release-19.2: storage: don't corrupt unsafe value buffer in replayTransactionalWrite

[nvanbenschoten]

Backport 1/1 commits from #52234.

/cc @cockroachdb/release

---

Fixes #49482.
Fixes #50386.
Fixes #50683.

This commit fixes a bug in replayTransactionalWrite that resulted in
incorrect "transaction %s with sequence %d has a different value %+v
after recomputing from what was written" errors being thrown when write
requests were reissued, often due to transaction refreshes after
partially successful batches. The bug was caused by an unsafe byte
buffer acquired through Iterator.UnsafeValue being used after its
iterator had been moved. Moving an iterator invalidates and potentially
corrupts unsafe memory. This caused a sanity check later in the function
to occasionally fail.

I had a hard time reproducing this in a unit test. Even with very large
keys, very large values, interspersed compactions, and a durable Pebble
instance, I couldn't create a situation where moving the iterator
actually caused corruption in value previously retrieved using
iter.UnsafeValue. Maybe I'm missing a trick there. Regardless, this
certainly seems to fix the failures observed in #50683, as I can no
longer reproduce it under stress.

This will need to be backported to release-20.1, release-19.2, and
release-19.1, as all three release branches are susceptible to the
spurious errors that this bug can result in.

Release note (bug fix): Large write requests no longer have a chance
of erroneously throwing a "transaction with sequence has a different
value" error.

[cockroach-teamcity]

This change is 

[petermattis]

I assume this backported cleanly modulo the test.

Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @petermattis)

[nvanbenschoten]

TFTR!

I assume this backported cleanly modulo the test.

Yes, this backported cleanly modulo the test.