Threshold should only produce positive updates

[frankmcsherry]

Threshold would set negative diffs to zero, but that confused other parts of the system. Specifically (for future ref) differential's consolidate method, which we used, doesn't seem to expect zero diffs and does not proactively filter out those that it does not create.

Fixes #2464

[benesch]

Hang on, this is a differential bug by my reckoning.

[benesch]

I have an alternative fix coming in shortly and then we can discuss!

[frankmcsherry]

Hang on, this is a differential bug by my reckoning.

You'll have to convince me. :) I can see it being a doc bug, but I think consolidate is working as intended (just, it doesn't state its prereqs well enough).

[benesch]

TimelyDataflow/differential-dataflow#264

The docs on consolidate state:

/// This method will sort `vec[offset..]` and then consolidate runs with the same first
/// element of a pair by accumulating the second elements of the pairs. Should the final
/// accumulation be zero, the element is discarded.

And yet if the slice has one element like [("a", 0)] consolidate will fail to remove that element.

[benesch]

Might as well do this too!

[cuongdo]

@frankmcsherry could you add a relevant release note to doc/user/release-notes.md?

[frankmcsherry]

Tbh, I'm not sure what to add. There was a bug that caused a crash, and it doesn't crash any more. There should be zero user-facing changes, other than fewer crashes. Any tips?

[frankmcsherry]

Looking at the description of the file, I don't think this qualifies as a "major bug fix". I'm up for being corrected on that, though.