me know which would be best fit for our loading solution. And
please add any other suitable solution if I missed.
Thanks & Regards,

Reply via email to