[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16947794#comment-16947794 ] Rafael commented on SPARK-15689: Could you clarify whether 4 item from the original list was implemented: _4. Nice-to-have: support additional common operators, including limit and sampling._ I wondering whether limit operation is optimized now, pushed to `PushedFilters: ` > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Fix For: 2.3.0 > > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian Jira (v8.3.4#803005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16636307#comment-16636307 ] Geoff Freeman commented on SPARK-15689: --- Thanks Wenchen. I was hoping that we might be able to somehow plug into the bucketing that Spark is doing. I'll keep looking into ways to plug in. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Fix For: 2.3.0 > > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[
https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16634871#comment-16634871
]
Wenchen Fan commented on SPARK-15689:
-
So {{SupportsReportPartitioning}} is not powerful enough to support custom hash
functions yet.
There are two major operators that may introduce shuffle: join and aggregate.
Aggregate only needs to have the data clustered, but doesn't care how, so the
data source v2 can support it, if your implementation catches
{{ClusteredDistribution}}. Join needs the data of the 2 children clustered by
the spark shuffle hash function, which is not supported by data source v2
currently.
> Data source API v2
> --
>
> Key: SPARK-15689
> URL: https://issues.apache.org/jira/browse/SPARK-15689
> Project: Spark
> Issue Type: New Feature
> Components: SQL
>Affects Versions: 2.3.0
>Reporter: Reynold Xin
>Assignee: Wenchen Fan
>Priority: Major
> Labels: SPIP, releasenotes
> Fix For: 2.3.0
>
> Attachments: SPIP Data Source API V2.pdf
>
>
> This ticket tracks progress in creating the v2 of data source API. This new
> API should focus on:
> 1. Have a small surface so it is easy to freeze and maintain compatibility
> for a long time. Ideally, this API should survive architectural rewrites and
> user-facing API revamps of Spark.
> 2. Have a well-defined column batch interface for high performance.
> Convenience methods should exist to convert row-oriented formats into column
> batches for data source developers.
> 3. Still support filter push down, similar to the existing API.
> 4. Nice-to-have: support additional common operators, including limit and
> sampling.
> Note that both 1 and 2 are problems that the current data source API (v1)
> suffers. The current data source API has a wide surface with dependency on
> DataFrame/SQLContext, making the data source API compatibility depending on
> the upper level API. The current data source API is also only row oriented
> and has to go through an expensive external data type conversion to internal
> data type.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16634650#comment-16634650 ] Geoff Freeman commented on SPARK-15689: --- I'm having trouble figuring out how to expose a custom hash function from my DataSourceV2. I'm trying to implement SupportsReportPartitioning, but I don't see how I can convert the physical.Partitioning that's required by outputPartitioning() into a HashPartitioning. I'd like to figure out how to pass this to DataSourceScanExec so that we can avoid shuffles. [~rxin] [~cloud_fan] Thanks! > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Fix For: 2.3.0 > > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16549724#comment-16549724 ] Reynold Xin commented on SPARK-15689: - Is there an umbrella ticket on improving the dsv2 api? If yes, please link to it from this Jira ticket. If not, we should create one. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Fix For: 2.3.0 > > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16234899#comment-16234899 ] Russell Spitzer commented on SPARK-15689: - I think knowing whether or not the count was occurring at the time of the pushdown would solve this. So aggregate pushdown is probably the cleanest solution. We have a similar problem to Spark, we can handle the pushdown efficently if we know it's a count that completely be handled by our filters but not if it isn't. Unfortunately we can see if we can satisfy all the filters when "unhandled filters" is called but don't know if the plan requires any additional columns. So really we would be ok if we just got the required "output" at that time. If we satisfy all the predicates and we know the output is empty we could handle the count pushdown. Having a specific Aggregate pushdown is probably cleaner though. