Document TableFunctionSplitProcessor thread-safety

[findepi]

No description provided.

[findepi]

CI #16882

[losipiuk]

does it mean that implementation of TableFunctionSplitProcessor need to ensure memory visibility of internal data structures which may be changed by one thread and then accessed by the other. Or are we sure this is ensured by the caller?

[findepi]

formally we would call on JLS's happen-before semantics and say that there is a happens-before relation between previous method call end and the next method call... i didn't want to be very formal here. but, i wanted to indicate the implementor should not take note of things iike thread id, or use ThreadLocal internally.

[kasiafi]

Generally, to reason about thread-safety, we must consider both TableFunctionSplitProcessor and TableFunctionProcessorProvider.
The LeafTableFunctionOperator calls TableFunctionProcessorProvider.getSplitProcessor(session, handle) for each split it has. The function author implements the TableFunctionProcessorProvider and they can decide on the lifecycle of the Processor. One extreme would be to keep a single TableFunctionSplitProcessor, and return it from each call to the provider -- and deal with multiple threads. The other extreme is to instantiate a new TableFunctionSplitProcessor for each call to the provider. The latter is easy and clear, and imo should be considered the default approach.

[findepi]

The LeafTableFunctionOperator calls TableFunctionProcessorProvider.getSplitProcessor(session, handle) for each split it has.

Can it provide a split already in this method call?
it would make it clear the processor serves one split.

(i understand we wanted the leaf processor to be similar to intermediate processor, but reality is that a function implementor implements only one of them at the same time, so making things just simpler would be beneficial)

then the TableFunctionSplitProcessor just provides Pages until it's done. So becomes equivalent to ConnectorPageSource. Maybe we could reuse that interface?

[alexjo2144]

Just throwing in my 2 cents that I also found this part of the interface a bit unintuitive when reviewing @homar's CDF table function implantation, specifically that process is initially called with a Split and then continues to be called with null arguments. It's something you only need to learn once, but at first glance I was expecting this to work more like a ConnectorPageSource.

[kasiafi]

Can it provide a split already in this method call?

It totally makes sense to call TableFunctionProcessorProvider.getSplitProcessor(session, handle, split), and remove the split argument from the TableFunctionSplitProcessor.process method. It should simplify the operator a lot.

[findepi]

i will take a stab.

What about replacing TableFunctionSplitProcessor with ConnectorPageSource, as a consequence of that?

[kasiafi]

That one seems to be designed for reading input. We'd have to think about how we implement getReadTimeNanos() etc. so that it makes sense for a particular table function. Or maybe subclass to ensure that those methods cannot be used. Even though they are never used anyway.

[findepi]

#17032

[losipiuk]

hmm - this is interesting contract :)

[kasiafi]

Yup, the states were primarily designed for the TableFunctionDataProcessor.
That processor gets one portion of data at a time, and it declares isUsedInput() when the portion is ingested. When it has ingested all the due data, it gets null until it's finished.
In case of the TableFunctionSplitProcessor, instead of input data, we have a Split. It is presented to the Processor as a single portion of data. If the processor declares isUsedInput(), it gets null in subsequent calls until it's finished.