search

apache / datafusion

Apache DataFusion SQL Query Engine
https://datafusion.apache.org/
Apache License 2.0
5.5k stars 1.02k forks source link

[EPIC] Streaming partitioned writes #6569

Open alamb opened 1 year ago

alamb commented 1 year ago

Is your feature request related to a problem or challenge?

This is a tracking epic for a collection of features related to writing data.

The basic idea is better / full support for writing data:

  1. to multiple (possibly Partitoned by value) files
  2. To different file types (parquet, csv, json, avro, arrow)
  3. In a streaming fashion (input doesn't need to be entirely buffered)
  4. From SQL (via INSERT, INSERT INTO, COPY, Etc)
  5. Stream to a target object_store (aka multi-part S3 upload)

This is partially supported today programmatically (see SessionContext::write_csv, etc)

Subtasks:

alamb commented 10 months ago

FYI @devinjdangelo I updated this ticket with various issues related to the write code you are working on

devinjdangelo commented 10 months ago

@alamb I opened #7298 to track improving statement level options/overrides relevant to this epic.

alamb commented 10 months ago

@alamb I opened https://github.com/apache/arrow-datafusion/issues/7298 to track improving statement level options/overrides relevant to this epic.

Thanks @devinjdangelo -- I added it to the list

devinjdangelo commented 10 months ago

An additional issue we should cut and add to this epic is allowing inserts to a sorted ListingTable. In the case of appending new files to a directory, I think it is as simple as having FileSinkExec require its input be sorted.

It can't really be supported efficiently for Append to existing file since it would require reading the existing file, sorting with the new data and rewriting the whole file. For this case, you could use insert overwrite instead if you really want to do this (which is another thing which we could cut a ticket to add support for).

Alternatively, we could have a check to see if 1) the table is sorted and 2) the input to FileSinkExec is sorted. If 1) is true but 2) is not, we would need to update the metadata about the table to indicate for subsequent queries it is no longer guaranteed to be sorted.

alamb commented 10 months ago

Filed https://github.com/apache/arrow-datafusion/issues/7354 to track

devinjdangelo commented 10 months ago

@alamb I opened #7442 to track adding support for parquet column level settings via SQL options.

alamb commented 10 months ago

@alamb I opened https://github.com/apache/arrow-datafusion/issues/7442 to track adding support for parquet column level settings via SQL options.

Thanks @devinjdangelo -- I added it to the list on this ticket

devinjdangelo commented 9 months ago

I added #7679 to track adding avro support for writes @alamb

devinjdangelo commented 9 months ago

@alamb I made some progress on inserts to sorted tables https://github.com/apache/arrow-datafusion/issues/7354

This also got me thinking about inserts to partitioned tables, so I opened issue to track: https://github.com/apache/arrow-datafusion/issues/7744

Lastly, I've been thinking we may want to deprecate and eventually remove the SessionContext write methods or alternatively hook them into FileSinkExec similarly to how the data frame write methods are set up. https://github.com/apache/arrow-datafusion/blob/0408c2b1596417ba55a636fa3c8a601ffbdb0e60/datafusion/core/src/execution/context.rs#L1271-L1278

alamb commented 9 months ago

This also got me thinking about inserts to partitioned tables, so I opened issue to track: https://github.com/apache/arrow-datafusion/issues/7744

Thank you -- I added https://github.com/apache/arrow-datafusion/issues/7744 to the list on this ticket

Lastly, I've been thinking we may want to deprecate and eventually remove the SessionContext write methods or alternatively hook them into FileSinkExec similarly to how the data frame write methods are set up.

That sounds like a reasonable idea to me. One challenge might be that the SessionContext::write_csv take an ExecutionPlan where the apis on DataFrame require a LogicalPlan.

Hooking them into FileSinkExec sounds like a great idea to consolidate the code. Shall I file a ticket for it?

devinjdangelo commented 9 months ago

@alamb I wrote up an issue describing what we discussed on #7743 regarding empty files being written out and some potential solutions.

https://github.com/apache/arrow-datafusion/issues/7767