storage: don't perform one-phase commit transactions after restarts

[nvanbenschoten]

Fixes #40466.

This commit adjusts the Replica-side logic around executing one-phase commit transactions. It prevents batches that could be executed on the 1PC fast-path from doing so if they contain transactions in any epoch other than their first epoch. We saw in #40466 that this could lead to intent abandonment if the txn had written intents in earlier epochs.

I believe that this is a new issue that was introduced when we moved from using the presence of a BeginTxn and EndTxn in the same batch to detect 1PC transactions to using the sequence numbers of writes and the presence of an EndTxn in a batch to detect 1PC transactions. That change was necessary for parallel commits. Before that change, I think logic in DistSender was preventing us from hitting this case because it would always split EndTxn requests from the rest of its batch when a txn had been restarted:

cockroach/pkg/kv/dist_sender.go

Lines 692 to 697 in 18bdfe1

	// To ensure that we lay down intents to prevent starvation, always
	// split the end transaction request into its own batch on retries.
	// Txns requiring 1PC are an exception and should never be split.
	if ba.Txn != nil && ba.Txn.Epoch > 0 && !require1PC {
	splitET = true
	}

[cockroach-teamcity]

This change is 

[tbg]

Reviewed 3 of 3 files at r1, 3 of 3 files at r2, 1 of 1 files at r3, 2 of 2 files at r4, 3 of 3 files at r5.
Reviewable status: complete! 0 of 0 LGTMs obtained

[nvanbenschoten]

bors r+

[craig]

Build succeeded

• GitHub CI (Cockroach)