[jira] [Comment Edited] (SQOOP-1771) Investigation CSV IDF FORMAT of the Array/NestedArray/ Set/ Map in Postgres and HIVE.

["Veena Basavaraj (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SQOOP-1771?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14221479#comment-14221479 ] 

Veena Basavaraj edited comment on SQOOP-1771 at 11/21/14 10:17 PM:
-------------------------------------------------------------------

Postgres format is very similar to JSON  but just needs more hand-rolling, instead of relying on a standard JSON library for both arrays and map,
Between hand-rolling and actually using jackson object mapper, the performance differences are highly unlikely to be different.
I would still prefer using a standard JSON library for encoding maps and nested arrays, so that the connectors can use the same standard as well.

Note this format is only going to help one data source that is post gres, everyone else still have to convert it to this format before they do setText or have to convert this back to their own format when they do getTex.

Unless there is concrete metrics it is hard to tell how many use cases will benefit from the post gres way of encoding.

If the group is still favoring the post gres representation, express your votes.

Maps and nested arrays would be
{code}
'{key:{value}}', '{{4,5,6},{7,8,9}}'

value in itself can be {4,5,6},{7,8,9}

No spaces, but just commas and everything is quoted

 You can put double quotes around any element value, and must do so if it contains commas or curly braces.!! so this has to handrolled as well! which jakson would do so automatically.
{code}


was (Author: vybs):
Postgres format is very similar to JSON  but just needs more hand-rolling, instead of relying on a standard JSON library for both arrays and map,
Between hand-rolling and actually using jackson object mapper, the performance differences are highly unlikely to be different.
I would still prefer using a standard JSON library for encoding maps and nested arrays, so that the connectors can use the same standard as well.

Note this format is only going to help one data source that is post gres, everyone else still have to convert it to this format before they do setText or have to convert this back to their own format when they do getTex.

Unless there is concrete metrics it is hard to tell how many use cases will benefit from the post gres way of encoding.

If the group is still favoring the post gres representation, express your votes.

Maps and nested arrays would be
{code}
'{key:{value}}', '{{4,5,6},{7,8,9}}'

value in itself can be {4,5,6},{7,8,9}

No spaces, but just commas and everything is quoted

 You can put double quotes around any element value, and must do so if it contains commas or curly braces.!! so this has to handrolled as well! which jaksocn would do so automatically.
{code}

> Investigation CSV IDF FORMAT of the Array/NestedArray/ Set/ Map in Postgres and HIVE.
> -------------------------------------------------------------------------------------
>
>                 Key: SQOOP-1771
>                 URL: https://issues.apache.org/jira/browse/SQOOP-1771
>             Project: Sqoop
>          Issue Type: Sub-task
>          Components: sqoop2-framework
>            Reporter: Veena Basavaraj
>             Fix For: 1.99.5
>
>
> https://cwiki.apache.org/confluence/display/SQOOP/Sqoop2+Intermediate+representation#Sqoop2Intermediaterepresentation-Intermediateformatrepresentationproposal
> The above document does not explicitly say the design goals for choosing the CSV IDF format for different types but with conversation on of the related tickets  RB : https://reviews.apache.org/r/28139/diff/#. Here are the considerations.
> Intermediate Data Format is more relevant when we transfer data between the FROM and TO and both do not agree on the same form of data as it is transferred via sqoop.
> The IDF API as of today exposes 3 types of setter, one for a generic type T, one for Text/String, one for object array.  
> {code}
>   /**
>    * Set one row of data. If validate is set to true, the data is validated
>    * against the schema.
>    * @param data - A single row of data to be moved.
>    */
>   public void setData(T data) {
>     this.data = data;
>   }
>   /**
>    * Get one row of data.
>    *
>    * @return - One row of data, represented in the internal/native format of
>    *         the intermediate data format implementation.
>    */
>   public T getData() {
>     return data;
>   }
>   /**
>    * Get one row of data as CSV.
>    *
>    * @return - String representing the data in CSV, according to the "FROM" schema.
>    * No schema conversion is done on textData, to keep it as "high performance" option.
>    */
>   public abstract String getTextData();
>   /**
>    * Set one row of data as CSV.
>    *
>    */
>   public abstract void setTextData(String text); 
> /**
>    * Get one row of data as an Object array.
>    *
>    * @return - String representing the data as an Object array
>    * If FROM and TO schema exist, we will use SchemaMatcher to get the data according to "TO" schema
>    */
>   public abstract Object[] getObjectData();
>   /**
>    * Set one row of data as an Object array.
>    *
>    */
>   public abstract void setObjectData(Object[] data);
>   /**
> {code} 
> NOTE : the java docs are not completely accurate, there is really no validation happening:). Second CSV in one way the IDF can be represented when it is TEXT.There can be other implementation of IDF as well such as AVRO or JSON, very similar to the serDe interface in HIVE that allows custom ways to store data, but in SQOOP it is custom ways to represent data as it flows vis SQOOP.  Another java doc fix... " String representing the data in CSV, according to the "FROM" schema.  * No schema conversion is done on textData, to keep it as "high performance" option.", this also is not accurate. The CSV format is a standard enforced by sqoop implementation, there is no one STANDARD CSV for all data types esp with nested types.. The FROM schema does not enforce any standard..
> Anyways, so the design considerations for the CSV IDF implementation seems to be the following. As I said before other IDF implementation can have other design goals and can be chosen by a particular connector to benefits data in and out of itself the most.
> 1. the setText/ getText are supposed to allow the FROM and TO to talk the same language and hence should have very minimal transformations as the data flows through SQOOP. This means that both FROM and TO agree to give data in the CSV IDF that is standardized in the wiki / spec/ docs and the read data in the same format. Transformation may have to happen before the setText() or after the getText, but nothing will happen in between when it flows through sqoop. If the FROM does a setText and the TO does a getObject then there is time spent it converting the elements within the CSV string to actual java objects. This means there is parsing and unescaping / unencoding happening in sqoop.
> 2. The current proposal seems to recommend the formats that are more prominent with the databases that have been explored in the list, but it is not really a complete set of all data sources/connectors sqoop may have in future. Most emphasis is on the relational DB stores since historically sqoop1 only supported that as the FROM source
> https://cwiki.apache.org/confluence/display/SQOOP/Sqoop2+Intermediate+representation#Sqoop2Intermediaterepresentation-Intermediateformatrepresentationproposal
> But overall the goal seem to be more on the side of sql dump and pg dump that use CSV format and the hope is such transfers in sqoop will happen more.
> 3. Avoiding any CPU cycles, there is no validation that will done to make sure that the data adheres to the CSV format. It is trust based system that the incoming data will follow the CSV rules as depicted in the link above 
> https://cwiki.apache.org/confluence/display/SQOOP/Sqoop2+Intermediate+representation#Sqoop2Intermediaterepresentation-Intermediateformatrepresentationproposa
> Next, having know these design goals, the format to encode the nested arrays and maps can be done in some ways. 
> 2 examples were explored below. HIVE and postgres. Details are given below in comments. One of the simplest ways was to use the universal JSON jackson api for nested arrays and maps.
> Postgres format  is very similar to that but just needs more hand-rolling instead of relying on a standard JSON library.  both for arrays and map, this format can be used as a standard.  Between this and actually using jackson object mapper, the performance differences are highly unlikely to be different. 
> I would still prefer using a standard JSON library for encoding maps and nested arrays, so that the connectors can  use the same standard as well.



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