rowflow: account for more memory usage of the row buffer

[yuzefovich]

Previously, we accounted for the memory usage of the row buffer from
which the rows are pushed from in the routers only when copying the rows
from the row container, but we also have another in-memory row buffer
that we can move the rows from which was missing the memory accounting.
This is now fixed which also deflakes the router disk spill test.

Fixes: #55848.

Release note: None

[cockroach-teamcity]

This change is 

[yuzefovich]

I'm open to suggestions on how to actually de-flake the test without removing it, but I can't seem to figure it out. Locally the tests fails after a few hundred of iterations under stress as is.

[asubiotto]

Thanks for looking at this. Does this also fix the race in #55848?

Could you describe the reason this fails in more detail? What does a rare case where this error does not occur look like? Even if the producer and consumer goroutines run concurrently, why does that affect the occurence of the error?

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @asubiotto)

[yuzefovich]

Yeah, it does fix the race too.

To be honest, I don't understand the reason why the test as is is flaky - sometimes the expected memory error doesn't occur which (sometimes). I'd appreciate if you took a look.

I thought that one explanation for the flakiness could be that the limit on test-monitor is set pretty high so that we can add a bout half of all rows to the in-memory container before spilling to disk. Once the spilling occurs, then the memory account of the row container is cleared, and the consumer goroutine never buffers up enough rows in rowBufToPushFrom account to hit that limit. I tried this diff to prototype a separate monitor with a limit of 1, but that still is flaky :/

Never mind, I think I figured it out here. Will kick off a nightly stress on a gceworker.

[yuzefovich]

No failures in 400k run in over an hour, so I assume that fixes it, RFAL.

[asubiotto]

Reviewable status: complete! 1 of 0 LGTMs obtained

[yuzefovich]

The build is red because of a logic test flake.

TFTR!

bors r+

[craig]

Build succeeded:

• GitHub CI (Cockroach)