GH-36092: [C++] Simplify concurrency in as-of-join node

[rtpsw]

What changes are included in this PR?

The key hasher invalidation and memo-store time updating are moved to the processing thread.

Are these changes tested?

Yes, by existing tests.

Are there any user-facing changes?

No.

• Closes: [C++] Simplify concurrency in as-of-join node #36092

[rtpsw]

cc @westonpace, @icexelloss

[rtpsw]

cc @westonpace, @icexelloss

[westonpace]

This makes sense to me. Let me verify I understand:

Before, on InputReceived, we would update the memo for an input based on the new batch.

Now, the only thing that really happens on InputReceived, is the batch is placed into the appropriate input queue. The update to the memo table happens when teh processing thread runs, as it is processing batches for an input.

[westonpace]

Suggested change

	bool update = current_time_ < ts;
	if (update) current_time_ = ts;
	current_time_ = std::min(current_time_, ts);

Minor nit: Maybe simpler?

[rtpsw]

The method still needs to return a Boolean, so we can't remove the bool update line, but we can replace the if-statement.

[icexelloss]

This makes sense to me. Let me verify I understand:

Before, on InputReceived, we would update the memo for an input based on the new batch.

Now, the only thing that really happens on InputReceived, is the batch is placed into the appropriate input queue. The update to the memo table happens when teh processing thread runs, as it is processing batches for an input.

Yes exactly

[icexelloss]

@rtpsw Looks like you added another atomic variable here:
https://github.com/apache/arrow/pull/35874/files#r1218191706

Can you fix that one too?

[icexelloss]

@rtpsw Looks like you added another atomic variable here: https://github.com/apache/arrow/pull/35874/files#r1218191706

Can you fix that one too?

Nvm looks like you fixed this

[icexelloss]

LGTM

[rtpsw]

This makes sense to me. Let me verify I understand:

Before, on InputReceived, we would update the memo for an input based on the new batch.

Now, the only thing that really happens on InputReceived, is the batch is placed into the appropriate input queue. The update to the memo table happens when teh processing thread runs, as it is processing batches for an input.

As @icexelloss said, this is correct. However, there is another difference to note. The pre-PR code invalidates the key hasher and updates the memo-store time given any batch, whereas the post-PR code does so given all but the first batch (after batches_processed_ is incremented). This is fine because the key hasher is initialized as invalid and the memo-store time is initialized to the lowest value.

[conbench-apache-arrow]

Conbench analyzed the 6 benchmark runs on commit 1b29d958.

There were no benchmark performance regressions. 🎉

The full Conbench report has more details.