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16234892#comment-16234892 ] Wenchen Fan commented on SPARK-15689: - Spark wants to get `unhandledFilters` first so that it can decide which column needs to be read. There may be a column only referred by a filter and if that filter is pushed down, Spark doesn't need to read that column. Do you have a strong requirement to see required columns before calling `unhandledFilters`? Can aggregate push down satisfy your case? > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16234820#comment-16234820 ] Russell Spitzer commented on SPARK-15689: - It does not, we can tell that a count (or similar request) occurs by seeing that no columns are requested from the underlying source. It the current api you can see this at the call to "buildScan" but unhandledFilters requires knowing the filtering without specifying the columns required. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16234811#comment-16234811 ] Wenchen Fan commented on SPARK-15689: - how would a count(agg function) exist in filter expression? > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16234640#comment-16234640 ] Russell Spitzer commented on SPARK-15689: - Something I just noticed, it may be helpful to also pass in "required columns" to the supportPushdownFilter. This could enable systems which can quickly estimate counts (but not actually materialize records) to respond to certain filters. In my example system, I can pick out a small number of records past on an index very quickly, but as the number of records as proportion of the data increases the usefulness of using the pushdown decreases and eventually becomes a hinderance on performance. With counts in particular it is almost always beneficial to use the index since no rows are returned but I cannot tell if a count is being preformed from the base supportsPushdown add in. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan >Priority: Major > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16224224#comment-16224224 ] Wenchen Fan commented on SPARK-15689: - ah missed this one, reopen this ticket. My concern is that, follow-ups should not block 2.3 release, while the basic data source v2 infrastructure should > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan > Labels: SPIP, releasenotes > Fix For: 2.3.0 > > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16224122#comment-16224122 ] Reynold Xin commented on SPARK-15689: - Why not put all of them as subtasks here? Also https://issues.apache.org/jira/browse/SPARK-22078 is not done. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan > Labels: SPIP, releasenotes > Fix For: 2.3.0 > > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16223988#comment-16223988 ] Wenchen Fan commented on SPARK-15689: - The basic read/write interfaces are done, I'm resolving this ticket and track the follow-ups in https://issues.apache.org/jira/browse/SPARK-22386 > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin >Assignee: Wenchen Fan > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16153834#comment-16153834 ] Apache Spark commented on SPARK-15689: -- User 'cloud-fan' has created a pull request for this issue: https://github.com/apache/spark/pull/19136 > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16153709#comment-16153709 ] Yanbo Liang commented on SPARK-15689: - [~cloud_fan] Great doc! I like the direction this is going in, especially for columnar read interface, schema inference interface and schema evolution. Thanks. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Affects Versions: 2.3.0 >Reporter: Reynold Xin > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16145309#comment-16145309 ] Reynold Xin commented on SPARK-15689: - That seems like an issue orthogonal to the API described here. Also I don't think we should break the old API. V2 can be added alongside v1. The Hive schema issue is due to Spark ignoring the Hive schema, because the Hive metastore (at least in the past) has been buggy and doesn't always accept valid data types specified in Spark. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16145304#comment-16145304 ] Thomas Graves commented on SPARK-15689: --- I haven't had a chance to look at this, someone just pointed this jira out to me due to SPARK-21841. If we are changing the api here can we make sure that it has hive compatibility or at least a way for users to specify to use the hive schema and cover SPARK-21841? > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: SPIP, releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16139479#comment-16139479 ] Wenchen Fan commented on SPARK-15689: - yea, `LogicalPlan` is an internal concept and we can't use it in a stable API. For making data source instances immutable, one problem is that, immutability can not be guaranteed by java/scala, we can only document it and ask users to follow, so we can't be 100% safe anyway. Another point is, I think implementing data source immutably is hard, especially when a data source implements many push down interfaces, as it needs to copy around many states. