[jira] [Commented] (SQOOP-1168) Sqoop2: Incremental From/To ( formerly called Incremental Import )

["Vinoth Chandar (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SQOOP-1168?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14223865#comment-14223865 ] 

Vinoth Chandar commented on SQOOP-1168:
---------------------------------------

>> yes, I believed so that the FROM/TO state should be per job. So the Job holds FromConfig and ToConfig objects that hold some of the state information. This can be extended to keep track of the info per run related to incremental reading and writing. Note anything prefixed with M is persisted
 +1 

>> So if the HDFS connector ended up producing a new file and later reconciled it would be upto it to document what its strategy is, Much simpler for the sqoop. 
You are right.. This is more of an implementation choice in each connector.. The question really is : do we support reconciliation of deltas in the out-of-the-box HDFS connector say.. 

>> AFAIK, there is a bug in the sqoop1 implementation for the last_modified and not completely sure if it really works end-end
I have not used sqoop1. :) . But, we are definitely running into problems with the last_modified based approach, due to long running transactions.. In such cases, the approach misses records between runs, if the transaction only commits after the job has run. Is this what you are referring to ? 




> Sqoop2: Incremental From/To ( formerly called Incremental Import )
> ------------------------------------------------------------------
>
>                 Key: SQOOP-1168
>                 URL: https://issues.apache.org/jira/browse/SQOOP-1168
>             Project: Sqoop
>          Issue Type: Bug
>            Reporter: Hari Shreedharan
>            Assignee: Veena Basavaraj
>             Fix For: 1.99.5
>
>
> Initial plan is to follow roughly the same design as Sqoop 1, except provide pluggability to start this through a REST API.
> This comment applies to every sub task
> {code}
> WIP ( so do not consider this as a final design)
> {code}
> Relevant code in Sqoop 1 ( to ensure parity with tests and features ). Also note that  Sqoop 1 was less generic than Sqoop2 in terms of the from and to. The From was a SQL and TO was HDFS/ Hive...so this is no longer true in SQOOP2. It cab be from HDFS to SQL or from MongoDB to MySQL.
> https://github.com/apache/sqoop/blob/trunk/src/java/org/apache/sqoop/tool/ImportTool.java
> https://github.com/apache/sqoop/blob/trunk/src/java/org/apache/sqoop/mapreduce/MergeJob.java
> https://github.com/apache/sqoop/blob/trunk/src/java/org/apache/sqoop/util/AppendUtils.java
> https://github.com/apache/sqoop/blob/trunk/src/test/com/cloudera/sqoop/TestAppendUtils.java
> https://github.com/apache/sqoop/blob/trunk/src/test/org/apache/sqoop/manager/oracle/OracleIncrementalImportTest.java
> https://github.com/apache/sqoop/blob/trunk/src/test/com/cloudera/sqoop/TestIncrementalImport.java
> Some things I considered.
> 1. Do we need the 2 modes,simpler sequential / append case based on a simple predicate ( that can be the default) and more generic random set of records ( based on a complex predicate ). The only reason these 2 modes might make sense,  is the simple append use case can be for majority of the part be handled by the sqoop. There is very little logic to add per connector.
> The more complex case where we can have random selection of records to read and even write them out in many different ways requires a custom logic. We can certainly provide some common implementation of doing such updates, such as writing a new dataset for the delta records and the merging duplicate row data sets based on the last updated value.
> 2. What are the phases of the Incremental read/ write process for each connector look like?. Should this be independent stage in the job lifecycle complementing the Extractor or another api on the Extractor to do deltaExtract or just another config object on the FromJobConfiguration where the computed predicate string ( combination of raw predicate and the last value from the FromConfig of the job)?
> Seems like just a field on the FromJobConfiguration suffices. It can be used both in the Partitioner and Extractor.
> The FromJobConfiguration  can have a dynamically created config object in its list that is reserved for all the incremental related fields. One of the field in it will be the following predicate.
> Say one way of supporting predicate is via a proper schema like this. 
> {code}
> { 
>  type : "append"
>  columns : "id"
>  operator : ">="
>  value : "34"
> }
> {code}
> The FromJobConfiguration might have  the actual query string WHERE id > = '34' in the string format. Alternatively. If we do not go with the predicates and use loose fields like in Sqoop1, then the FromJobConfiguration will hold these fields and its values.  The FromConfig will hold the last processed value, so that it can be used in the subsequent runs to indicate from where to start.
> 3. Similarly in terms of Writing /Loading, the ToJobConfiguration will have information in it to indicate if this was a delta write. Based on the predicate type/ modes ( SEQUENTIAL( APPEND)  / RANDOM) , the loader can decide how to write the data. 
> 4. Does storing the last-processed-value in the job submission make it easier? in that case we would need 2 fields, the last_read_value for the FROM side and the last_written_value for the TO. I would consider submission, since these values are a result of job run...but still store predicate in the job config ( with the given last value )
> 5. Even the last step, very inaptly names Destroyer, after successful might need this info to finally record the values in the submission / configs and other related stats of the incremental read/ writes ..So the FROM and the TO destroyer will need to do this part.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)