storage: strip Transaction proto from individual Response messages

[nvanbenschoten]

This commit ensures that we strip Transaction protos from individual Response message headers before they are put on this wire. These Transaction objects are already used to update the Transaction proto on the BatchResponse header, so there's no reason to include these in individual Responses as well.

Before this change, the following Response types contained the extra information:

• EndTransactionResponse
• BeginTransactionResponse
• HeartbeatTxnResponse
• QueryIntentResponse

This reduces total network traffic in a cluster running non-1PC transactions by up to 3.00%. This was determined through the following experiment:

Cluster: 3x n1-standard-16 VMs
Load gen: 1 per node to avoid leaseholder effects (33% of KV RPCs are local)
          on same hosts as crdb nodes to avoid client<->gateway network traffic
Setup: `workload init indexes --secondary-indexes=1`
Workload: (on each VM) `workload run indexes --secondary-indexes=1 --payload=1 --concurrency=60 --max-rate=2000 --ramp=1m --duration=3m`
          --max-rate ensures that both trials run the exact same number of txns

The workload was run first without this change (on the left) and next with this change (on the right). The table was dropped in between.

We see from the aggregated results that without this change, the cluster transmitted 7.01 GB on the network over the course of the benchmark. With this change, the cluster transmitted 6.80 GB on the network over the course of the benchmark. This indicates a reduction in network traffic by 3.00%.

We can do out the math here to see whether the results of this experiment add up. The change avoids transmitting 2 Transaction protos per transaction over the network. These were in the responses of EndTxn(STAGING) and of EndTxn(COMMITTED) (see parallel commits). In the cluster, we were running exactly 6,000 transactions per second. So this change avoids 12,000 Transaction protos per second. Assuming these protos have an average size of 116 bytes (the encoded size of nonZeroTxn in roachpb/data_test.go), we would expect this to result in a reduction in network throughput of 1.4 MB/s. This is roughly what we see in the third graph.

Release note (performance improvement): Individual response messages in a response batch no longer each contain information about transaction state changes.

[cockroach-teamcity]

This change is 

[tbg]

Nice.

Reviewed 4 of 4 files at r1.
Reviewable status: complete! 1 of 0 LGTMs obtained

[nvanbenschoten]

bors r+

[craig]

Build succeeded

• GitHub CI (Cockroach)