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16138607#comment-16138607 ] Reynold Xin commented on SPARK-15689: - Not the author but my guess is that the other approach wouldn't be stable, neither source nor binary. It relies on internal logical plans. Note that it would be possible to incorporate the other approach in this as well, if we add a function that takes in a logical plan and returns a logical plan. Wenchen, one thing I have been thinking is whether we should make such user implementations of data sources immutable. That is, all methods (e.g. Push filters) would return a new instance of the source. It would be safer. I don't know if it is worth the complexity though. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16138509#comment-16138509 ] Andrew Ash commented on SPARK-15689: Can the authors of this document add a section contrasting the approach with the one from https://issues.apache.org/jira/browse/SPARK-12449 ? In that approach the data source receives an entire arbitrary plan rather than just parts of the plan. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16137998#comment-16137998 ] Liang Chen commented on SPARK-15689: Cool! Very look forward to API V2 > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16137985#comment-16137985 ] Wenchen Fan commented on SPARK-15689: - yea, we plan to. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16137982#comment-16137982 ] Liang Chen commented on SPARK-15689: Thanks for the doc. all these features will be considered in Spark 2.3 ? > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16130863#comment-16130863 ] Wenchen Fan commented on SPARK-15689: - good doc attached! > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16130494#comment-16130494 ] Dongjoon Hyun commented on SPARK-15689: --- Thank you for the document, too! > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16130355#comment-16130355 ] Russell Spitzer commented on SPARK-15689: - Thanks [~cloud_fan] for posting the design doc it was a great read and I like a lot of the direction this is going in. It would helpful if we could have access to the doc as a google doc or some other editable/comment-able form though to encourage discussion. I left some comments on the prototype but one thing I think could be a great addition would be a joinInterface. I ended up writing up one of these specifically for Cassandra and had to do a lot of plumbing to get it to fit into the rest of the Catalyst ecosystem so I think this would be a great time to plan ahead in Spark design. The join interface would look a lot like a combination of the read and write apis, given a row input and a set of expressions the relationship should return rows that match those expressions OR fallback to just being a read relationship if none of the expressions can be satisfied by the join (leaving all the expressions to be evaluated in spark). > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > Attachments: SPIP Data Source API V2.pdf > > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-15689) Data source API v2
[ https://issues.apache.org/jira/browse/SPARK-15689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16052321#comment-16052321 ] Russell Spitzer commented on SPARK-15689: - I've been trying to work with making Catalyst Cassandra partitioning aware. There seem to be two major blocks on this. The first is that DataSourceScanExec is unable to learn what the underlying partitioning should be from the BaseRelation it comes from. I'm currently able to get around this by using the DataSourceStrategy plan and then transforming the resultant DataSourceScanExec. The second is that the Partitioning trait is sealed. I want to define a new partitioning which is Clustered but is not hashed based on certain columns. It would look almost identical to the HashPartitioning class except the expression which returns a valid PartitionID given expressions would be different. So for V2 I would really like the ability to specify the physical partitioning and as well be able to define new custom partitioning. > Data source API v2 > -- > > Key: SPARK-15689 > URL: https://issues.apache.org/jira/browse/SPARK-15689 > Project: Spark > Issue Type: New Feature > Components: SQL >Reporter: Reynold Xin > Labels: releasenotes > > This ticket tracks progress in creating the v2 of data source API. This new > API should focus on: > 1. Have a small surface so it is easy to freeze and maintain compatibility > for a long time. Ideally, this API should survive architectural rewrites and > user-facing API revamps of Spark. > 2. Have a well-defined column batch interface for high performance. > Convenience methods should exist to convert row-oriented formats into column > batches for data source developers. > 3. Still support filter push down, similar to the existing API. > 4. Nice-to-have: support additional common operators, including limit and > sampling. > Note that both 1 and 2 are problems that the current data source API (v1) > suffers. The current data source API has a wide surface with dependency on > DataFrame/SQLContext, making the data source API compatibility depending on > the upper level API. The current data source API is also only row oriented > and has to go through an expensive external data type conversion to internal > data type. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
