Epic: pre-launch performance testing safekeeper and follow ups

[stepashka]

Motivation

When we launch, Compute -> Safekeeper performance should be roughly on par with streaming postgres replication.

DoD

we have rough numbers that compare the performance of Zenith with PG streaming replication
low hanging fruits for improvements are implemented, tasks for bigger issues are created and linked.

Tasks

• Basic safekeeper performance numbers #799
• wal_proposer is stuck for a long time with high CPU usage #780
• unnecessary queue traversals in WAL proposer
• Switch to async proposer <-> safekeeper protocol #858
• fix performance issue with multiple connections #960
• Low local performance with single safekeeper compared to single synchronous replica #1141
• Optimize flushes and AppendResponses in safekeeper #1144

Follow ups for the future that came up from this:

• Try packing more data into each AppendRequestMessage for better performance #1005

[petuhovskiy]

I think we reached DoD, should we close this epic now? @stepashka

[stepashka]

reopened to address the discoveries from #1141

[petuhovskiy]

#1141 is fixed now