Thanks for clarifying, Wenchen. I think that's what I expected.

As for the abstraction, here's the way that I think about it: there are two important parts of a scan: the definition of what will be read, and task sets that actually perform the read. In batch, there's one definition of the scan and one task set so it makes sense that there's one scan object that encapsulates both of these concepts. For streaming, we need to separate the two into the definition of what will be read (the stream or streaming read) and the task sets that are run (scans). That way, the streaming read behaves like a factory for scans, producing scans that handle the data either in micro-batches or using continuous tasks.

To address Jungtaek's question, I think that this does work with continuous. In continuous mode, the query operators keep running and send data to one another directly. The API still needs a streaming read layer because it may still produce more than one continuous scan. That would happen when the underlying source changes and Spark needs to reconfigure. I think the example here is when partitioning in a Kafka topic changes and Spark needs to re-map Kafka partitions to continuous tasks.


On Fri, Aug 31, 2018 at 5:12 PM Wenchen Fan <> wrote:
Hi Ryan,

Sorry I may use a wrong wording. The pushdown is done with ScanConfig, which is not table/stream/scan, but something between them. The table creates ScanConfigBuilder, and table creates stream/scan with ScanConfig. For streaming source, stream is the one to take care of the pushdown result. For batch source, it's the scan.

It's a little tricky because stream is an abstraction for streaming source only. Better ideas are welcome!

On Sat, Sep 1, 2018 at 7:26 AM Ryan Blue <> wrote:
Thanks, Reynold!

I think your API sketch looks great. I appreciate having the Table level in the abstraction to plug into as well. I think this makes it clear what everything does, particularly having the Stream level that represents a configured (by ScanConfig) streaming read and can act as a factory for individual batch scans or for continuous scans.

Wenchen, I'm not sure what you mean by doing pushdown at the table level. It seems to mean that pushdown is specific to a batch scan or streaming read, which seems to be what you're saying as well. Wouldn't the pushdown happen to create a ScanConfig, which is then used as Reynold suggests? Looking forward to seeing this PR when you get it posted. Thanks for all of your work on this!


On Fri, Aug 31, 2018 at 3:52 PM Wenchen Fan <> wrote:
Thank Reynold for writing this and starting the discussion!

Data source v2 was started with batch only, so we didn't pay much attention to the abstraction and just follow the v1 API. Now we are designing the streaming API and catalog integration, the abstraction becomes super important.

I like this proposed abstraction and have successfully prototyped it to make sure it works.

During prototyping, I have to work around the issue that the current streaming engine does query optimization/planning for each micro batch. With this abstraction, the operator pushdown is only applied once per-query. In my prototype, I do the physical planning up front to get the pushdown result, and
add a logical linking node that wraps the resulting physical plan node for the data source, and then swap that logical linking node into the logical plan for each batch. In the future we should just let the streaming engine do query optimization/planning only once.

About pushdown, I think we should do it at the table level. The table should create a new pushdow handler to apply operator pushdowm for each scan/stream, and create the scan/stream with the pushdown result. The rationale is, a table should have the same pushdown behavior regardless the scan node.


On Fri, Aug 31, 2018 at 2:00 PM Reynold Xin <> wrote:
I spent some time last week looking at the current data source v2 apis, and I thought we should be a bit more buttoned up in terms of the abstractions and the guarantees Spark provides. In particular, I feel we need the following levels of "abstractions", to fit the use cases in Spark, from batch, to streaming.

Please don't focus on the naming at this stage. When possible, I draw parallels to what similar levels are named in the currently committed api:

0. Format: This represents a specific format, e.g. Parquet, ORC. There is currently no explicit class at this level.

1. Table: This should represent a logical dataset (with schema). This could be just a directory on the file system, or a table in the catalog. Operations on tables can include batch reads (Scan), streams, writes, and potentially other operations such as deletes. The closest to the table level abstraction in the current code base is the "Provider" class, although Provider isn't quite a Table. This is similar to Ryan's proposed design.

2. Stream: Specific to streaming. A stream is created out of a Table. This logically represents a an instance of a StreamingQuery. Pushdowns and options are handled at this layer. I.e. Spark guarnatees to data source implementation pushdowns and options don't change within a Stream. Each Stream consists of a sequence of scans. There is no equivalent concept in the current committed code.

3. Scan: A physical scan -- either as part of a streaming query, or a batch query. This should contain sufficient information and methods so we can run a Spark job over a defined subset of the table. It's functionally equivalent to an RDD, except there's no dependency on RDD so it is a smaller surface. In the current code, the equivalent class would be the ScanConfig, which represents the information needed, but in order to execute a job, ReadSupport is needed (various methods in ReadSupport takes a ScanConfig).

To illustrate with pseudocode what the different levels mean, a batch query would look like the following:

val provider = reflection[Format]("parquet")
val table = provider.createTable(options)
val scan = table.createScan(scanConfig) // scanConfig includes pushdown and options
// run tasks on executors

A streaming micro-batch scan would look like the following:

val provider = reflection[Format]("parquet")
val table = provider.createTable(options)
val stream = table.createStream(scanConfig)

while(true) {
  val scan = streamingScan.createScan(startOffset)
  // run tasks on executors

Vs the current API, the above:

1. Creates an explicit Table abstraction, and an explicit Scan abstraction.

2. Have an explicit Stream level and makes it clear pushdowns and options are handled there, rather than at the individual scan (ReadSupport) level. Data source implementations don't need to worry about pushdowns or options changing mid-stream. For batch, those happen when the scan object is created.

This email is just a high level sketch. I've asked Wenchen to prototype this, to see if it is actually feasible and the degree of hacks it removes, or creates.

Ryan Blue
Software Engineer

Ryan Blue
Software Engineer