Oh, I see what you mean. Yeah, I agree that having BigQueryIO use TableRow as the native format was a suboptimal decision in retrospect, and I agree that it would be reasonable to provide ability to go through Avro GenericRecord instead. I'm just not sure how to provide it in an API-compatible way - that would be particularly challenging since BigQueryIO is a beast in terms of amount of code and intermediate transforms involved. But if you have ideas, they would be welcome.
On Sat, Sep 9, 2017 at 11:18 AM Steve Niemitz <[email protected]> wrote: > Ah that makes sense wrt splitting, but is indeed confusing! Thanks for the > explanation. :) > > wrt native types and TableRow, I understand your point, but could also > argue that the raw avro records are just as "native" to the BigQuery > connector as the TableRow JSON objects, since both are directly exposed by > BigQuery. > > Maybe my use-case is more specialized, but I already have a good amount of > code that I used pre-Beam to process BigQuery avro extract files, and avro > is significantly smaller and more performant than JSON, which is why I'm > using it rather than just using TableRows. > > In any case, if there's no desire for such a feature I can always replicate > the functionality of BigQueryIO in my own codebase, so it's not a big deal, > it just seems like a feature that would be useful for other people as well. > > On Sat, Sep 9, 2017 at 1:55 PM, Reuven Lax <[email protected]> > wrote: > > > On Sat, Sep 9, 2017 at 10:53 AM, Eugene Kirpichov < > > [email protected]> wrote: > > > > > This is a bit confusing - BigQueryQuerySource and BigQueryTableSource > > > indeed use the REST API to read rows if you read them unsplit - > however, > > in > > > split() they run extract jobs and produce a bunch of Avro sources that > > are > > > read in parallel. I'm not sure we have any use cases for reading them > > > unsplit (except unit tests) - perhaps that code path can be removed? > > > > > > > I believe split() will always be called in production. Maybe not in unit > > tests? > > > > > > > > > > About outputting non-TableRow: per > > > https://beam.apache.org/contribute/ptransform-style- > > > guide/#choosing-types-of-input-and-output-pcollections, > > > it is recommended to output the native type of the connector, unless > it's > > > impossible to provide a coder for it. This is the case for > > > AvroIO.parseGenericRecords, but it's not the case for TableRow, so I > > would > > > recommend against it: you can always map a TableRow to something else > > using > > > MapElements. > > > > > > On Sat, Sep 9, 2017 at 10:37 AM Reuven Lax <[email protected]> > > > wrote: > > > > > > > Hi Steve, > > > > > > > > The BigQuery source should always uses extract jobs, regardless of > > > > withTemplateCompatibility. What makes you think otherwise? > > > > > > > > Reuven > > > > > > > > > > > > On Sat, Sep 9, 2017 at 9:35 AM, Steve Niemitz <[email protected]> > > > wrote: > > > > > > > > > Hello! > > > > > > > > > > Until now I've been using a custom-built alternative to > > BigQueryIO.Read > > > > > that manually runs a BigQuery extract job (to avro), then uses > > > > > AvroIO.parseGenericRecords() to read the output. > > > > > > > > > > I'm investigating instead enhancing the actual BigQueryIO.Read to > > allow > > > > > something similar, since it appears a good amount of the plumbing > is > > > > > already in place to do this. However I'm confused at some of the > > > > > implementation details. > > > > > > > > > > To start, it seems like there's two different read paths: > > > > > - If "withTemplateCompatibility" is set, a similar method to what I > > > > > described above is used; an extract job is started to export to > avro, > > > and > > > > > AvroSource is used to read files and transform them into TableRows. > > > > > > > > > > - However, if not set, the BigQueryReader class simply uses the > REST > > > API > > > > to > > > > > read rows from the tables. This method, I've seen in practice, has > > > some > > > > > significant performance limitations. > > > > > > > > > > It seems to me that for large tables, I'd always want to use the > > first > > > > > method, however I'm not sure why the implementation is tied to the > > > oddly > > > > > named "withTemplateCompatibility" option. Does anyone have insight > > as > > > to > > > > > the implementation details here? > > > > > > > > > > Additionally, would the community in general be accepting to > > > enhancements > > > > > to BigQueryIO to allow the final output to be something other than > > > > > "TableRow" instances, similar to how AvroIO.parseGenericRecords > > takes a > > > > > parseFn? > > > > > > > > > > Thanks! > > > > > > > > > > > > > > >
