Skip state independent checks in prepare proposal

[rootulp]

Context

celestia-app/app/prepare_proposal.go

Lines 36 to 38 in 1d73e33

	// TODO: we can remove all state independent checks from the ante handler here such as signature verification
	// and only check the state dependent checks like fees and nonces as all these transactions have already
	// passed CheckTx.

[rootulp]

It's not clear to me what the benefits are from doing this. It seems like a small performance improvement that we should only pursue if we have data indicating that:

• (a) prepare proposal takes too long and
• (b) the ante handler step in prepare proposal is taking too much time relative to the rest of the operations

[evan-forbes]

it seems a bit risky to me as well. It's true that we've ran check tx successfully on all of the txs, however it seems like it would be easy to let txs slip through the cracks. For example, aren't we skipping signature verification when we call ReCheckTx? if that's the case, is it possible that a transaction gets included that invaldates the nonce of a tx that makes it to Prepare? What happens if a validator isn't running ReCheckTx?

we can quantify how long this is taking with further tracing and profiling, then re-evaluate.

[cmwaters]

It's not clear to me what the benefits are from doing this. It seems like a small performance improvement that we should only pursue if we have data indicating that:

Yeah I agree that we should benchmark how long these operations take to get a better idea of the relative times. Maybe this is a good first step. We are benchmarking a lot of the network latencies as we're assuming that's were the biggest gains can be made, but it may also be a good idea to benchmark more of the computational work i.e. state execution, square construction, checkTx and the other abci calls as these do end up having significant impact on the total system (imagine waiting 3 seconds to process a block before you can continue to gossip it to peers that haven't seen the block yet).

[evan-forbes]

adding a plot here from a previous analysis for posterity, when the square size is increased, there are dramatic increases in the amount time required for nodes to process and prepare a proposal block. Worse yet, prepare and process proposal cannot overalp (without pipelining ofc) and are blocking, meaning that any increase in time verifying the data root is doubled.

This plot shows square sizes that are 512x512 ODS (anything greater than 32MB)

[rootulp]

Cool plot! IMO we need to drill into Prepare/ProcessProposal to see if the ante handler is a bottleneck before we do this issue.

[cmwaters]

Would be interested to see how much time the application is also handling CheckTx requests in a height