uses varint encoding for vote-state lockout offsets

[behzadnouri]

Problem

VoteStateUpdate has large serialized size.

Summary of Changes

The commit removes CompactVoteStateUpdate and instead reduces serialized
size of VoteStateUpdate using varint encoding for vote-state lockout
offsets.

[AshwinSekar]

Since offset already has known bounds does this save us much vs using fixed? And is there a performance hit when serializing / deserializing varints?

Makes the code cleaner so I think it's a better solution, curious about the metrics.

[AshwinSekar]

can be u8

[behzadnouri]

done.

any reason the original VoteStateUpdate defines a u32?

[AshwinSekar]

Was copied over from the original Vote which defined confirmation_count as u32. @carllin was there a specific reason for that?

[carllin]

Cool stuff, would be good to get a size comparison between this and the existing compactvotestate for a full tower (i.e. a vote with 32 slots)

[behzadnouri]

Since offset already has known bounds does this save us much vs using fixed? And is there a performance hit when serializing / deserializing varints?

yeah, might not save much. It depends on the actual value of offsets.
There is only one lockouts vector so we save the vector length encoding for the other two.
There is also no root_to_first_vote_offset: u64.
Also in cases where 21 bits are enough to encode the offset we save one byte compared to lockouts_32.

On the other hand there are instances we might need an extra byte due to most significant bit being used for varint encoding.
So at the end depends on the distribution of offsets.

The commit also removes compact/uncompact back-and-forth, and simplifies the code a lot. So it might be better in that regard even if the actual bytes savings are not significant.

[bw-solana]

Collected some data comparing the encoding methods with a validator running on testnet. Here's the average serialized size of vote state for each (sample size of 57k votes observed):

• VoteStateUpdate = 430B
• CompactVoteStateUpdate = 176B
• serde_compact_vote_state = 116B <-- Note my experiment added enum wrapper (extra 4B)

I was seeing very little variation for each. Very sparingly size would dip to the following minimum values

• VoteStateUpdate = 418B
• CompactVoteStateUpdate = 174B
• serde_compact_vote_state = 114B

Data can be found here.

Pull request has been modified.